* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: eager patches + torch.compile cache speed phase
Adds the opt-in speed path for the GGUF diffusion transformer behind a
selectable speed mode (default off, so output is unchanged until a profile
is chosen):
- diffusion_eager_patches.py: shared eager fast-paths (channels_last,
attention/backend selection, fused norms and QKV) installed at load and
rolled back on unload or failed load.
- diffusion_compile_cache.py / diffusion_gguf_compile.py: a persistent
torch.compile cache and the GGUF-transformer compile wiring.
- diffusion_arch_patches.py: architecture-specific patches.
- diffusion_patch_backend.py: shared install/restore plumbing.
- diffusion_speed.py: speed-profile planning.
Tests for each module plus the benchmarking and probe scripts used to
measure speed, memory, and accuracy of the path.
* Studio diffusion: image workflows (safetensors, image-conditioned, editing) + Images UI
Backend:
- Load non-GGUF safetensors models: full bnb-4bit pipelines and single-file
fp8 transformers, gated to the unsloth org plus a curated allowlist.
- Image-conditioned workflows built with Pipeline.from_pipe so they reuse the
loaded transformer/VAE/text-encoder with no extra VRAM: img2img, inpaint,
outpaint, and a hires-fix upscale pass.
- Instruction editing as its own family kind (Qwen-Image-Edit-2511,
FLUX.1-Kontext-dev) and FLUX.2-klein reference conditioning (single and
multi-reference) plus klein inpaint.
- Auto-resize odd-sized inputs to a multiple of 16 (and resize the matched
mask) so img2img/inpaint/edit no longer reject non-/16 uploads. Bound the
decoded image size and cap upscale output to avoid OOM on large inputs.
- Fixes: from_pipe defaulting to a float32 recast that crashed torchao
quantized transformers; image-conditioned calls forcing the slider size
onto the input image. Native sd.cpp engine rejects image-conditioned and
reference requests it cannot serve.
Frontend:
- Redesigned Images page with capability-gated workflow tabs (Create,
Transform, Inpaint, Extend, Upscale, Reference, Edit), a brush mask editor,
client-side outpaint, and a multi-reference picker.
- Advanced options moved to a right-docked panel mirroring Chat: closed by
default, toggled by a single fixed top-bar button that stays in place.
sd.cpp installer: pin the release, verify each download's sha256, add a
download timeout, and make the source repo configurable for a future mirror.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: correct the Advanced panel comment (closed by default, fixed toggle)
* Studio diffusion: LoRA adapters for the Images workflow
Add community LoRA support across both diffusion backends, the single
biggest step toward broad image-workflow coverage.
Backend
- New shared module core/inference/diffusion_lora.py: adapter discovery
(local scan + curated catalog + owner/name[:file] Hub refs), download
via hf_hub_download_with_xet_fallback, alias sanitization, native
managed-dir materialization with collision-broken aliases, prompt-tag
injection (deduped against user-typed tags), and a supports_lora gate.
- Native sd-cli: resolve + materialize selected LoRAs into a per-run
managed dir, inject <lora:ALIAS:w> tags, pass --lora-model-dir with
--lora-apply-mode auto. The arg builder already emitted these flags.
- Diffusers: non-fused load_lora_weights + set_adapters manager, tracked
on the pipe so an unchanged selection is a no-op and a model swap
resets; cleared on unload. Never fuses (breaks quantized transformers
and blocks live weight tweaks).
- Gated off where unsupported: torchao fp8/int8 dense, GGUF-via-diffusers,
and native Qwen-Image (no LoRA name-conversion branch upstream).
- Request contract: optional loras on DiffusionGenerateRequest; empty or
omitted is identical to today. supports_lora surfaced in status; chosen
LoRAs persisted in gallery recipe metadata.
- New GET /api/models/diffusion-loras for the picker (family-filtered).
Frontend
- Repeatable multi-LoRA picker (adapter select + weight slider 0..2 +
remove), gated by the loaded model's supports_lora and family, max 8.
Tests
- New test_diffusion_lora.py (14): helpers, request validation, native
tag/dir wiring, diffusers set_adapters manager, supports_lora matrix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: ControlNet for the Images workflow (diffusers)
Add ControlNet conditioning, the #2 most-used diffusion workflow after
LoRA, on the diffusers backend for the families with ControlNet pipelines
(FLUX.1 and Qwen-Image), with Union models as the default picks.
Backend
- New core/inference/diffusion_controlnet.py: family-gated discovery
(curated Union models + local dirs + bare owner/name repos), resolution
to a loadable repo/dir, control-image preprocessing (passthrough +
a dependency-free canny edge map), and a supports_controlnet gate.
- diffusion.py: a ControlNet manager parallel to the LoRA one. Loads the
(small) ControlNet model once via from_pretrained (cached by id) and
builds the family's ControlNet pipeline via Pipeline.from_pipe(base,
controlnet=model), reusing the resident base modules at their loaded
dtype (no reload, no recast). Passes the control image + conditioning
scale + guidance start/end at generate time; cleared on unload.
- Families: FLUX.1 -> FluxControlNetPipeline/Model, Qwen-Image ->
QwenImageControlNetPipeline/Model. Others declare none (gated off).
- Gated off for the native engine, GGUF-via-diffusers, and torchao
fp8/int8 dense (same rule as LoRA). v1 conditions txt2img only.
- Request contract: optional controlnet on DiffusionGenerateRequest;
supports_controlnet in status; the choice persisted in gallery meta.
- New GET /api/models/diffusion-controlnets for the picker.
Frontend
- A ControlNet control in the Images rail (model select + control-image
upload + control-type select + strength slider), gated by the loaded
model's supports_controlnet + family, shown for text-to-image.
Tests
- New test_diffusion_controlnet.py (10): discovery/resolve/preprocess/gate
helpers, request validation, family wiring, and the diffusers pipe
manager (loads once, caches, from_pipe with controlnet, rejects
unsupported families).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio ControlNet: show the picker on the Create tab (workflow id is 'create', not 'txt2img')
The ControlNet control gated on workflow === 'txt2img', but the Images workflow tab ids are create/transform/inpaint/extend/upscale/reference/edit -- there is no 'txt2img'. So the picker never rendered even with a ControlNet-capable model loaded. Gate on 'create' (the text-to-image tab) for both the picker and the request wiring. Found via a live Playwright capture of the running Studio.
* Studio: do not force diffusers pipelines cross-tagged gguf into the GGUF variant expander
Some diffusers image repos (e.g. unsloth/Qwen-Image-2512-unsloth-bnb-4bit) carry a
stray "gguf" tag on the Hub but ship no .gguf files. The model search classified
them as GGUF from the bare tag, so the picker rendered the GGUF variant expander,
which then dead-ended at "No GGUF variants found." Trust the bare gguf tag only when
the repo is not a diffusers pipeline; the -GGUF name suffix and real gguf metadata
(populated via expand=gguf) remain authoritative, so genuine GGUF repos are unaffected.
* Studio Images: load non-curated unsloth/on-device diffusers repos instead of no-op
handleModelSelect only loaded curated safetensors ids and GGUF variant picks; any other
non-GGUF pick (an on-device diffusers folder, or a future unsloth diffusers image repo
surfaced by search) silently did nothing. Treat such a pick as a full diffusers pipeline
load when the id is unsloth-hosted or on-device (the backend infers the family + base repo
and gates loads to unsloth/* or local paths), and show a clear message otherwise instead
of silently ignoring the click. Curated and GGUF paths are unchanged.
* Studio Images: keep curated safetensors models in Recommended after download
The curated bnb-4bit / fp8 diffusion rows were filtered out of the Images picker's
Recommended list once cached (curatedSafetensorsRows dropped anything in downloadedSet),
so they vanished from the picker after the first load and could only be found by typing an
exact search. The row already renders a downloaded badge, matching how GGUF Recommended
rows stay visible when cached. Drop the exclusion so the curated safetensors always list.
* Studio diffusion LoRA: sanitize dots out of adapter aliases
The LoRA alias is used as the diffusers PEFT adapter name, and PEFT rejects names
containing "." (module name can't contain "."). sanitize_alias kept dots, so a LoRA whose
filename carries a version tag (e.g. Qwen-Image-2512-Lightning-8steps-V1.0-bf16) failed to
apply with a 400. Replace dots too; the alias stays a valid native <lora:NAME:w> filename
stem. Adds regression coverage for internal dots.
* Studio Images: clarify the GGUF transformer-quant Advanced control
Renamed the confusing "Transformer quant / GGUF default" control to "GGUF speed mode"
with an "Off (run the GGUF)" default, and reworded the hint to state plainly that FP8/INT8/
FP4 load the FULL base model (larger download + more VRAM) rather than re-packing the GGUF,
falling back to the GGUF if it can't fit. Behavior unchanged; labels/hint only.
* Studio Images: list on-device unsloth diffusion models in the picker
The Images picker's On Device tab hid every non-GGUF cached repo whenever a
task filter was active, so downloaded unsloth diffusion pipelines (bnb-4bit
and FP8 safetensors) never showed up there. List cached repos that pass the
task gate, limited under a filter to unsloth-hosted ones so base repos (which
fail the diffusion load trust gate) don't appear only to dead-end on click.
Chat behavior is unchanged: the task gate still drops image repos there.
* Studio: hide single-file image checkpoints from the chat model picker
The chat picker treats a cached repo as an image model, and hides it, only
when it ships a diffusers model_index.json. Single-file, ComfyUI, and
ControlNet image checkpoints (an FP8 Qwen-Image, a z-image safetensors, a
Qwen-Image ControlNet) carry none, so they surfaced as loadable chat models.
Fall back to resolving the repo id against the known diffusion families, the
same resolver the Images backend loads from, so these checkpoints are tagged
text-to-image and stay in the Images picker only.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: add the FLUX.2-dev model family
Loading unsloth/FLUX.2-dev-GGUF failed because detect_family knew only the
Qwen3-based FLUX.2-klein, so FLUX.2-dev (the full, Mistral-based Flux2Pipeline)
resolved to nothing and the load errored. Add a flux.2-dev family: Flux2Pipeline
+ Flux2Transformer2DModel over the black-forest-labs/FLUX.2-dev base repo (gated,
reachable with an HF token), with its FLUX.2 32-channel VAE and Mistral text
encoder wired for the sd-cli path from the open Comfy-Org/flux2-dev mirror.
text-to-image only: diffusers 0.38 ships no Flux2 img2img / inpaint pipeline for
dev. Frontend gets sensible dev defaults (28 steps, guidance 4), distinct from
klein's turbo defaults. Verified live: GGUF load resolves the family + gated base
repo and generates a real 1024x1024 image on GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: clearer error for an unsupported diffusion model
When a repo id resolves to no diffusion family the load raised 'Could not infer a
diffusion family... Pass family_override (z-image)', which points at an unrelated
family and doesn't say what is supported. Replace it with a message that lists the
supported families (from a new supported_family_names helper) and notes that video
models and image models whose diffusers transformer has no single-file loader are
not supported. Applies to both the diffusers and native sd.cpp load paths. Also
refreshes two stale family-registry comments that still called FLUX.2-dev omitted.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove stray async task scratch outputs committed by mistake
* ControlNet: reject filesystem-like ids and do not cache a model past an unload race
Two review findings on the ControlNet path:
- resolve_controlnet's bare-repo fallback accepted any id with a slash, so a
path-shaped id (/tmp/x, ../x) reached from_pretrained as a local directory.
Restrict the fallback to a strict owner/name HF repo id shape.
- _controlnet_pipe now re-checks the cancel event after the blocking
from_pretrained: an unload that raced the download had already cleared the
caches, so caching the late module would pin it past the unload.
* ControlNet: address review findings on the diffusers path
- resolve_controlnet enforces catalog family compatibility so a direct API call
cannot load a ControlNet built for another family through the wrong pipeline.
- Unknown ControlNet ids now surface as a 400 (call site maps FileNotFoundError
to ValueError) instead of a generic 500.
- strength 0 disables ControlNet entirely, so a no-op selection never pays the
download / VRAM cost; the control image is decoded and validated BEFORE the
ControlNet is resolved or built, so a malformed image fails fast for the same reason.
- ControlNet loads use the base compute dtype (state.dtype is a display string,
not a torch.dtype, so it silently fell back to float32) and honor the base
offload policy via group offloading instead of forcing the module resident.
- Empty/malformed HF token coerced to anonymous access.
- Flux Union ControlNet control_mode mapped from the selected control type.
- resolve_controlnet drops the unused hf_token/cancel_event params.
- ControlNetSpec validates guidance_start <= guidance_end (clean 422).
- Images UI ControlNet Select shows its placeholder when nothing is selected.
Adds regression tests for family enforcement and the union control-mode map.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Diffusion LoRA: harden resolution, native tag precedence, and diffusers teardown
Address review findings on the LoRA path:
- resolve_one: normalise a blank/whitespace hf_token to None (anonymous access)
and reject a client-supplied weight file with traversal / absolute path.
- resolve_specs: convert FileNotFoundError from an unknown/stale id to ValueError
so the route returns 400 instead of a generic 500.
- _scan_local: disambiguate local adapters that share a stem (foo.safetensors vs
foo.gguf) so each is uniquely addressable.
- inject_prompt_tags: the backend-validated weight now wins over a user-typed
<lora:ALIAS:...> for a selected adapter; unselected user tags are left alone.
- diffusers _apply_loras: reject a .gguf adapter with a clear error before touching
the pipe (diffusers loads safetensors only).
- _unload_locked: drop the explicit unload_lora_weights() on teardown; the pipe is
dropped wholesale (freeing adapters), so the previous call could race an in-flight
denoise on the same pipe.
- Images page: use a stable LoRA key and clear the selection (not just the options)
when the catalog refresh fails.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Diffusion: guard trust check against OSError and validate conditioning inputs
- _is_trusted_diffusion_repo: wrap Path.exists() so a repo id with invalid
characters (or a bare owner/name id) can't raise OSError; treat any failure as
not-a-local-path and fall through to the unsloth/ allowlist. validate_load_request
still raises the clear FileNotFoundError for a genuinely missing local pick.
- generate(): reject mask_image / upscale / reference_images supplied without an
input image, and reject reference_images on a family that does not support
reference conditioning, instead of silently degrading to txt2img / img2img.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review findings on the image-workflows PR
Keep diffusion.py importable without torch: the compile/arch patch modules
import torch at module level, so import them lazily at their load/unload
call sites instead of at module load. This restores the torchless contract
so get_diffusion_backend() works on a CPU/native sd.cpp install.
Match family reject keywords and aliases as whole path/name segments, not
raw substrings, so an unrelated word like edited, edition, or kontextual no
longer misroutes or hides a valid base image model, while supported edit
families (Qwen-Image-Edit, FLUX Kontext) still resolve. Mirror the same
segment matching in the picker task filter.
Route FLUX.2-dev native guidance through --guidance like the other FLUX
families rather than --cfg-scale. Reject native upscale requests that have
no input image. Read image header dimensions and reject over-limit inputs
before decoding pixels, so a crafted small-payload image cannot spike
memory. Reject an upscale that would shrink the source below its input
size. Validate the model_kind against the filename extension before the
GPU handoff. Estimate a local diffusers pipeline's size from its on-disk
weights so auto memory planning does not skip offload and OOM. Report
workflows: [txt2img] from the native backend status so the Create tab
stays enabled for a loaded native model. Clamp the outpaint canvas to the
backend's 4096px decode limit.
Adds regression tests for segment matching and kind/extension validation.
* Harden diffusion LoRA handling on the diffusers and native paths
Reject LoRA on a torch.compile'd diffusers transformer (Speed=default/max):
diffusers requires the adapter loaded before compilation, so applying one to
the already-compiled module fails with adapter-key mismatches. The status
gate now hides the picker and generate raises a clear message instead.
Convert a cancelled Hub LoRA download (RuntimeError Cancelled) to the
diffusion cancellation sentinel in resolve_specs, so an unload/superseding
load during resolution maps to a 409 instead of a generic server error.
Drop weight-0 LoRA rows before the native support gate so a request carrying
only disabled adapters stays a no-op on families where native LoRA is
unsupported, matching the diffusers path.
Reject duplicate LoRA ids in the request model: both apply paths suffix
colliding names, so a repeated id would stack the same adapter past its
per-adapter weight bound.
Strip all user-typed <lora:...> prompt tags on the native path (only the
selected adapters are materialized in the managed lora-model-dir, so an
unselected tag can never resolve), and restore saved LoRA selections from a
gallery recipe so restore reproduces a LoRA image.
* Harden ControlNet resolve, gallery metadata, and the control-type picker
Check cancellation immediately after a ControlNet from_pretrained and before
any device placement, so an unload/eviction that raced the download does not
allocate several GB onto the GPU after the load was already cleared.
Require a loadable weight or shard index (not just config.json) before a local
ControlNet folder is advertised, so an interrupted copy is hidden instead of
failing deep in from_pretrained as a generic 500.
Do not record a strength-0 ControlNet in the gallery recipe: it is treated as
disabled and skipped, so the image is unconditioned and the metadata must not
claim a ControlNet was applied.
Build the control-type picker from the selected ControlNet's advertised
control_types instead of a hardcoded passthrough/canny pair, so a union model
with a precomputed depth or pose map sends the correct control_mode.
* Address further Codex findings on the image-workflows PR
- Persist the actual output image size in the gallery recipe instead of the
request sliders: Transform/Inpaint/Edit derive the size from the uploaded
image, Extend grows the canvas, and Upscale resizes it, so the sliders
recorded (and later restored) the wrong dimensions for those workflows.
- Reject a remote '*-GGUF' repo loaded as a full pipeline (no single-file
name) in validate_load_request, so the unloadable pick fails before chat is
evicted rather than deep in from_pretrained.
- Only publish an image-conditioned from_pipe wrapper to the shared aux cache
when the load is still current: from_pipe runs under the generate lock but
not the state lock, so an unload racing its construction could otherwise
cache a wrapper over torn-down modules that a later load would reuse.
- Verify the Windows CUDA runtime archive checksum before extracting it, like
the main sd-cli archive, so a corrupt or tampered runtime is rejected rather
than extracted next to the binary.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: eager patches + torch.compile cache speed phase
Adds the opt-in speed path for the GGUF diffusion transformer behind a
selectable speed mode (default off, so output is unchanged until a profile
is chosen):
- diffusion_eager_patches.py: shared eager fast-paths (channels_last,
attention/backend selection, fused norms and QKV) installed at load and
rolled back on unload or failed load.
- diffusion_compile_cache.py / diffusion_gguf_compile.py: a persistent
torch.compile cache and the GGUF-transformer compile wiring.
- diffusion_arch_patches.py: architecture-specific patches.
- diffusion_patch_backend.py: shared install/restore plumbing.
- diffusion_speed.py: speed-profile planning.
Tests for each module plus the benchmarking and probe scripts used to
measure speed, memory, and accuracy of the path.
* Studio diffusion: image workflows (safetensors, image-conditioned, editing) + Images UI
Backend:
- Load non-GGUF safetensors models: full bnb-4bit pipelines and single-file
fp8 transformers, gated to the unsloth org plus a curated allowlist.
- Image-conditioned workflows built with Pipeline.from_pipe so they reuse the
loaded transformer/VAE/text-encoder with no extra VRAM: img2img, inpaint,
outpaint, and a hires-fix upscale pass.
- Instruction editing as its own family kind (Qwen-Image-Edit-2511,
FLUX.1-Kontext-dev) and FLUX.2-klein reference conditioning (single and
multi-reference) plus klein inpaint.
- Auto-resize odd-sized inputs to a multiple of 16 (and resize the matched
mask) so img2img/inpaint/edit no longer reject non-/16 uploads. Bound the
decoded image size and cap upscale output to avoid OOM on large inputs.
- Fixes: from_pipe defaulting to a float32 recast that crashed torchao
quantized transformers; image-conditioned calls forcing the slider size
onto the input image. Native sd.cpp engine rejects image-conditioned and
reference requests it cannot serve.
Frontend:
- Redesigned Images page with capability-gated workflow tabs (Create,
Transform, Inpaint, Extend, Upscale, Reference, Edit), a brush mask editor,
client-side outpaint, and a multi-reference picker.
- Advanced options moved to a right-docked panel mirroring Chat: closed by
default, toggled by a single fixed top-bar button that stays in place.
sd.cpp installer: pin the release, verify each download's sha256, add a
download timeout, and make the source repo configurable for a future mirror.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: correct the Advanced panel comment (closed by default, fixed toggle)
* Studio diffusion: LoRA adapters for the Images workflow
Add community LoRA support across both diffusion backends, the single
biggest step toward broad image-workflow coverage.
Backend
- New shared module core/inference/diffusion_lora.py: adapter discovery
(local scan + curated catalog + owner/name[:file] Hub refs), download
via hf_hub_download_with_xet_fallback, alias sanitization, native
managed-dir materialization with collision-broken aliases, prompt-tag
injection (deduped against user-typed tags), and a supports_lora gate.
- Native sd-cli: resolve + materialize selected LoRAs into a per-run
managed dir, inject <lora:ALIAS:w> tags, pass --lora-model-dir with
--lora-apply-mode auto. The arg builder already emitted these flags.
- Diffusers: non-fused load_lora_weights + set_adapters manager, tracked
on the pipe so an unchanged selection is a no-op and a model swap
resets; cleared on unload. Never fuses (breaks quantized transformers
and blocks live weight tweaks).
- Gated off where unsupported: torchao fp8/int8 dense, GGUF-via-diffusers,
and native Qwen-Image (no LoRA name-conversion branch upstream).
- Request contract: optional loras on DiffusionGenerateRequest; empty or
omitted is identical to today. supports_lora surfaced in status; chosen
LoRAs persisted in gallery recipe metadata.
- New GET /api/models/diffusion-loras for the picker (family-filtered).
Frontend
- Repeatable multi-LoRA picker (adapter select + weight slider 0..2 +
remove), gated by the loaded model's supports_lora and family, max 8.
Tests
- New test_diffusion_lora.py (14): helpers, request validation, native
tag/dir wiring, diffusers set_adapters manager, supports_lora matrix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: do not force diffusers pipelines cross-tagged gguf into the GGUF variant expander
Some diffusers image repos (e.g. unsloth/Qwen-Image-2512-unsloth-bnb-4bit) carry a
stray "gguf" tag on the Hub but ship no .gguf files. The model search classified
them as GGUF from the bare tag, so the picker rendered the GGUF variant expander,
which then dead-ended at "No GGUF variants found." Trust the bare gguf tag only when
the repo is not a diffusers pipeline; the -GGUF name suffix and real gguf metadata
(populated via expand=gguf) remain authoritative, so genuine GGUF repos are unaffected.
* Studio Images: load non-curated unsloth/on-device diffusers repos instead of no-op
handleModelSelect only loaded curated safetensors ids and GGUF variant picks; any other
non-GGUF pick (an on-device diffusers folder, or a future unsloth diffusers image repo
surfaced by search) silently did nothing. Treat such a pick as a full diffusers pipeline
load when the id is unsloth-hosted or on-device (the backend infers the family + base repo
and gates loads to unsloth/* or local paths), and show a clear message otherwise instead
of silently ignoring the click. Curated and GGUF paths are unchanged.
* Studio Images: keep curated safetensors models in Recommended after download
The curated bnb-4bit / fp8 diffusion rows were filtered out of the Images picker's
Recommended list once cached (curatedSafetensorsRows dropped anything in downloadedSet),
so they vanished from the picker after the first load and could only be found by typing an
exact search. The row already renders a downloaded badge, matching how GGUF Recommended
rows stay visible when cached. Drop the exclusion so the curated safetensors always list.
* Studio diffusion LoRA: sanitize dots out of adapter aliases
The LoRA alias is used as the diffusers PEFT adapter name, and PEFT rejects names
containing "." (module name can't contain "."). sanitize_alias kept dots, so a LoRA whose
filename carries a version tag (e.g. Qwen-Image-2512-Lightning-8steps-V1.0-bf16) failed to
apply with a 400. Replace dots too; the alias stays a valid native <lora:NAME:w> filename
stem. Adds regression coverage for internal dots.
* Studio Images: clarify the GGUF transformer-quant Advanced control
Renamed the confusing "Transformer quant / GGUF default" control to "GGUF speed mode"
with an "Off (run the GGUF)" default, and reworded the hint to state plainly that FP8/INT8/
FP4 load the FULL base model (larger download + more VRAM) rather than re-packing the GGUF,
falling back to the GGUF if it can't fit. Behavior unchanged; labels/hint only.
* Studio Images: list on-device unsloth diffusion models in the picker
The Images picker's On Device tab hid every non-GGUF cached repo whenever a
task filter was active, so downloaded unsloth diffusion pipelines (bnb-4bit
and FP8 safetensors) never showed up there. List cached repos that pass the
task gate, limited under a filter to unsloth-hosted ones so base repos (which
fail the diffusion load trust gate) don't appear only to dead-end on click.
Chat behavior is unchanged: the task gate still drops image repos there.
* Studio: hide single-file image checkpoints from the chat model picker
The chat picker treats a cached repo as an image model, and hides it, only
when it ships a diffusers model_index.json. Single-file, ComfyUI, and
ControlNet image checkpoints (an FP8 Qwen-Image, a z-image safetensors, a
Qwen-Image ControlNet) carry none, so they surfaced as loadable chat models.
Fall back to resolving the repo id against the known diffusion families, the
same resolver the Images backend loads from, so these checkpoints are tagged
text-to-image and stay in the Images picker only.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: add the FLUX.2-dev model family
Loading unsloth/FLUX.2-dev-GGUF failed because detect_family knew only the
Qwen3-based FLUX.2-klein, so FLUX.2-dev (the full, Mistral-based Flux2Pipeline)
resolved to nothing and the load errored. Add a flux.2-dev family: Flux2Pipeline
+ Flux2Transformer2DModel over the black-forest-labs/FLUX.2-dev base repo (gated,
reachable with an HF token), with its FLUX.2 32-channel VAE and Mistral text
encoder wired for the sd-cli path from the open Comfy-Org/flux2-dev mirror.
text-to-image only: diffusers 0.38 ships no Flux2 img2img / inpaint pipeline for
dev. Frontend gets sensible dev defaults (28 steps, guidance 4), distinct from
klein's turbo defaults. Verified live: GGUF load resolves the family + gated base
repo and generates a real 1024x1024 image on GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: clearer error for an unsupported diffusion model
When a repo id resolves to no diffusion family the load raised 'Could not infer a
diffusion family... Pass family_override (z-image)', which points at an unrelated
family and doesn't say what is supported. Replace it with a message that lists the
supported families (from a new supported_family_names helper) and notes that video
models and image models whose diffusers transformer has no single-file loader are
not supported. Applies to both the diffusers and native sd.cpp load paths. Also
refreshes two stale family-registry comments that still called FLUX.2-dev omitted.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove stray async task scratch outputs committed by mistake
* Diffusion LoRA: harden resolution, native tag precedence, and diffusers teardown
Address review findings on the LoRA path:
- resolve_one: normalise a blank/whitespace hf_token to None (anonymous access)
and reject a client-supplied weight file with traversal / absolute path.
- resolve_specs: convert FileNotFoundError from an unknown/stale id to ValueError
so the route returns 400 instead of a generic 500.
- _scan_local: disambiguate local adapters that share a stem (foo.safetensors vs
foo.gguf) so each is uniquely addressable.
- inject_prompt_tags: the backend-validated weight now wins over a user-typed
<lora:ALIAS:...> for a selected adapter; unselected user tags are left alone.
- diffusers _apply_loras: reject a .gguf adapter with a clear error before touching
the pipe (diffusers loads safetensors only).
- _unload_locked: drop the explicit unload_lora_weights() on teardown; the pipe is
dropped wholesale (freeing adapters), so the previous call could race an in-flight
denoise on the same pipe.
- Images page: use a stable LoRA key and clear the selection (not just the options)
when the catalog refresh fails.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Diffusion: guard trust check against OSError and validate conditioning inputs
- _is_trusted_diffusion_repo: wrap Path.exists() so a repo id with invalid
characters (or a bare owner/name id) can't raise OSError; treat any failure as
not-a-local-path and fall through to the unsloth/ allowlist. validate_load_request
still raises the clear FileNotFoundError for a genuinely missing local pick.
- generate(): reject mask_image / upscale / reference_images supplied without an
input image, and reject reference_images on a family that does not support
reference conditioning, instead of silently degrading to txt2img / img2img.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review findings on the image-workflows PR
Keep diffusion.py importable without torch: the compile/arch patch modules
import torch at module level, so import them lazily at their load/unload
call sites instead of at module load. This restores the torchless contract
so get_diffusion_backend() works on a CPU/native sd.cpp install.
Match family reject keywords and aliases as whole path/name segments, not
raw substrings, so an unrelated word like edited, edition, or kontextual no
longer misroutes or hides a valid base image model, while supported edit
families (Qwen-Image-Edit, FLUX Kontext) still resolve. Mirror the same
segment matching in the picker task filter.
Route FLUX.2-dev native guidance through --guidance like the other FLUX
families rather than --cfg-scale. Reject native upscale requests that have
no input image. Read image header dimensions and reject over-limit inputs
before decoding pixels, so a crafted small-payload image cannot spike
memory. Reject an upscale that would shrink the source below its input
size. Validate the model_kind against the filename extension before the
GPU handoff. Estimate a local diffusers pipeline's size from its on-disk
weights so auto memory planning does not skip offload and OOM. Report
workflows: [txt2img] from the native backend status so the Create tab
stays enabled for a loaded native model. Clamp the outpaint canvas to the
backend's 4096px decode limit.
Adds regression tests for segment matching and kind/extension validation.
* Harden diffusion LoRA handling on the diffusers and native paths
Reject LoRA on a torch.compile'd diffusers transformer (Speed=default/max):
diffusers requires the adapter loaded before compilation, so applying one to
the already-compiled module fails with adapter-key mismatches. The status
gate now hides the picker and generate raises a clear message instead.
Convert a cancelled Hub LoRA download (RuntimeError Cancelled) to the
diffusion cancellation sentinel in resolve_specs, so an unload/superseding
load during resolution maps to a 409 instead of a generic server error.
Drop weight-0 LoRA rows before the native support gate so a request carrying
only disabled adapters stays a no-op on families where native LoRA is
unsupported, matching the diffusers path.
Reject duplicate LoRA ids in the request model: both apply paths suffix
colliding names, so a repeated id would stack the same adapter past its
per-adapter weight bound.
Strip all user-typed <lora:...> prompt tags on the native path (only the
selected adapters are materialized in the managed lora-model-dir, so an
unselected tag can never resolve), and restore saved LoRA selections from a
gallery recipe so restore reproduces a LoRA image.
* Address further Codex findings on the image-workflows PR
- Persist the actual output image size in the gallery recipe instead of the
request sliders: Transform/Inpaint/Edit derive the size from the uploaded
image, Extend grows the canvas, and Upscale resizes it, so the sliders
recorded (and later restored) the wrong dimensions for those workflows.
- Reject a remote '*-GGUF' repo loaded as a full pipeline (no single-file
name) in validate_load_request, so the unloadable pick fails before chat is
evicted rather than deep in from_pretrained.
- Only publish an image-conditioned from_pipe wrapper to the shared aux cache
when the load is still current: from_pipe runs under the generate lock but
not the state lock, so an unload racing its construction could otherwise
cache a wrapper over torn-down modules that a later load would reuse.
- Verify the Windows CUDA runtime archive checksum before extracting it, like
the main sd-cli archive, so a corrupt or tampered runtime is rejected rather
than extracted next to the binary.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: eager patches + torch.compile cache speed phase
Adds the opt-in speed path for the GGUF diffusion transformer behind a
selectable speed mode (default off, so output is unchanged until a profile
is chosen):
- diffusion_eager_patches.py: shared eager fast-paths (channels_last,
attention/backend selection, fused norms and QKV) installed at load and
rolled back on unload or failed load.
- diffusion_compile_cache.py / diffusion_gguf_compile.py: a persistent
torch.compile cache and the GGUF-transformer compile wiring.
- diffusion_arch_patches.py: architecture-specific patches.
- diffusion_patch_backend.py: shared install/restore plumbing.
- diffusion_speed.py: speed-profile planning.
Tests for each module plus the benchmarking and probe scripts used to
measure speed, memory, and accuracy of the path.
* Studio diffusion: image workflows (safetensors, image-conditioned, editing) + Images UI
Backend:
- Load non-GGUF safetensors models: full bnb-4bit pipelines and single-file
fp8 transformers, gated to the unsloth org plus a curated allowlist.
- Image-conditioned workflows built with Pipeline.from_pipe so they reuse the
loaded transformer/VAE/text-encoder with no extra VRAM: img2img, inpaint,
outpaint, and a hires-fix upscale pass.
- Instruction editing as its own family kind (Qwen-Image-Edit-2511,
FLUX.1-Kontext-dev) and FLUX.2-klein reference conditioning (single and
multi-reference) plus klein inpaint.
- Auto-resize odd-sized inputs to a multiple of 16 (and resize the matched
mask) so img2img/inpaint/edit no longer reject non-/16 uploads. Bound the
decoded image size and cap upscale output to avoid OOM on large inputs.
- Fixes: from_pipe defaulting to a float32 recast that crashed torchao
quantized transformers; image-conditioned calls forcing the slider size
onto the input image. Native sd.cpp engine rejects image-conditioned and
reference requests it cannot serve.
Frontend:
- Redesigned Images page with capability-gated workflow tabs (Create,
Transform, Inpaint, Extend, Upscale, Reference, Edit), a brush mask editor,
client-side outpaint, and a multi-reference picker.
- Advanced options moved to a right-docked panel mirroring Chat: closed by
default, toggled by a single fixed top-bar button that stays in place.
sd.cpp installer: pin the release, verify each download's sha256, add a
download timeout, and make the source repo configurable for a future mirror.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: correct the Advanced panel comment (closed by default, fixed toggle)
* Studio: do not force diffusers pipelines cross-tagged gguf into the GGUF variant expander
Some diffusers image repos (e.g. unsloth/Qwen-Image-2512-unsloth-bnb-4bit) carry a
stray "gguf" tag on the Hub but ship no .gguf files. The model search classified
them as GGUF from the bare tag, so the picker rendered the GGUF variant expander,
which then dead-ended at "No GGUF variants found." Trust the bare gguf tag only when
the repo is not a diffusers pipeline; the -GGUF name suffix and real gguf metadata
(populated via expand=gguf) remain authoritative, so genuine GGUF repos are unaffected.
* Studio Images: load non-curated unsloth/on-device diffusers repos instead of no-op
handleModelSelect only loaded curated safetensors ids and GGUF variant picks; any other
non-GGUF pick (an on-device diffusers folder, or a future unsloth diffusers image repo
surfaced by search) silently did nothing. Treat such a pick as a full diffusers pipeline
load when the id is unsloth-hosted or on-device (the backend infers the family + base repo
and gates loads to unsloth/* or local paths), and show a clear message otherwise instead
of silently ignoring the click. Curated and GGUF paths are unchanged.
* Studio Images: keep curated safetensors models in Recommended after download
The curated bnb-4bit / fp8 diffusion rows were filtered out of the Images picker's
Recommended list once cached (curatedSafetensorsRows dropped anything in downloadedSet),
so they vanished from the picker after the first load and could only be found by typing an
exact search. The row already renders a downloaded badge, matching how GGUF Recommended
rows stay visible when cached. Drop the exclusion so the curated safetensors always list.
* Studio Images: clarify the GGUF transformer-quant Advanced control
Renamed the confusing "Transformer quant / GGUF default" control to "GGUF speed mode"
with an "Off (run the GGUF)" default, and reworded the hint to state plainly that FP8/INT8/
FP4 load the FULL base model (larger download + more VRAM) rather than re-packing the GGUF,
falling back to the GGUF if it can't fit. Behavior unchanged; labels/hint only.
* Studio Images: list on-device unsloth diffusion models in the picker
The Images picker's On Device tab hid every non-GGUF cached repo whenever a
task filter was active, so downloaded unsloth diffusion pipelines (bnb-4bit
and FP8 safetensors) never showed up there. List cached repos that pass the
task gate, limited under a filter to unsloth-hosted ones so base repos (which
fail the diffusion load trust gate) don't appear only to dead-end on click.
Chat behavior is unchanged: the task gate still drops image repos there.
* Studio: hide single-file image checkpoints from the chat model picker
The chat picker treats a cached repo as an image model, and hides it, only
when it ships a diffusers model_index.json. Single-file, ComfyUI, and
ControlNet image checkpoints (an FP8 Qwen-Image, a z-image safetensors, a
Qwen-Image ControlNet) carry none, so they surfaced as loadable chat models.
Fall back to resolving the repo id against the known diffusion families, the
same resolver the Images backend loads from, so these checkpoints are tagged
text-to-image and stay in the Images picker only.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: add the FLUX.2-dev model family
Loading unsloth/FLUX.2-dev-GGUF failed because detect_family knew only the
Qwen3-based FLUX.2-klein, so FLUX.2-dev (the full, Mistral-based Flux2Pipeline)
resolved to nothing and the load errored. Add a flux.2-dev family: Flux2Pipeline
+ Flux2Transformer2DModel over the black-forest-labs/FLUX.2-dev base repo (gated,
reachable with an HF token), with its FLUX.2 32-channel VAE and Mistral text
encoder wired for the sd-cli path from the open Comfy-Org/flux2-dev mirror.
text-to-image only: diffusers 0.38 ships no Flux2 img2img / inpaint pipeline for
dev. Frontend gets sensible dev defaults (28 steps, guidance 4), distinct from
klein's turbo defaults. Verified live: GGUF load resolves the family + gated base
repo and generates a real 1024x1024 image on GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: clearer error for an unsupported diffusion model
When a repo id resolves to no diffusion family the load raised 'Could not infer a
diffusion family... Pass family_override (z-image)', which points at an unrelated
family and doesn't say what is supported. Replace it with a message that lists the
supported families (from a new supported_family_names helper) and notes that video
models and image models whose diffusers transformer has no single-file loader are
not supported. Applies to both the diffusers and native sd.cpp load paths. Also
refreshes two stale family-registry comments that still called FLUX.2-dev omitted.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove stray async task scratch outputs committed by mistake
* Diffusion: guard trust check against OSError and validate conditioning inputs
- _is_trusted_diffusion_repo: wrap Path.exists() so a repo id with invalid
characters (or a bare owner/name id) can't raise OSError; treat any failure as
not-a-local-path and fall through to the unsloth/ allowlist. validate_load_request
still raises the clear FileNotFoundError for a genuinely missing local pick.
- generate(): reject mask_image / upscale / reference_images supplied without an
input image, and reject reference_images on a family that does not support
reference conditioning, instead of silently degrading to txt2img / img2img.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review findings on the image-workflows PR
Keep diffusion.py importable without torch: the compile/arch patch modules
import torch at module level, so import them lazily at their load/unload
call sites instead of at module load. This restores the torchless contract
so get_diffusion_backend() works on a CPU/native sd.cpp install.
Match family reject keywords and aliases as whole path/name segments, not
raw substrings, so an unrelated word like edited, edition, or kontextual no
longer misroutes or hides a valid base image model, while supported edit
families (Qwen-Image-Edit, FLUX Kontext) still resolve. Mirror the same
segment matching in the picker task filter.
Route FLUX.2-dev native guidance through --guidance like the other FLUX
families rather than --cfg-scale. Reject native upscale requests that have
no input image. Read image header dimensions and reject over-limit inputs
before decoding pixels, so a crafted small-payload image cannot spike
memory. Reject an upscale that would shrink the source below its input
size. Validate the model_kind against the filename extension before the
GPU handoff. Estimate a local diffusers pipeline's size from its on-disk
weights so auto memory planning does not skip offload and OOM. Report
workflows: [txt2img] from the native backend status so the Create tab
stays enabled for a loaded native model. Clamp the outpaint canvas to the
backend's 4096px decode limit.
Adds regression tests for segment matching and kind/extension validation.
* Address further Codex findings on the image-workflows PR
- Persist the actual output image size in the gallery recipe instead of the
request sliders: Transform/Inpaint/Edit derive the size from the uploaded
image, Extend grows the canvas, and Upscale resizes it, so the sliders
recorded (and later restored) the wrong dimensions for those workflows.
- Reject a remote '*-GGUF' repo loaded as a full pipeline (no single-file
name) in validate_load_request, so the unloadable pick fails before chat is
evicted rather than deep in from_pretrained.
- Only publish an image-conditioned from_pipe wrapper to the shared aux cache
when the load is still current: from_pipe runs under the generate lock but
not the state lock, so an unload racing its construction could otherwise
cache a wrapper over torn-down modules that a later load would reuse.
- Verify the Windows CUDA runtime archive checksum before extracting it, like
the main sd-cli archive, so a corrupt or tampered runtime is rejected rather
than extracted next to the binary.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: persistent sd-server for the native engine (load once, serve many)
The native sd.cpp tier ran sd-cli one-shot per image, so begin_load only resolved
asset paths and every generation re-spawned sd-cli and reloaded the multi-GB GGUF
from disk (a batch of N = N full reloads). This makes it a resident backend backed by
stable-diffusion.cpp's persistent sd-server, mirroring the chat backend's llama-server
lifecycle:
- begin_load spawns sd-server once (the model loads there) and polls /v1/models until
ready; unload kills it.
- generate submits ONE async /sdcpp/v1/img_gen job for the whole batch (no reload),
polls it to completion, and decodes the returned images. Step progress and ETA come
from the server's stdout (the job JSON has no per-step field).
- The one-shot sd-cli path is kept as an automatic fallback: it is used when sd-server
is absent, and also when a present sd-server fails to start, so behavior is never
worse than before. The public backend surface is unchanged, so routes/router need no
change.
New: sd_cpp_server.py (SdCppServer manager: spawn/readiness/job-submit-poll/cancel/stop,
process spawned inside the drain thread so PR_SET_PDEATHSIG binds to the interpreter, not
a transient thread; empty scratch dir for the server's per-request LoRA/upscaler/embd
scans). Extended: sd_cpp_engine.py (find_sd_server_binary), sd_cpp_args.py
(build_sd_cpp_server_command + build_img_gen_request), sd_cpp_backend.py (server/one-shot
modes, ensure_sd_server_binary upgrades existing sd-cli-only installs), and the prebuilt
installer (locate + chmod sd-server, which ships in the same archive as sd-cli).
Verified on a B200 (Z-Image-Turbo-GGUF, CUDA sd-server): one model load across multiple
generations (server pid stable, a single 'listening on:'), a batch served from one job
with distinct per-image seeds, the second generation faster than the first, and
unload/reload spawning a fresh process. 105 sd.cpp + 81 diffusion tests pass.
Addresses the review of the Phase 16 native-engine PR.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio native diffusion: harden the persistent sd-server path
Addresses review findings on the sd-server backend:
- Router: treat a runnable sd-server as native availability, so an
sd-server-only install (no sd-cli) still routes to the native engine
instead of silently falling back to diffusers.
- Backend: probe the sd-server binary before the multi-GB asset download,
falling back to one-shot sd-cli up front when it cannot run.
- Backend: a lazily cached one-shot fallback engine no longer pins the
backend to one-shot; only an explicitly injected engine does, so a
now-available server can be used on the next load.
- Backend: mask explicit seeds to sd.cpp's signed int64 range before
submitting a server job (large seeds were rejected/wrapped in server
mode only), and split batches above the server's per-job limit into
chunks, each with a timeout proportional to its image count.
- Backend/server: make server startup cancellable. stop() signals an abort
event before taking the lifecycle lock so a blocking readiness wait bails
promptly; unload() stops a not-yet-committed pending server.
- Backend: status() clears stale loaded state when the resident server has
exited, so clients reload instead of hammering a dead process with 500s.
- Server: abandon a poll whose best-effort cancel is not honored within a
grace window (releasing the generate lock), report a pre-submit
stop/cancel as cancellation (409, not 500), and verify JSON responses are
the expected type before indexing.
- Server: use a bounded deque for the stdout tail buffer.
- Add native_mode to DiffusionStatusResponse so the field is not dropped by
the response model.
Adds regression tests for each behavioral change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* sd-server: harden native lifecycle and GPU install path
- Treat a crashed sd-server probe (signal death / non-127 nonzero) as
unavailable so a broken prebuilt falls back to diffusers instead of
routing to a server that dies on startup.
- Drop stale loaded state when a resident server has exited before a
generate, returning the recoverable not-loaded path rather than a 500.
- Reject incomplete server batches (fewer blobs than requested) like the
one-shot path instead of silently dropping images.
- Bound the server log tail in place (keep the deque(maxlen)) and bypass
HTTP(S) proxies for the loopback client (trust_env=False).
- Honor a stop() that arrives after the server is published but before
start() takes the lock, so a cancelled load cannot leak a spawned model
process.
- Map a closed-client RuntimeError during poll to a cancellation when the
generation is being cancelled, so unload races surface as 409 not 500.
- Stop a timed-out server job (best-effort cancel then teardown) so an
abandoned generation cannot keep denoising and block later loads.
- Install the accelerator-matched sd-server build (ROCm/Vulkan/CUDA) and
probe the resident server before auto-installing sd-cli, so a server-only
or GPU host does not fetch the wrong or an unused binary.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* inference: add passthrough tool-call healing core (heal_gate, heal_openai_message, StreamToolCallHealer, nudge helpers)
Small GGUF models often emit tool calls as text (<tool_call>{...}</tool_call>,
Gemma <|tool_call>, <function=> XML) instead of structured tool_calls. Studio's
enable-tools loop already heals these, but the client-tool passthrough
(unsloth run --disable-tools, unsloth start agents) relays them verbatim, so
the agent sees prose and the turn dies.
This module is the shared response-side repair layer the passthrough routes
will call: promote parsed text-form calls to structured calls, but only for
function names the client actually declared; coerce arguments through the same
canonical-key healing as the tool loop; never touch the upstream request body
(llama-server KV/slot reuse stays byte-identical). StreamToolCallHealer is the
streaming buffer-and-repair state machine: prose forwards immediately, only a
partial-signal tail or a suspected tool block is held, false alarms flush
verbatim, and a 64 KiB bound caps memory. nudge_should_retry/nudge_messages
support an opt-in single-retry nudge for non-streaming routes (wired later).
Kill-switch: UNSLOTH_DISABLE_TOOL_CALL_HEALING=1. Reuses
core/tool_healing.parse_tool_calls_from_text, strip_tool_call_markup, and
tool_loop_controller.coerce_tool_arguments unchanged.
* inference: heal text-form tool calls on the OpenAI and Responses passthrough
Wire the passthrough healing core into /v1/chat/completions and /v1/responses,
default ON whenever the request declares client tools:
Non-streaming: heal_openai_message runs inside the existing response-mutation
loop; a promoted call flips finish_reason to tool_calls and nulls the content,
and the verbatim-bytes fast path still applies when nothing was healed.
/v1/responses non-streaming inherits this through openai_chat_completions.
Streaming: a StreamToolCallHealer per stream. Ordinary prose relays
byte-for-byte (a fast path keeps upstream bytes when the healer passes a chunk
through whole); once a tool signal appears, content is held, and at the
finish/[DONE] boundary either synthetic delta.tool_calls chunks replace the
markup (finish_reason rewritten to tool_calls, including the synthetic-finish
path) or a false alarm flushes the held text verbatim. Structured upstream
deltas put the healer to sleep after flushing anything held, so grammar-mode
responses stay byte-identical. The Responses stream feeds healed calls through
the same per-call state machinery as structured deltas (indexes live in a
disjoint range so a healed call can never merge into a structured call's
state), and the visible/reasoning split runs first so reasoning text is never
promoted. parallel_tool_calls=false caps healed calls on every path.
The upstream request body is never touched and healing issues no extra
generation, so llama-server slot/KV-cache reuse is unchanged. Opt-out per
request with auto_heal_tool_calls=false (Responses reads it from the
extra-body); requests without tools relay verbatim.
* inference: heal text-form tool calls on the Anthropic /v1/messages passthrough
Streaming: AnthropicPassthroughEmitter.enable_healing(allowed_tools) routes
content deltas through the shared StreamToolCallHealer. A promoted call closes
any open text block (only the safe prose prefix ever streamed into it), opens a
synthetic tool_use block with a fresh toolu_* id, carries one input_json_delta,
and closes; finish() then forces stop_reason to tool_use unless a truncation
(max_tokens) wins. Structured upstream deltas flush anything held and put the
healer to sleep, so grammar-mode responses are untouched, as is every stream
where enable_healing is never called (Studio's own loop, no-tools requests).
disable_parallel_tool_use caps healed calls too.
Non-streaming: the OpenAI message dict is healed BEFORE block building, so the
existing tool_use promotion loop and stop_reason line treat promoted calls
exactly like native ones (finish_reason length still maps to max_tokens). The
legacy tool-XML strip still runs on remaining text, so opted-out requests keep
today's cleanup behavior byte-for-byte.
auto_heal_tool_calls is now a typed field on AnthropicMessagesRequest
(default True, mirroring Chat Completions) and threads into both passthrough
calls. Healing never touches the upstream request body.
* inference: opt-in single-retry tool-call nudge on the non-streaming passthrough
When the model clearly tried to call a tool (a tool signal in the text) but
healing produced nothing usable, re-ask once: the retry body is the original
body plus an assistant turn (the model's own failed text) and a short user
nudge naming the declared tools. The prompt prefix stays byte-identical, so
llama-server reuses the slot's KV cache and only the two-message suffix is
prefilled. The retry replaces the original response only when it actually
yields a promotable or structured call; on any error or still-garbage output
the original response is returned unchanged. Exactly one retry, non-streaming
OpenAI and Anthropic passthroughs only (a stream has already emitted bytes).
OPT-IN per user decision: nudge_tool_calls=true per request (typed on both
ChatCompletionRequest and AnthropicMessagesRequest, lifted from the Responses
extra-body), or UNSLOTH_TOOL_CALL_NUDGE=1 to flip the process default.
auto_heal_tool_calls=false disables healing AND the nudge.
Also align the non-streaming heal on allow_incomplete=True: the response is
final, so a trailing unclosed tool block is a model failure worth repairing,
matching the enable-tools loop's drain semantics.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* inference: never assume the upstream response shape in the nudge helpers
llama-server error bodies can carry message: null (or no choices at all), and
_last_assistant_text / response_has_promotable_calls / nudge_should_retry
called .get() on the message without a dict check, so a malformed upstream
response raised an AttributeError the surrounding except tuples did not catch,
failing the request instead of degrading to 'nothing to heal'. Route the shape
probing through one _first_choice_message helper that returns None for any
non-dict message, and add a parametrized test over the malformed shapes.
* inference: constrain healing by tool_choice, preserve length finish_reason, keep healed event order in Responses streams
Three review findings on the passthrough healer:
- heal_gate now honors the request's tool_choice: "none" disables healing
outright and a forced function narrows the promotion allowlist to that
one function, so healing can never contradict the request's tool-choice
constraint. Wired through the OpenAI chat (stream and non-stream),
Responses, and Anthropic (converted shape) passthroughs.
- The OpenAI non-streaming heal only upgrades finish_reason "stop" to
"tool_calls"; a truncated generation keeps "length" (the healed call
stays attached) matching the streaming and Anthropic paths.
- The Responses stream emits healer events in order instead of collapsing
all text ahead of the healed calls, so text after a healed call no longer
jumps ahead of the function_call item and output indexes are claimed in
the order the model produced them.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* inference: all-or-nothing promotion when a response mixes declared and undeclared text-form calls
Promoting a subset used to strip ALL tool markup from the content, which
silently deleted the text of any call naming an undeclared tool. The heal
now declines entirely when any parsed call is unpromotable, so the whole
message relays verbatim (pre-PR behavior) and no bytes are ever lost. In
streaming, a declared call that completed before an undeclared one arrived
is already emitted; the late undeclared markup still flushes as raw text.
The nudge helpers mirror the same contract via a shared predicate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* tests: wrap long lines in the Responses healing tests to the project style
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* inference: span-exact healing, disjoint healed stream indexes, per-call Responses message items, allowlisted nudge acceptance
Four review findings on the passthrough healer:
- parse_tool_calls_from_text gains an optional with_spans return so healing
removes EXACTLY the promoted calls' markup. This supersedes the previous
all-or-nothing rule: declared calls promote and every unpromoted byte
(undeclared calls, unparseable closed blocks, suppressed alternate
formats such as a <function=...> block after a JSON call) relays as text.
The stream healer also processes one block per pass, so text between two
healed calls keeps its document position instead of trailing them.
- The OpenAI chat stream shifts native tool-call delta indexes past any
already-emitted healed calls; clients merge deltas by index, so a healed
call and a later native call can no longer merge into one.
- A healed call in the Responses stream closes the open message item and
trailing text opens a fresh one with a later output index, matching the
native stream shape; response.completed snapshots every message item
with its own text.
- The nudge retry only replaces the original response when the retry's
structured call names a DECLARED tool; a hallucinated undeclared call is
not an improvement.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stop the heal path folding trailing prose into a closed function call
parse_tool_calls_from_text(allow_incomplete=True) cut a <function=...> body only
at an end-anchored </function>, so a fully closed call followed by trailing prose
(<function=..>..</parameter></function> words) folded </parameter></function> and
the prose into the tool argument and deleted the prose from visible content. The
strict path (allow_incomplete=False) already cut at the real </function> via rfind.
Do the same in both modes: trim the body at the real </function> when present and
end the removal span there, falling back to the end-anchored strip and body_end
only when the call is genuinely truncated. Add a regression test.
* inference: one shared single-call budget for healed and native calls
Codex round 5: the parallel-call caps counted healed and native calls
separately, so a healed text-form call followed by a native structured
delta double-emitted on all three streaming surfaces when the client
disabled parallel calls.
- OpenAI SSE: once a healed call went out with parallel_tool_calls
false, native tool_call deltas are dropped instead of index-shifted.
- Anthropic emitter: native deltas skip block allocation when the
healed-plus-native count already filled the single slot, and healed
emission counts open native states too.
- Responses stream: native deltas that survived the chunk-level cap are
skipped once a healed call claimed the slot.
Also adds a span assertion for the closed-</function> trailing-prose
parse fixed in the previous commit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: relay undeclared text-form calls as text on Anthropic non-streaming
heal_openai_message promotes only declared text-form tool calls and
span-trims just their markup, deliberately leaving every unpromoted byte
(undeclared text-form calls included) in the content to relay as text.
The Anthropic non-streaming builder then ran a blanket _TOOL_XML_RE strip
over that content unconditionally, deleting the undeclared block before
building the text part, so Anthropic clients silently lost a call the
OpenAI non-streaming path preserves. The strip was harmless when healing
was all-or-nothing but became data loss once healing turned span-exact.
Gate the legacy strip on whether healing promoted a call, matching the
OpenAI passthrough and the intent already stated in the comment above.
Add a route-level regression test for the mixed declared+undeclared case.
* inference: require fully declared nudge retries; keep unpromoted Anthropic text
Codex round 6, two findings:
- response_has_promotable_calls accepted a nudge retry when any one
structured call named a declared tool, so a mixed retry (hallucinated
undeclared call plus a declared one) replaced the original and the
caller forwarded the undeclared call, or with parallel_tool_calls
false could keep only it. All structured retry calls must be declared.
- The Anthropic non-streaming builder still ran the legacy _TOOL_XML_RE
strip after span-exact healing, deleting undeclared or malformed call
text that healing deliberately preserved. The legacy strip now runs
only when healing is off (no declared tools, or opted out), matching
the OpenAI passthrough.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* inference: keep unpromoted Anthropic text whenever healing is active
The previous commit skipped the legacy strip only when a call was
actually promoted, so an undeclared-only (or malformed-only) response
was still silently emptied: exactly the dead-turn shape this path
exists to fix, and inconsistent with the OpenAI passthrough, which
relays those bytes verbatim. Gate the strip on healing being active
instead; opt-out and no-tools requests keep the legacy strip.
* Fix schema-aware tool healing for PR #6801
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix passthrough healing ordering for PR #6801
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix stream finish ordering for PR #6801
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
* Studio: add OpenAI-compatible /v1/images/generations endpoint
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fall back to the loaded base repo for image-generation defaults
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restrict the image-generation 503 to genuine unload races
* Route /v1/images/generations through the active diffusion engine
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Memory planning and dense-quant path: size a local diffusers base's
resident companions from its on-disk VAE and text-encoder weights instead
of folding them to zero, feed the distilled variant hint into the runtime
headroom estimate so turbo and schnell models are not over-reserved, place
group-offload companions resident before attaching the transformer hooks
so a failed placement falls back to whole-module offload instead of
crashing, and bail out of the dense transformer download before it starts
when the requested quant scheme is unsupported so the load falls back to
GGUF cleanly.
sd.cpp stack: scrub the native path lease secret from sd-cli child env,
redact native load-progress errors, forward the resolved accelerator when
auto-installing a forced-native binary, release stale diffusion GPU
ownership on CPU-native loads, and remove the sd.cpp install tree on
uninstall.
Prequant and scripts: reject prequant artifacts missing base_model_id
when a base is requested, expanduser before checkpoint existence checks,
record and validate the int8 exclusion filter and fp8 fast-accum in
checkpoint metadata, make verify_prequant_backend allowlist its local
checkpoint and fail on missing or bad LPIPS and on load-peak regressions,
average only finite PSNR values in diffusion_quality, and reset the
process-wide attention backend between perf probe variants.
API and UI: normalize attention_backend casing before Literal validation,
close hidden popovers when leaving the Images page, and clear the stale
quant label when loading a direct local GGUF file.
Backend:
- Sanitize a blank hf_token to None in begin_load and load_pipeline, so the
default empty Studio token loads anonymously instead of 401ing as an explicit
empty credential.
- Free the ACTIVE diffusion engine before LLM training and in the delete-cached
guard: on a native (sd_cpp) selection the diffusers singleton reports
unloaded, so training could start against a live sd-cli generation and
delete-cached could remove a GGUF the native engine is using. Both now go
through diffusion_engine_router.get_active_diffusion_engine().
- Refuse delete-cached while a background image load is downloading the repo
(or its companion base): status().loaded is False in that window, but the
delete would yank blobs from under the in-flight assembly. Both engines
expose the in-flight ids via a new loading_repo_ids().
- Cap request seeds at 2**53-1: seeds round-trip through JSON gallery recipes,
where JavaScript rounds larger integers, so a restored recipe generated a
different image. Random seeds were already masked to this range.
- Add the task field to CachedModelRepo: the handler sets it for cached
diffusers image repos but response_model silently dropped it, letting
image-only repos pass the chat picker's task gate.
Frontend:
- Offset sequential run seeds by the batch size: the native engine seeds image
j of a run at seed+j, so a +1 run offset regenerated the previous run's
batch-mates.
- Revert the optimistic quant selection when a load fails to start.
- Stop disabling the Images page on chat-only hosts: the native sd.cpp engine
exists exactly for the no-GPU route.
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
---------
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Fix Gemma 4 GGUF OpenAI API streams
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Avoid duplicate Responses stream disconnect watcher
* Keep reasoning-only Responses output hidden
* Address Gemma stream review comments
* Avoid Responses stream task-group cleanup
* Harden OpenAI chat completion streams
* Address OpenAI stream review issues
* Clean up Studio OpenAI stream helpers
* Fix Studio passthrough cold stream timeout
* Fix tool parser compatibility exports lint
* Preserve audio stream disconnect cancellation
* Avoid synthetic finish after passthrough errors
* Address stream cleanup and Gemma parser reviews
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gemma 4: parse bare-string tool args and keep safetensors tools for native <|tool_call>
- Quote bare unquoted string values in Gemma native tool-call args (e.g.
{location:Tokyo,unit:celsius}) so they parse; JSON scalars stay typed.
- Stop _detect_safetensors_features from suppressing supports_tools for
templates that emit Gemma native <|tool_call>, which the shared parser
now reads.
- Add tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden Gemma tool-call parsing and stream-error detection
Address three issues in the Gemma-native tool-call path:
- _quote_gemma_object_keys stopped a bare (unquoted) string value at the
first comma, so an argument like `location:New York, NY` was split
mid-value and the synthesized JSON failed to parse, dropping the whole
tool call. A bare value now ends only at `}` or a comma that begins the
next `key:` pair.
- parse_tool_calls_from_text scanned the entire response for Gemma markers
even inside a tool call already parsed from a `<tool_call>{...}` JSON
block, so a marker-like string inside an argument (data) was promoted to
a second, unintended tool call. Matches inside an already-consumed call
span are now skipped.
- _openai_passthrough_stream relied on _monitor_openai_sse_line to flag a
stream error, which returns early when monitor_id is None
(skip_api_monitor), so an upstream error chunk left saw_stream_error
unset and the synthetic-finish guard emitted a successful finish_reason
after a failed stream. Error chunks are now detected independently of API
monitoring.
Adds tests/test_gemma_tool_parse_edge_cases.py covering the comma and
marker-injection cases.
* Emit the terminal finish_reason chunk in GGUF streams
The OpenAI chat-completions GGUF tool stream and plain stream both built a
final ChatCompletionChunk carrying finish_reason but never yielded it, so
clients received the optional usage chunk and [DONE] with no chunk carrying
finish_reason. OpenAI-compatible consumers rely on that terminal choice to
distinguish stop/length/tool_calls. Yield it before the usage chunk and
[DONE], matching the other streaming paths.
* Parse tool calls in document order and skip nested markers both ways
Unify the JSON- and Gemma-format tool-call passes into a single
position-ordered scan:
- Calls are now emitted in byte order across both formats, so a mixed
output like `<|tool_call>call:create{...}<tool_call|> ... <tool_call>
{"name":"read",...}</tool_call>` executes create before read, matching
the order they appear in (tools run in returned order).
- A candidate that starts inside an already-accepted call's span is
skipped, in both directions: a JSON marker inside a Gemma argument and a
Gemma marker inside a JSON argument are treated as data, not promoted to
a second executable tool call.
Extends tests/test_gemma_tool_parse_edge_cases.py with the ordering and
JSON-in-Gemma nesting cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Quote bare Gemma array elements; order finish before trailing usage
- _quote_gemma_object_keys skipped array values, so a Gemma call with a
bare-string array argument like labels:[bug,ui] produced invalid JSON and
the whole tool call was dropped. Array values are now scanned and bare
string elements quoted, while numbers, quoted strings, and JSON literals
are preserved.
- In the OpenAI passthrough stream, a trailing usage-only chunk
(stream_options.include_usage) that arrived before any finish chunk was
relayed before the synthetic finish, producing usage -> finish -> [DONE].
Emit the synthetic finish before that usage chunk so the order matches the
other streams (finish -> usage -> [DONE]).
Extends tests/test_gemma_tool_parse_edge_cases.py with the bare-array cases.
* Harden Gemma array parsing, XML-parameter guard, and stream teardown
Address five review findings on the Gemma tool-call and OpenAI passthrough
streaming paths:
- parse_tool_calls_from_text collected JSON and Gemma markers without the
_inside_open_parameter guard, so a marker embedded in an existing
<function=...><parameter=...> value was promoted to a separate tool call.
Candidates that start inside an open XML parameter are now skipped, matching
the guard the XML-style parser already applies.
- _quote_gemma_array_elements preserved array elements starting with { or [
verbatim, so an array of objects (items:[{path:a}]) or a nested array failed
json.loads and the whole call was dropped. Object and nested-array elements
are now normalised recursively.
- _openai_passthrough_stream synthesized a finish chunk before a trailing
usage-only chunk and set saw_finish_reason, which made the EOF guard skip the
[DONE] sentinel. The EOF path now emits [DONE] whenever the upstream omitted
it, even after a finish chunk was already synthesized.
- /generate/stream drove generation through asyncio.to_thread with no
disconnect watcher, so a client disconnect during a long generation went
unnoticed until the next send. It now runs _await_disconnect_then_cancel
against the request, matching the other local streaming endpoints.
- _SameTaskStreamingResponse closed the body iterator with aclose() on a
send-side disconnect, raising GeneratorExit so the generators' cancellation
handlers (which finish the api_monitor entry) never ran. It now throws
CancelledError, falling back to aclose() when athrow is unavailable.
Extends tests/test_gemma_tool_parse_edge_cases.py with array-of-objects,
nested-array, and marker-inside-XML-parameter cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Watch disconnects on Anthropic streams; keep timestamps in Gemma values
Two follow-ups on the streaming and tool-parse paths:
- _anthropic_tool_stream and _anthropic_plain_stream drove generation through
asyncio.to_thread(next, gen, ...) and only polled is_disconnected() between
events, so a client disconnect during prefill or a long generation/tool step
held the decode slot until the next event or a failed send. Both now run the
_await_disconnect_then_cancel watcher used by the other local streams, stop it
in finally, and break promptly when cancel_event is set.
- _GEMMA_NEXT_KEY_RE treated any comma followed by word-chars-then-colon as the
next key, so a bare value such as "meet at 10:00, 11:00 tomorrow" was split
into bogus keys. The next-key token must now be identifier-shaped (start with
a letter or underscore), so a comma before a timestamp, ratio, or other
numeric-then-colon text stays part of the value.
Adds a timestamp-in-bare-value regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard nested markers, reset on disconnect, clean unstarted streams
Three follow-ups on the tool-parse and streaming paths:
- parse_tool_calls_from_text only skipped markers that fell inside a span it
had already parsed successfully, so when an unquoted Gemma argument contained
a literal marker (code:<|tool_call>call:terminal{...}<tool_call|>) the outer
object failed to normalize, its span was never recorded, and the inner marker
was promoted to a standalone terminal call. Candidates nested inside any other
candidate's brace span are now skipped regardless of whether the enclosing
candidate parsed, so a marker in malformed outer data is never executed.
- /generate/stream skipped backend.reset_generation_state() when the disconnect
watcher set cancel_event between chunks: the loop broke and the finally's reset
is guarded on cancel_event being unset. A subprocess backend kept decoding
after the client left. The cancel-break path now resets the backend.
- _SameTaskStreamingResponse threw CancelledError / called aclose() on the body
iterator on a send-side disconnect, but neither runs the try/finally of a
generator that never started (early disconnect on http.response.start), so the
passthrough's eagerly-opened upstream httpx stream and cancel-registry entry
leaked. It now tracks whether the body started and, when it did not, runs an
optional unstarted_cleanup hook; the OpenAI passthrough wires it to close the
upstream resp/client and exit the cancel tracker.
Adds a nested-unquoted-marker regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Recognise the Gemma 4 separate-drafter MTP family, auto-download the drafter with retry, fall back to n-gram with a clear reason when it cannot be resolved, and retry the download on reload. Gemma 3n (ships no drafter) and embedded-MTP models (Qwen) are unaffected.
Fixes#6406
* Studio: Auto disables MTP for MLA models (GLM-5.2 et al.); UNSLOTH_MLA_MTP_ENABLED to re-enable
Studio's Auto speculative mode promotes any embedded-MTP model >=3B to
--spec-type draft-mtp. For MLA models (GLM-5.2/DeepSeek/Kimi) that is a
regression: llama.cpp's MLA/DSA MTP path keeps a duplicated full target-KV
context and recomputes the sparse-attention indexer every draft step, so it
runs ~2x slower than no speculation (GLM-5.2 UD-IQ1_S bench: 27 vs 45 tok/s,
flat across draft depth 1..6 and 96-100% acceptance, on both prose and code).
vLLM/SGLang get a speedup from the same model, so this is a llama.cpp
implementation gap, not a model property.
Auto now drops embedded MTP for MLA models and falls back to ngram-mod (or
spec-off when the binary lacks ngram-mod), mirroring the existing sub-3B
fallback. The metadata separator is kv_lora_rank: it is present on MLA models
and absent on non-MLA embedded-MTP models (Qwen3.x-MTP), whose MTP module is
structurally identical but fast, so a "full layer" heuristic cannot tell them
apart. Qwen MTP, separate drafters (Gemma, --model-draft), and non-MTP models
are unchanged.
Explicit overrides still engage the slower MTP route: choosing MTP / MTP+Ngram
in Settings, or passing --spec-type in extra args. UNSLOTH_MLA_MTP_ENABLED=1
re-enables Auto promotion for MLA once the upstream path is optimized.
A new spec_fallback_reason value "mla_mtp_disabled" surfaces this as an
Auto-mode policy downgrade (not a binary/update problem), with a settings
banner that points users at the MTP override. It is deliberately kept out of
the "Update llama.cpp" affordance since updating does not help.
Tests: resolver-matrix rows for MLA->ngram-mod / MLA-no-ngram->off /
non-MLA-Qwen->draft-mtp / MLA-separate-drafter->draft-mtp /
non-MTP-MLA->default / forced mtp|mtp+ngram on MLA->draft-mtp / env flag;
kv_lora_rank metadata fixtures; and reload-skip coverage (Auto ngram-mod is
idempotent, forced mtp bounces a reload).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: fix Bypass Permissions menu freeze and show decimal GB for model sizes
Bypass Permissions freeze: the warning dialog lived inside the composer
"+"/More dropdown and kept the menu mounted via onSelect preventDefault,
so confirming or cancelling the dialog left both popovers frozen open.
Lift the dialog out of the menu into a store-driven
BypassPermissionsConfirmDialog mounted at a stable spot in the composer.
The menu item now closes normally on select and just toggles a new
bypassConfirmOpen store flag, so the popovers dismiss as expected.
Model search sizes: formatBytes divided bytes by 1024 but labelled the
result "GB", so unsloth/GLM-5.2-GGUF:UD-IQ1_S showed 201.8 GB where
Hugging Face reports 217 GB. Switch the search display to decimal
(base-1000) units to match what Hugging Face reports. The GPU-fit math
stays base-1024 since VRAM capacity is binary.
* Studio: address review feedback and add GLM-5.2 high/max/disabled thinking
Review feedback on the Bypass Permissions and size-format changes:
- Mount the Bypass Permissions warning dialog once at the chat-page root
instead of inside each Composer. It is driven by global store state, so
the per-composer mount meant Compare mode (multiple composers) rendered
duplicate dialogs and the shared-composer menu had none. A single root
mount fixes both.
- Defer opening the dialog past Radix's menu-close focus restoration with
setTimeout(0), so the dropdown does not steal focus back and break the
dialog's focus trap.
- Clamp the unit index in formatBytes so units[i] cannot go out of bounds
past TB (and to absorb log() float error at exact powers of 1000).
GLM-5.2 reasoning levels:
GLM-5.2's template gates thinking with enable_thinking and also reads a
reasoning_effort level ('high' or 'max'), so it needs high / max /
disabled rather than the binary toggle it got before (its style was
detected as enable_thinking, which made 'high' unreachable). Add a new
reasoning style 'enable_thinking_effort' that reuses the effort dropdown
but, unlike gpt-oss, can be fully disabled:
- detect_reasoning_flags classifies a template that has both
enable_thinking and reasoning_effort, extracting the discrete levels
from the quoted effort literals it branches on. Templates with only one
of the two (gpt-oss, Qwen3, DeepSeek, GLM-4.6) are unchanged.
- _request_reasoning_kwargs maps the new style to enable_thinking plus an
in-range reasoning_effort; disabling sends enable_thinking=false. The
gpt-oss reasoning_effort path is left untouched.
- The backend reports reasoning_effort_levels on the load/status response;
the frontend carries them through to the effort dropdown and sends
enable_thinking + reasoning_effort for this style.
Verified: backend reasoning kwargs render the real GLM-5.2 template to
"Reasoning Effort: High/Max" (thinking) and an empty <think></think>
(disabled); tsc, eslint, i18n parity and the production build all pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address review feedback on reasoning effort and formatBytes
- chat-adapter localReasoningEffort: accept 'minimal' so a template that
branches on it (extracted into reasoning_effort_levels) is sent through
instead of being coerced to 'low' and then dropped by the backend.
- formatBytes: return '0 B' for non-finite / non-positive sizes (missing
metadata -> NaN, Infinity, negatives) and clamp the unit index lower
bound to 0, so sub-1-byte values can't produce a negative index.
* Studio: hybrid reasoning none gate and decimal GB in load progress
- _request_reasoning_kwargs: for enable_thinking_effort models, treat a
raw reasoning_effort='none' (OpenAI 'no reasoning' sentinel) as the
enable_thinking=false off gate, so a direct API caller can disable
thinking even without passing enable_thinking. The frontend already
sends enable_thinking=false; this only affects raw API callers.
- use-chat-model-runtime: the download / 'X of Y GB in memory' load
progress divided bytes by 1024**3 but labelled GB, so it disagreed with
the model picker and Hugging Face. Use decimal GB (1e9) to match.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: carry hybrid reasoning levels on all load paths and harden formatBytes
Review follow-ups on the enable_thinking_effort work:
- Every model-load path now copies reasoning_effort_levels and derives
supportsReasoningOff, via a shared reasoningCapsFromLoad() helper. The
shared/Compare composer load and the three chat-adapter auto-load paths
previously set only reasoningStyle, so a GLM-style hybrid model loaded
through Compare or first-chat auto-load fell back to the default
low|medium|high and lost its Max / Off controls.
- The local send path clamps the effort to the loaded model's advertised
levels (clampReasoningEffortToLevels) instead of a hard-coded list. A
stale "max" carried over from an external provider no longer reaches a
pure reasoning_effort (gpt-oss) model that only accepts none|low|medium|
high, where the backend would have dropped it.
- formatBytes divides iteratively instead of via Math.log, which has float
error at exact powers of 1000 (log(1e12)/log(1000) = 3.9999... would
label 1 TB as "1000 GB"). Keeps the non-finite/non-positive guard.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: free chat model VRAM at training start only when the GPU is tight
The training start route unconditionally tore down the transformers/MLX
inference subprocess before training, and never stopped the llama.cpp GGUF
server at all, so a loaded GGUF chat model kept holding VRAM for the whole
run. Conversely the HF model was always unloaded even when there was plenty
of room to keep it.
Make the unload VRAM aware and cover every inference backend:
- Add routes/training_vram.py with summarize_resident_chat(),
can_keep_chat_during_training() and free_chat_models_for_training(). The
keep/unload decision reuses the same estimator and live per device free
VRAM reader the training GPU selection already uses (auto_select_gpu_ids,
estimate_required_model_memory_gb, get_visible_gpu_utilization), so the
probe agrees with the placement computed later in start_training.
- When a chat model is resident and training fits alongside it with a
conservative margin (required_gb * 1.15 + 4 GB), keep it loaded so the
user can train and chat at the same time; on a multi GPU box training
lands on a different GPU and both coexist. Otherwise unload the HF/MLX
orchestrator and the llama.cpp GGUF server before training starts.
- The export subprocess shutdown stays unconditional and now runs first so
its freed VRAM is reflected in the decision.
Default deny: non CUDA backends, unestimable models, or any probe error
fall back to the previous always unload behavior.
Adds tests/test_training_vram_coexistence.py and updates two existing route
tests in test_gpu_selection.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: per-GPU floor for explicit GPU lists + don't unload chat on invalid gpu_ids
Address review feedback on the chat coexistence probe:
- Explicit gpu_ids mode now enforces a per-GPU floor in addition to the
aggregate free-VRAM check, mirroring auto_select_gpu_ids' min_per_gpu_N.
Without it, an uneven split such as free [45, 10] for a 40 GB job passed
the aggregate threshold and kept chat loaded even though the 10 GB GPU
could not hold its training shard, risking an OOM.
- Invalid explicit gpu_ids (ids outside the visible set, or a UUID/MIG
mask) make resolve_requested_gpu_ids raise. That request is rejected with
a 400 before training starts, so leave the resident chat model untouched
instead of unloading it.
- Tighten the target_modules / gpu_ids type hints to List[str] / List[int].
Adds tests for the per-GPU floor (uneven split unloads, even split keeps)
and for invalid gpu_ids keeping the chat model loaded.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: only free chat VRAM once training will start; handle in-flight and CPU-only chat
Address the second review pass on the chat-coexistence path:
- Run the chat/export VRAM teardown as a before_spawn hook inside
TrainingBackend.start_training, fired only after the start guards pass.
Previously the route freed chat VRAM before calling start_training, so a
refused start (e.g. a lingering pump thread) would tear down the resident
chat model even though no training job began.
- Treat an in-flight HF chat load (loading_models set, no active model yet)
as not safely sizeable: free it rather than risk both OOMing as the load
keeps allocating after training starts.
- Do not count or tear down a GGUF llama-server confirmed to run entirely on
CPU (_gpu_offload_active is False): it holds no VRAM, so killing it cannot
help training fit.
Adds tests for the before_spawn hook (runs on start, skipped when a
subprocess is alive or a pump thread will not die, survives a hook error),
the in-flight load flag, and the CPU-only GGUF exclusion in both the resident
summary and the unload path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: treat any in-flight chat load (HF swap / mid-start GGUF) as unsafe to keep
Tighten the in-flight detection in summarize_resident_chat so the keep check
never sizes a load that is still allocating:
- Flag loading on ANY non-empty loading_models, not only when active_model_name
is empty. load_model adds the new model to loading_models before clearing the
old active_model_name, so a replacement load during a swap was previously
sized as a normal resident and could OOM as the new model finishes loading.
- Flag a GGUF server that is active but not yet healthy (is_loaded False) as
in-flight: it is still mmaping/offloading layers, so its final VRAM footprint
is unknown.
Consolidates the signal into a single resident["loading"] flag; the route frees
the chat model whenever it is set. Adds tests for the replacement HF load and
the mid-start GGUF cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten comments in chat/training VRAM coexistence (comments only)
* Studio: run before_spawn VRAM hook only after GPU-selection validation
Reviewers found the before_spawn hook fired before prepare_gpu_selection
validated gpu_ids (and before config build), so a refused start (invalid
gpu_ids -> 400, or a bad grad-clip value) could still tear down chat/export
VRAM. Move the hook to immediately before proc.start(), once all synchronous
validation and process construction have passed. This also fixes the route's
in-flight-chat loading branch, since that teardown runs inside the same hook.
Add test_hook_skipped_when_gpu_selection_rejects.
* Studio: recompute GPU auto-selection after the before_spawn VRAM hook
Codex P2: with before_spawn moved after prepare_gpu_selection, placement was
frozen against the pre-teardown VRAM state while the hook freed export/chat
afterward. Auto-selection could pin training onto a GPU the hook then cleared
(or onto a kept chat model). Split validation from placement: explicit gpu_ids
are still validated before the hook (raise -> 400, no teardown; explicit
placement is VRAM-independent), but VRAM-dependent auto-selection now runs
after the hook so it sees the freed memory.
Add test_auto_placement_runs_after_hook and test_explicit_placement_validated_before_hook.
* Studio: allow chatting during training (lift sidebar gate + VRAM-aware load guard) (#6335)
* Studio: allow chatting during training (lift sidebar gate + VRAM-aware load guard)
The sidebar disabled New Chat, project, and home navigation while a training
run was active, so users could not chat during training even though the backend
serves inference fine alongside a run. This removes that gate and adds a backend
guard so the one genuinely risky operation, loading a new local chat model
mid-training, is refused with a clear 409 when it would not fit beside the run.
Frontend (app-sidebar.tsx): drop the chatDisabled = isTrainingRunning gate and
its consumers. Navigation triggers no model load on its own, so chat stays
usable during training.
Backend (routes/training_vram.py, routes/inference.py): add
can_load_chat_during_training plus a load/validate guard that sizes the same
effective load the loader performs (LoRA 4-bit to 16-bit resolved first, HF auto
placement via auto_select_gpu_ids, explicit multi-GPU per-GPU floor, GGUF sized
from on-disk shards and companions or the selected remote variant). It is a
no-op when training is inactive, never blocks external providers or
already-resident models, and default-denies only on a CUDA sizing failure so a
load can never OOM the run. Validate refuses early with the real settings so the
frontend does not unload the resident chat model for a load that would be
rejected.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address review feedback for chat-during-training load guard
- Run the load/validate VRAM guard via asyncio.to_thread so the sync
nvidia-smi + HF metadata work never blocks the event loop.
- Size the GGUF KV cache at the requested context (_estimate_gguf_kv_gb)
and add it to the local GGUF estimate so large-context picks are not
under-counted.
- Keep the requested quantization when adapter_config.json is malformed
(not a JSON object) instead of raising in _effective_load_in_4bit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: size the training load guard at the launcher's effective GGUF context
The GGUF KV-cache estimate used max_seq_length only, but the llama.cpp
launcher honors a user --ctx-size/-c in llama_extra_args. A load such as
max_seq_length=4096 with --ctx-size 131072 was sized against a 4k cache
while the server allocates 131k, so the guard could approve a long-context
GGUF load that then OOMs training. Size the guard's KV at the larger of
max_seq_length and the parsed --ctx-size (reusing the launcher's own
parse_ctx_override), keeping the conservative f16 cache so the estimate is
never smaller than what the server allocates.
The chat model picker also validated with the raw max_seq_length while
/load sizes with resolveLoadMaxSeqLength, so validate could pass, unload
the current model, then have /load reject the native-context load. Validate
now uses the same effective context; the load path is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: size the GGUF training guard at the server parallel-slot count
The KV-cache estimate assumed a single slot, but llama-server allocates the
cache across --parallel slots (app.state.llama_parallel_slots). On a Studio
launched with --parallel N>1 the guard under-sized the cache N-fold and could
approve a GGUF chat load that then OOMs training. Thread the same slot count
the loader uses into the guard's KV estimate; default 1 leaves single-slot
setups unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim comments for chat-during-training guard
* Studio: keep chat generation alive across navigation; Train spinner + Return to Chat
Hoist the base chat runtime above the routed outlet so navigating to Train (or any tab) no longer aborts an in-flight generation; only an explicit Stop cancels. Add a Train sidebar spinner and swap New Chat to Return to Chat while a run is active, with a lightweight completion watch so the spinner clears from any tab. Also respawn a chat llama-server killed mid-session and guard unreadable HF cache dirs that 500'd the hub model list.
* Studio: show Return to Chat on the Train tab whenever a chat is live
Previously the top sidebar item only swapped to Return to Chat while training was running; on the Train tab with an idle/just-finished run it stayed New Chat, which started a fresh thread and cancelled an in-flight generation. Show Return to Chat (and navigate back, preserving the run) whenever a generation is running or its thread is still active, or training is in progress.
* Studio: keep a running chat alive when starting a New Chat
Starting a New Chat (or switching threads) while a generation was in flight
remounted the single-chat runtime provider, which detached the in-flight run
and cut the previous chat off (it showed up frozen / empty when reopened).
Key the single-chat view by project instead of by thread or new-chat nonce so
the provider stays mounted and assistant-ui switches to a fresh thread in place.
The previous generation keeps streaming in the background and autosaves on
completion, and returning to that thread reattaches the live run instead of
reloading a half-saved one.
Also:
- "Return to Chat" now lands on the thread that is still generating rather than
the empty new chat that became active after New Chat.
- Skip the explicit /inference/cancel POST when an abort comes from a runtime
detach (navigation / background switch) rather than an explicit Stop, so a
backgrounded generation is never cancelled behind the scenes.
* Studio: make model export non-blocking and inline
The Export tab opened a full-screen modal that trapped focus, could not be
closed or cancelled while running, and showed no progress. It also stopped
training and unloaded the chat model before loading, so export could not run
alongside them.
Export now mirrors the training runtime pattern:
- Inline panel embedded where the Export Model button was, with no modal or
backdrop, so the rest of the UI stays usable during an export.
- Global export runtime store plus an app-root lifecycle hook, so a run keeps
going and streaming across navigation and is reflected on the Export nav item
from any tab.
- The worker log stream now stays connected across the load to export phase
boundary instead of stranding on "Waiting for worker output".
- Progress bar driven by phase and quant index (quant N of M for GGUF), with
elapsed time and a working Cancel.
- load-checkpoint no longer stops training or unloads inference; export loads in
its own subprocess in parallel and surfaces out-of-memory as a clear error.
- Add POST /api/export/cancel and is_export_active on /api/export/status.
* Studio: show Return to Chat on the Export tab too
Extend the New Chat to Return to Chat swap to the Export route so leaving a
running chat for Export offers a way back to the live generation, matching the
Train tab.
* Studio: smooth out Export animations and polish the panel
- Drop the height-based reveal animations (source switch, run panel, quant
picker, hub fields) that caused flashing and reflow; use instant swaps and
quick opacity fades instead.
- Method and quant cards now transition colors only, with no transition-all or
hover lift, so selecting a method or quant is crisp instead of jumpy.
- Auto-scroll the export panel into view when it opens and add a scroll-to-bottom
button when its output is below the fold, like Chat.
- Show Return to Chat on the Export tab while an export is running, matching how
training drives it on the Train tab.
- Surface the current phase or stage in the live output before the first worker
line arrives so the panel never looks stuck while progress is advancing.
* Studio: show Return to Chat on every non-chat tab
Generalize the Return to Chat swap from just Train/Export to any non-chat route
(Recipes, Projects, Hub, ...) so a running or active chat is always one click
away, instead of showing New Chat there.
* Studio: stream export logs over the Cloudflare tunnel; drop janky export animations
Exporting over a --secure Cloudflare quick tunnel showed "connecting..." with no
logs while the progress bar advanced. Cloudflare buffers text/event-stream and
only flushes when the stream closes, so the SSE log stream never reached the
browser during the run (direct localhost is unaffected, which is why this only
showed up over the tunnel).
Add a tunnel-safe JSON poll fallback (GET /api/export/logs?since=) that the
runtime lifecycle hook polls while a run is active. Short JSON responses are not
buffered by the proxy, so logs show up in near real time over the tunnel. It
shares the orchestrator's monotonic seq cursor with the SSE stream and the store
de-dupes by seq, so the two transports run together (SSE on localhost, poll over
the tunnel) without double-printing. A successful poll marks the panel
"streaming" instead of leaving it stuck on "connecting...".
Also remove the framer-motion AnimatePresence reveals from the export config and
run panel (quant picker, hub fields, the inline run panel, and the live log
section). The expand/slide animations flashed and felt clunky; the sections now
render in place.
* Studio: recover export over the Cloudflare tunnel when the blocking POST times out (524)
A model export over a --secure Cloudflare quick tunnel showed "Request failed
(524)" even though the export succeeded on the backend (the GGUF was written).
Cloudflare returns 524 when a single request takes longer than ~100s to respond,
and a GGUF conversion routinely runs for minutes, so the blocking per-method
export POST is cut off while the backend keeps going.
Confirm completion via short status polls instead of relying on the long POST
response (the same approach that fixed log streaming):
- The orchestrator records each finished op's outcome (status / output_path /
error) with a monotonic seq, exposed on GET /api/export/status.
- parseJson now preserves the HTTP status; a 524/520/522/523/502/503 or a
status-less network drop is classified as a recoverable transport error.
- runExport wraps each phase (load, every export method, each GGUF quant): on a
recoverable failure it keeps the run alive (logs keep streaming, the panel
shows "reconnecting...") and polls status until the still-running op finishes,
then settles from the recorded result, recovering the output path for the
success banner. A real 4xx still fails immediately; localhost still uses the
fast POST response. applyBackendStatus also settles a reloaded run from the
last-op record.
Verified over the tunnel: a 3m14s gemma-4-E4B-it GGUF export now ends on the
success banner with the output path instead of 524.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep the export method + logs visible after navigating away mid-export
While an export was running, navigating to another tab and back to Export
remounted the page and reset the local form state (exportMethod, quant levels),
so the method card showed unselected and the run panel's log area was hidden
until the card was re-clicked. The run itself lives in the global store and was
unaffected.
Seed exportMethod / quantLevels from the active run's summary via lazy useState
initializers on (re)mount, and gate the panel's log area on the live run
(isExporting / logLines / the run's method) rather than only the local form
selection. The card stays selected and the logs/progress stay visible across
navigation; nothing changes when no run is active.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: address export/training review findings
- Export: guard Start against an empty GGUF quant selection so an inline-panel
run with no quant can't settle as success with no file produced.
- Export: thread the source HF token into the background load so gated/private
HF source exports (and gated bases) authenticate, matching the consent path.
- Export: only settle a recovered (non-owned) run as a finished export when the
last backend op was an export, not a standalone load_checkpoint.
- Training: free the export subprocess whenever an export is active, not only
once a checkpoint is loaded, so an in-flight export load can't race training
for VRAM (current_checkpoint is unset during the load phase).
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Harden model fetching: consent gate for trust_remote_code
Add a load-path consent gate that scans a model's auto_map repository code
before it executes and blocks CRITICAL/HIGH findings unless the user pins
approval of that exact code version. Capability detection stays code-free,
reading raw config.json instead of AutoConfig.
- Scan config.json and tokenizer_config.json auto_map, nested local helpers,
and external owner/name--module repos; fail closed on partial downloads.
- Gate inference, training, and export workers, including the MLX path and a
LoRA's base model, and report requires_trust_remote_code from the raw config
so chat and auto-load surface the dialog.
- Verify trusted-org auto-enable against the Hub with the request token and key
the verdict cache by token; reject local-path and spoofed names.
- Add a consent dialog showing the flagged file, line, and surrounding code.
- Thread hf_token through the scan and load paths for gated repos.
* Address review: token handling, tokenizer/LoRA scan coverage, rollback
- Send the HF token for remote-code scans in the POST body, not the URL, so it
never lands in a log or browser history.
- Collect tokenizer_config.json auto_map files directly instead of relying only
on the repo file listing.
- Resolve a LoRA's base model for the validate flag and the scan endpoint so the
dialog scans the code the workers actually gate.
- Pass the request token to the training YAML trusted-org auto-enable.
- Resend a previously approved fingerprint when rolling back to a custom-code
model after a failed switch.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Consent UX: drop legacy chat toggle, fix decline copy, purge declined downloads
The per-model consent dialog is now the single approval path for custom
(auto_map) code in chat, so three leftovers from before it existed are removed:
- Remove the "Enable custom code" switch from Chat Settings and stop persisting
trust_remote_code, so a previously saved blanket-on cannot linger and load a
model without going through per-version review. The flag stays as an internal
YAML/preset default (e.g. first-party auto-enable); the load path still gates
every custom-code load on a fingerprint only the dialog produces.
- Reword the decline message and the auto-load toast to describe approving the
model's code from the dialog, not a missing settings toggle.
- On decline, purge the repo the scan downloaded so untrusted code is not left
on disk. A new /api/models/discard-remote-code endpoint deletes only a
metadata-only cache entry the scan created; it refuses local paths, loaded
models, and any repo with weight files cached, so a model the user already had
or pre-downloaded is always left untouched. The frontend only calls it when
the scan reported created_by_scan.
Adds discard-endpoint tests (delete metadata-only, refuse on weights/gguf,
refuse local, no-op when not cached) and a created_by_scan payload assertion.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Export: remove the user-facing trust remote code toggle
The Export page kept a "Trust remote code" switch (default on) next to the HF
token field. Like chat, custom (auto_map) code should be approved per model
through the load-time review dialog, not a persistent blanket switch, so the
toggle is removed. The export load path already routes through the same consent
dialog: an HF source now starts with trust_remote_code off and only enables it
when the user approves the scanned code in the dialog (a local checkpoint the
user exported stays trusted by default). With the dialog unreachable and no
approval, an HF source loads with trust_remote_code off, which fails closed
rather than running unreviewed code.
* Block loads of repos with unsafe files using Hugging Face's security scan
The trust_remote_code consent gate covers one load-time RCE vector (a repo's
auto_map Python). It does not cover the other: a malicious pickle inside a weight
file (pytorch_model.bin, *.pkl, *.dat) deserializes during from_pretrained even
with trust_remote_code False, so a repo with a normal config plus a poisoned
pickle slips past the existing gate.
Add a metadata-only malware gate that uses Hugging Face's own scan (picklescan +
ClamAV), read via model_info(securityStatus=True).security_repo_status. It never
downloads, opens, or unpickles the flagged files; it only reads the Hub's verdict
and surfaces the flagged file names. New evaluate_file_security runs
unconditionally (independent of trust_remote_code) in every load path (inference,
training SFT/MLX, export), blocking the load when a file is flagged
unsafe/suspicious/malicious. The /remote-code-scan preflight and the validate
endpoint also report the result so the consent dialog opens as a hard block (no
override) listing the flagged files, even for a repo with no custom code.
Policy: hard block with no user override; fail open when the scan is unavailable
(offline/unscanned) so legitimate loads are not broken; no first-party exemption
(a poisoned pickle in a compromised trusted repo still blocks); local paths and
GGUF are skipped (no Hub scan, non-pickle format). Blocking does not gate on
scansDone, since that is often false for clean repos and a file already flagged
unsafe is unsafe regardless.
Adds test_file_security.py covering the block/allow/fail-open/skip matrix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: scan list-form tokenizer auto_map, gate unsafe files on all load paths
Fixes from a 10-reviewer pass on the model-fetching hardening:
- The remote-code scanner skipped tokenizer auto_map encoded as a [slow, fast]
list (transformers' standard tokenizer shape, e.g.
{"AutoTokenizer": ["owner/repo--tokenization_x.Slow", null]}). External
tokenizer code in that form was never fetched, scanned, or fingerprinted, so an
AutoTokenizer(trust_remote_code=True) load could run it. _auto_map_refs now
flattens string, list, and nested values. Adds a regression test.
- Compare-mode chat loads and background auto-load only gated on
requires_trust_remote_code, so a repo flagged unsafe by the Hub scan but with no
custom code skipped the hard-block dialog. Both now also gate on
requires_security_review, matching the main chat path.
- The /remote-code-scan and /validate routes collapsed a LoRA adapter to its base
before the malware scan, so unsafe files in the adapter repo itself were missed
in the pre-load review (the workers already scan both). Both routes now run the
file-security scan over the adapter and the base.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Require approval for all HIGH remote code, fail closed when unscannable
Tighten the load-time security gates based on review:
Consent gate
- HIGH-severity auto_map code now requires explicit, per-version approval for
every repo, including first-party unsloth/nvidia. The org is no longer a
blanket bypass: a compromised first-party repo with HIGH code still warrants
review. CRITICAL stays a hard block; clean code still loads after the consent
prompt.
- Fail closed when auto_map code is present but cannot be fully fetched or
listed to scan (gated, offline, transient, or a repo-listing failure that
could hide an imported helper). We cannot fingerprint code we cannot see, so
this is a non-approvable block, retryable once the repo is reachable.
- Scan auto_map from every config that can carry one (model, tokenizer, image
and feature processor, processor, video processor), not just config.json and
tokenizer_config.json, so a custom-processor model is not missed. The file
list is the single source of truth in remote_code_scan and is pinned to the
transformers filename constants by a guard test.
- Distinguish a genuine 404 (config truly absent) from a transient error: only
the latter forces a scan, so a repo with no config is correctly a no-op.
Malware gate
- Scan a remote repo even when its name ends in .gguf; only local paths skip the
Hub scan, so a repo cannot dodge the scan by naming itself "*.gguf".
- Correct the docstring: a file already flagged unsafe blocks regardless of
scansDone; the only fail-open path is an unavailable scan.
Coverage
- Resolve a remote LoRA adapter's base model (not just local directories) so the
base, where the code and weights actually execute, is scanned in validate,
the scan route, and the training and export workers.
- Gate the embedding training path (FastSentenceTransformer) with the malware
and consent checks, matching the other load paths.
Tests updated and added for each change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope malware gate to the load-path vector; stop false-blocking first-party models
Follow-up hardening from a second review pass + a broad live model matrix
(unsloth/* , nvidia/* , third-party, and the eicar malware repo).
Malware / unsafe-file gate
- Scope the block to the actual RCE vector: a root-level file in a code-executing
format. from_pretrained deserializes weight files at the repo ROOT, so a flag is
only a load-path pickle vector there. Two exclusions, because neither is loaded:
inert formats (safetensors is tensor-only, gguf is non-pickle, configs/text/
images) and files in subdirectories. This keeps eicar blocked (its *.pkl/*.dat/
eicar_test_file sit at the repo root) while no longer false-blocking legitimate
first-party repos: nvidia/Nemotron-H-8B-Base-8K ships root safetensors plus NeMo
pickle checkpoints under nemo/ that the loader never touches, and the Hub flags
both; the gate previously hard-blocked it.
- Unknown / future non-"safe" levels now fail closed (block) instead of being
silently allowed, so Hub schema drift cannot introduce a bypass; in-progress
("pending"/"scanning"/"error") levels stay non-blocking to avoid false blocks.
Consent gate
- Ignore a STALE own-repo auto_map target that is absent from the repo listing (an
older config pointing at a file the repo no longer ships) instead of failing the
whole repo closed as unscannable. The present .py are still fully scanned, which
is the stronger coverage, and a file that is not there cannot execute. This
unblocks first-party models like unsloth/PaddleOCR-VL (its tokenizer_config.json
names processing_ppocrvl.py while the repo ships processing_paddleocr_vl.py). A
referenced .py that IS present but cannot be fetched, and a repo-listing failure,
still fail closed.
Remote LoRA base resolution
- Distinguish a genuine 404 (not a LoRA / repo absent -> None) from a transient
error: the transient case is retried once, then logged as a WARNING (a missed
base is scanned by neither gate) rather than silently skipped.
Discard endpoint
- Treat .onnx and .ckpt as weights so a repo whose only heavy artifact is one of
those is never eligible for the declined-download purge.
Tests added for each: load-path scoping (safetensors/subdir/Nemotron-H shapes,
unknown-level fail-closed, pending non-block), stale own-repo auto_map ref, remote
LoRA transient retry, and the empty-config-list (all-404 -> []) semantics.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make LoRA-base transient-warning test robust to logging backend
Assert on the logger object directly instead of capsys, so the test does not
depend on whether the real structlog logger or the module-stub logger is active
(which varies with test collection order).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Allow a repo with auto_map but no executable code (e.g. GGUF) instead of blocking
A config can declare an auto_map yet the repo ship NO executable .py -- most
commonly a GGUF repo whose config.json carries an auto_map copied from the original
model (e.g. unsloth/Llama-3_1-Nemotron-Ultra-253B-v1-GGUF references
modeling_decilm.py, which the GGUF-only repo does not contain). A GGUF model loads
through llama.cpp, which never executes auto_map, and transformers cannot run a file
that is not present, so there is nothing to scan and trust_remote_code is a no-op.
The fail-closed change treated this empty result the same as "code is present but we
could not fetch it" and hard-blocked the load. Distinguish the two: repo_remote_code_files
now RAISES RemoteCodeUnscannable when code is present but cannot be fully fetched or
listed (offline / gated / transient / a present .py that 404s / a listing failure),
and returns an empty dict only when the listing succeeded and the repo genuinely ships
no executable .py. The consent gate blocks on the exception (fail closed) and allows the
empty case as a no-op. Real unscannable code still hard-blocks; eicar and CRITICAL/HIGH
custom code are unaffected.
Verified against all 37 unsloth/*Nemotron* models (two GGUF repos were false-blocked,
now load) and the existing matrix (eicar still blocks; DeepSeek-OCR / NVLM-D-72B still
prompt approvable consent). Tests updated to expect the raise for unscannable cases and
added for the no-executable-code no-op.
* Ignore vestigial auto_map in GGUF repos (llama.cpp never runs it)
A GGUF repo's config.json is often copied verbatim from the original
transformers model, auto_map and all, but a GGUF load goes through
llama.cpp which never executes auto_map, so the config is inert. Treat
a direct .gguf reference, and a repo that ships .gguf weights with no
.safetensors, as having no remote code so the consent flow is never
triggered. A mixed repo with both .gguf and .safetensors is still gated,
since the safetensors variant would load through transformers where
auto_map does run. The check sits behind the existing auto_map-present
gate so normal models pay no extra repo listing.
* Add scanner-result copy to the remote-code consent dialog
Make the consent dialog state the scan outcome in plain language for
every model. When the static scan finds nothing, reassure the user with
'Our automatic scanner did not flag any worrying files, but please
double check.' (shown only for the clean, approvable case). When the
scan flags custom code or unsafe files, label the list with 'Our
automatic scanner flagged issues including:'. The Hugging Face
attribution for unsafe files stays in the dialog description.
* Close GGUF-suffix consent bypass for repo ids ending in .gguf
The .gguf short-circuit in _config_has_auto_map skipped the scan for any
model name ending in .gguf, including a bare two-segment repo id like
'evil/model.gguf'. Such a repo can still ship safetensors plus auto_map
Python that transformers would execute, so skipping the scan was an
asymmetric bypass (file_security already scans those repos). Restrict the
short-circuit to genuine direct GGUF file references via
_is_direct_gguf_file_ref: a local .gguf path, or a remote repo_id plus
filename (three or more segments). A two-segment repo id named *.gguf now
falls through to the config scan and _is_gguf_repo file inspection, so it
only skips consent when it actually ships .gguf weights and no safetensors.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align consent dialog body with the title and fix narrow-width overflow
The scan results (the 'Our automatic scanner...' label, finding/unsafe
cards, and the clean-scan reassurance) sat at the dialog's left padding
while the title and description were indented past the status icon, so
the body did not line up under the description. Move the title,
description and results into one column to the right of the icon so they
share a left edge, and let that column fill its width so the description
no longer wraps early.
Also stop a wide code snippet from pushing the dialog off-screen on
narrow viewports: AlertDialogHeader is a grid with place-items-center,
which sized the content row to its content; give the row w-full so it
fills the track, and add min-w-0 down the results chain so the snippet
scrolls inside its card instead of widening the dialog. Verified aligned
and contained from mobile portrait through ultrawide.
* Treat a repo as GGUF-only only when it ships no transformers weights
_is_gguf_repo excluded only .safetensors, so a repo with a .gguf and a
pytorch_model.bin (or .pt/.pth/.h5/.msgpack/.onnx/.ckpt) and no
safetensors was treated as GGUF-only and skipped the consent scan, even
though transformers can load that weight set and execute the repo's
auto_map code. Require the absence of ANY transformers-loadable weight
before treating the repo as a llama.cpp-only GGUF load. A genuine
GGUF-only repo (only .gguf) is still inert; a mixed repo with any pickle
or safetensors weight is gated. Adds a regression test across all the
non-safetensors weight formats.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Block flagged subdir weight shards referenced by a root index
The malware gate treated every subdirectory file as non-loadable, but
from_pretrained deserializes a subdir shard a root index references
(pytorch_model.bin.index.json -> shards/...-00001-of-00002.bin). Read the
root weight indexes and block a flagged subdir pickle the weight_map
points at; a flagged subdir pickle no index lists (NeMo nemo/*.distcp)
stays non-blocking, and an inconclusive index lookup fails closed.
* Pass hf_token to the export checkpoint load
ExportBackend.load_checkpoint scanned with hf_token in the worker but
loaded the weights unauthenticated, so a gated/private checkpoint passed
preflight then 401'd at from_pretrained. Add hf_token to load_checkpoint
and forward token to every from_pretrained branch; the worker passes the
command's hf_token.
* Scope created_by_scan to every HF cache the discard searches
created_by_scan used get_cache_path (active HF_HUB_CACHE only) while
/discard-remote-code deletes across active, legacy, and default caches. A
repo the user already had in a legacy/default cache was marked
scan-created and deleted on decline. Check all three caches for the repo
dir before declaring the scan created it.
* Scan the full .py closure of external auto_map repos
An auto_map cross-repo ref (owner/name--module.Class) only had its entry
file downloaded, but transformers also fetches that file's relative
imports from the same repo, so a dangerous helper.py was left outside the
scanned fingerprint. List each external repo's .py and scan the whole set
(plus the referenced entry files); fail closed if the repo cannot be
listed or fetched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fail closed when a weight index cannot be fully read
_indexed_shard_paths treated a partial result as definitive: if one weight
index read cleanly but another failed transiently, it returned the shard
paths it did see. A flagged subdirectory pickle listed only by the index we
could not read would then be classed as "not a load input" and skipped,
re-opening the very fail-open this guard was added to close.
Return None whenever any index read is inconclusive, even if another read
cleanly, so the caller blocks the already-flagged subdir pickle. A repo that
ships no index files raises EntryNotFoundError for each (never inconclusive)
and still returns an empty set.
* Match cached repos case-insensitively in the created_by_scan guard
_repo_in_any_hf_cache resolved casing only against the active cache and then
probed every cache with an exact directory name. A case-variant already
present in a legacy or default cache (models--Unsloth--Foo for a scan of
unsloth/foo) was missed, so the repo was marked created_by_scan and deleted
on decline -- but discard_remote_code_download deletes case-insensitively,
so that delete would hit the user's pre-existing cache entry. Detect
case-insensitively too, mirroring the deletion path.
* Skip remote-code and security review for selected GGUF variants
validate_model ran the trust_remote_code and Hugging Face security-scan
preflight against the repo even when the selected artifact is a .gguf. A
GGUF loads through llama.cpp, which never executes the repo's auto_map
Python and never deserializes root pickle weights, so repo-level Transformers
artifacts (a config.json with auto_map, or an unsafe pytorch_model.bin next
to the .gguf in a mixed repo) are inert for that load. Gating the GGUF on
them is a false positive. Run both preflights only for non-GGUF loads.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope the malware gate to actual load roots and serialized files
Two fixes to evaluate_file_security so it neither misses a load-path pickle nor
false-blocks an inert file:
- Honor subdirectory load roots. Spark-TTS / BiCodec call from_pretrained on the
snapshot's LLM subdirectory, so a flagged pickle directly under it is a
root-level load artifact there. A new load_subdirs parameter (set from the
model's audio type via security_load_subdirs) reclassifies those files relative
to the load root and looks for weight indexes under it, so a flagged shard in
that subdir is no longer skipped as "not root-level".
- Exempt source files. A root .py is never deserialized by from_pretrained;
executable repo code runs only through auto_map, which the remote-code consent
gate scans. Flagging a Python helper here would false-block a repo that merely
ships a build or train script.
* Scan a LoRA adapter and base as one consent unit, and gate MEDIUM code
A LoRA load runs both the adapter's and the base's repo code. The consent gate
scanned them separately and pinned one fingerprint per repo, so an adapter that
shipped its own auto_map code was either never shown in the dialog (which only
saw the base) or impossible to approve with the base's fingerprint.
evaluate_remote_code_consent_for_targets now scans all of a load's repos as a
single combined unit and pins ONE fingerprint over the union of their code, so
approving the load approves every repo's code together. evaluate_remote_code_consent
becomes a thin single-target wrapper, and an unscannable target fails the whole
load closed.
Also gate MEDIUM findings: like HIGH they now block pending pinned approval, so a
direct API caller cannot run flagged code by setting trust_remote_code=True
without consenting. Only a clean scan loads without a fingerprint.
* Preflight a LoRA load's adapter and base as one combined consent scan
scan_model_remote_code rewrote a LoRA adapter to its base and scanned only the
base for remote code, so the dialog never surfaced an adapter's own auto_map
code. Scan the adapter and base together through
preflight_remote_code_consent_for_targets, which pins one combined fingerprint
the worker gate accepts. The malware preflight is also scoped to each target's
load subdirectories.
* Apply combined consent and subdir-aware malware scan in load workers
Each load worker (inference, export, training) evaluated remote-code consent
once per target with a single shared fingerprint, so a LoRA adapter that ships
its own auto_map code could not be approved by the base's fingerprint. They now
scan the adapter and base together via evaluate_remote_code_consent_for_targets,
which pins one combined fingerprint over the union of their code. The malware
scan in each worker is also scoped to the model's load subdirectories so a
flagged pickle under a from_pretrained load subdir is not missed.
* Report a consistent trust_remote_code requirement after a model loads
validate_model reports requires_trust_remote_code from the YAML default OR the
raw auto_map, but the load, already-loaded, and status responses reported only
the YAML default. A custom-code model approved and loaded via auto_map was then
reported as not requiring trust_remote_code, so the frontend stored false and a
later retry or rollback sent trust_remote_code=false and failed.
A shared resolver reports the same requirement for a loaded model (a value
stored at load time, else the trust_remote_code the load used, else the YAML
default, else the raw auto_map check), and the load response persists it so the
status and already-loaded paths stay consistent. The selected-GGUF security
review is also scoped to the model's load subdirectories.
* Run the consent gate on training resume and for YAML-only trust_remote_code
Three frontend gaps left a model loading without the trust_remote_code it needs:
- The shared consent helper returned early when the scan found no auto_map and no
unsafe files, dropping a requirement that comes from a model's Studio YAML
default (e.g. GLM-4.7-Flash). It now grants the caller's requirement with an
empty pin instead of sending trust_remote_code=false.
- Resume-from-history called startTraining directly with no consent gate, so a
resumed run whose model needs custom code (or an old run with no approved
fingerprint) hit the worker block with no dialog. It now runs the same gate as
a fresh start.
- HF export passed requiresTrustRemoteCode=false for every HF source, so a
YAML-only model could not flip the flag before export. It now signals the
requirement for HF sources.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover both LoRA repos in validate, report GGUF as inert, purge all declined repos
Three follow-on gaps from the combined adapter+base consent work:
- validate_model resolved requires_trust_remote_code from the base alone, so a
LoRA adapter that ships its OWN auto_map code (with a plain base) was reported
as not needing trust_remote_code and the consent dialog never opened. It now
checks the [adapter, base] target set, matching the scan route and the workers
(which already gate both) and the security review already running over both.
- The already-loaded, loaded, and status responses for a selected GGUF reported
requires_trust_remote_code from the model's YAML default. A GGUF loads through
llama.cpp, which never executes the repo's auto_map Python, so the requirement
is inert for that load. They now report False, matching validate_model (which
already skips both gates for GGUF) so a status refresh cannot flip the flag
back on.
- The remote-code scan downloads both the adapter's and the base's config, but
created_by_scan tracked only the primary, so a base the scan was first to pull
into the cache was left on disk when the user declined. The scan now reports
scan_created_repos (every repo it newly cached) and the decline cleanup purges
each; created_by_scan stays for older clients. The frontend falls back to the
primary flag when the list is absent.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scan the repo the load fetches, purge external code on decline, harden consent pins
Six follow-on hardening fixes from a fresh review pass over the gate:
- The malware gate scanned the literal "Spark-TTS-0.5B/LLM" alias, but the trainer
downloads it as unsloth/Spark-TTS-0.5B and loads LLM/, so the alias 404'd and
failed open, missing a flagged LLM/ pickle. evaluate_file_security now resolves
the alias to the repo the loader fetches and scans LLM/ as a load root.
- security_load_subdirs relied only on tokenizer detection, which fails on an
unresolved alias or offline; it now also honors the Studio YAML audio_type
default, so a BiCodec LLM/ load root is not missed.
- The remote-code scan downloads external auto_map repos (owner/name--module.Class),
but the decline cleanup tracked only the model/adapter/base, leaving the external
untrusted code cached. The scan now enumerates external auto_map repos and reports
the ones it created in scan_created_repos, so a decline purges them too.
- External auto_map refs failed the whole load closed on a stale or mis-derived
dotted ref (sub.mod.py vs the real sub/mod.py) even though the actual file was
present and scanned. They now drop such refs when the repo listing is real, exactly
like the own-repo path; an empty/incomplete listing still fetches and fails closed.
- The combined consent fingerprint keyed code by the raw target string, so the scan
endpoint's canonicalized casing and a worker's raw user input produced different
pins for identical code, rejecting a valid approval. Hub repo ids are now folded to
lowercase in the key (local paths stay case-sensitive), so the pin tracks the code.
- Export threaded hf_token into the weight load but not into detect_audio_type /
is_vision_model, so a gated multimodal base 404'd in detection and fell through to
the text loader. Both probes now use the same token.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Thread the token through check-vision and guard the gate's parallel sites
The /check-vision endpoint classified a model without the hf_token, so a gated or
private vision model 404'd in the probe and was reported as a plain text model --
the same dropped-token shape as the export probes, at a sibling site. It now passes
the token like the neighboring /check-embedding endpoint.
Add deterministic consistency guards (tests/test_security_gate_consistency.py) that
enumerate the gate's parallel sites mechanically instead of relying on a review to
spot a missed sibling: every is_vision_model / is_embedding_model / detect_audio_type
caller under routes/ and core/ must thread the token, every GGUF response must report
trust_remote_code via the resolver or False (never the raw YAML default), and every
load worker that runs the malware or consent gate must resolve the LoRA base. A new
site that drops the token or mis-reports the requirement now fails CI directly.
* Narrow the LLM alias rewrite and make audio detection token-aware
Three fixes from the confirmatory review, one a regression from the previous round:
- _load_scan_target rewrote EVERY remote repo ending in "/LLM" to unsloth/<parent>,
so a real third-party repo named "<owner>/LLM" was scanned as unsloth/<owner>
while the loader still fetched the real repo -- a fail-open hole introduced when
the Spark-TTS alias handling was added. It now rewrites only a registry-known
bicodec alias; every other "/LLM" repo is scanned as itself.
- detect_audio_type cached results under the bare model name, so an unauthenticated
probe of a gated/private repo cached None and poisoned a later authenticated call
with the token. The cache is now keyed by (normalized_name, token_fingerprint),
matching the vision cache.
- The training fallback /check-vision call dropped the hf_token, misclassifying a
gated/private VLM when the config endpoint failed. It now passes the token, like
the getModelConfig call it falls back from; checkEmbeddingModel takes the token too.
Extend the consistency guards: every capability cache must be keyed by a tuple
including the token, so a cache re-declared as Dict[str, ...] fails CI.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Document the broad .py scan as deliberate and enforce it with a test
The remote-code scanner scans every .py in a repo once an auto_map exists, not
just the auto_map entry's static import closure. This is intentional: the entry
module can reach a sibling via an absolute import, importlib, or exec, none of
which a static relative-import closure follows, so closure-only scanning would be
a real bypass of a load-time RCE gate. The broad scan never under-scans; the cost
is that an unrelated benign script can over-block, which is the safe failure
direction (HIGH stays approvable; only CRITICAL hard-blocks).
Spell this out at both the local and remote scan sites so the choice reads as
deliberate, and add a test asserting an unrelated, never-imported .py is still
scanned -- so a future narrowing to the static closure fails CI.
* Purge a declined remote LoRA adapter the scan downloaded
scan_model_remote_code probed the created-by-scan state AFTER resolving the base,
but get_base_model_from_lora_identifier downloads a remote adapter's own
adapter_config.json, so the adapter looked already-cached and was dropped from
scan_created_repos. On decline the adapter -- including the auto_map .py the
preflight fetched -- was left on disk, defeating the "untrusted code is not left
on disk" guarantee for the adapter itself.
Snapshot the primary's cache state BEFORE base resolution and use it when marking
the adapter scan-created; on any probe error treat it as pre-existing so a decline
never deletes it. The base and external repos are unaffected (their configs are not
downloaded before their own probe). Add a test that models the mid-scan download
side effect, which the prior static-stub tests did not.
* Clear remote-code approval when the training model changes
Switching the training model from an approved custom-code model to a clean one
kept the previous model's trust_remote_code=true and approved fingerprint in the
store: setSelectedModel reset visionImageSize on a true switch but not the
remote-code approval. The clean model then trained with trust_remote_code=true,
which bypasses the compiler and disables fused cross-entropy.
Reset trustRemoteCode and approvedRemoteCodeFingerprint on a true model switch.
The new model's own YAML default is re-applied by loadAndApplyModelDefaults, and a
custom-code model still re-opens the consent dialog before training starts, so the
only change is that a clean model no longer inherits a stale approval.
* Trim verbose comments across the model-fetching hardening changes
Condense the explanatory comments and docstrings introduced across the
trust_remote_code consent gate, the malware/unsafe-file gate, the remote-code
scanner, the load workers, the model routes, and the security frontend into
fewer, tighter lines while preserving every security rationale (fail-open vs
fail-closed direction, the deliberate broad-scan anti-bypass note, the
empty-vs-unscannable distinction, stale-ref handling, and the alias-rewrite
spoof guard).
Comments and docstrings only. No code, logic, identifiers, or test behaviour
changed; verified comment-only via the AST/TypeScript checker (40/40), with the
backend test suite and frontend tsc green.
* Do not cache transient audio-detection failures
detect_audio_type cached _detect_audio_from_tokenizer's result
unconditionally, so a transient read failure (network error or 5xx,
returned as None) poisoned the cache and the later successful probe never
ran. Mirror the vision cache: _detect_audio_from_tokenizer now returns
(audio_type, definitive) and the caller caches only definitive results.
A read that succeeds with no audio tokens, or clean 404s for every
tokenizer path, stays a cacheable None; only a genuine transient failure
(connection error, timeout, 5xx, malformed body) skips the cache so the
next call retries.
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: add 'Load on selection' toggle to configure load options before loading
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: seed staged speculative decoding from the standing default
* Studio: address PR review for load-on-selection staging
* Studio: handle direct GGUF staging and stale-stage edge cases from load-on-selection review
* Studio: cancel replaced staged downloads and keep staged pick on load failure
* Studio: centralize staged-download cancel and guard staged-load restore
* fix: address staged GGUF load review
* fix: honor staged GGUF load metadata
* fix: clarify load-on-selection tooltip
Keep the load-on-selection hint visually anchored to the control and make the on/off behavior explicit without changing the broader deferred-load flow.
* Studio: reset orphaned staged knobs on abandon and cap Max Tokens to staged context
* Studio: remove dead code and cancel staged download when loading a different model
* fix: surface staged model in run settings before deferred load
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
* Studio: Add inline confirmation (Allow/Always allow/Deny) for tool calls
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix race in tool-call confirmation gate
* Studio: gate built-in tool calls and harden the confirmation handshake
The Allow / Always allow / Deny controls only lived in the fallback tool
card, but the built-in tools (web search, python, terminal, code
execution, image generation) render with their own components and so
never showed the buttons. Those calls paused after tool_start with no way
to approve them, hanging until the 1 hour timeout. Only MCP tools, which
use the fallback renderer, actually worked.
Render the controls for every tool card by wrapping each registered tool
component (and the fallback) in thread.tsx with a shared
ToolConfirmationControls, so the gate applies uniformly.
Also make the handshake robust:
- The gate keys on a per-call approval_id minted by the backend and
echoed in tool_start, instead of session_id alone, so a stale or
concurrent confirmation can no longer resolve the wrong call.
- The approval slot is registered before tool_start is yielded, closing
the race where a fast click or an auto "Always allow" could reach the
backend before the waiter existed.
- The frontend resolves with the same session id the request was sent
with (plus the approval_id), fixing the new-thread mismatch where the
confirmation targeted a different session than the blocked stream.
- The confirm endpoint returns {resolved}; the UI keeps the buttons and
shows a retry hint until the backend confirms a match, instead of
hiding them on a failed or mistargeted post.
- The gate runs after the disabled-tool and duplicate-call checks, so a
call that will not execute is not put up for approval. A denied call is
still excluded from duplicate detection, so re-issuing and approving it
works.
- "Always allow" is scoped per session to match the backend gate.
Add backend tests for the approval registry, the SSE no-deadlock
handshake, and the loop integration (allow, deny, disabled, duplicate,
re-issue after deny).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Move "Confirm tool calls" to the Tools section
* Studio: add Bypass Permissions (skip confirmation, disable tool sandbox)
Adds an opt-in Bypass Permissions toggle next to Confirm tool calls. When on,
no tool call shows a confirmation prompt and the python/terminal sandbox is
disabled: safety checks, command blocklist, and resource limits are skipped.
Secret env vars are still stripped and HOME stays repointed at the session
workdir. Default off keeps current behavior, and it takes precedence over
Confirm tool calls. Enabling it requires accepting a warning each time.
* Studio: harden Bypass Permissions secret handling and fix Anthropic tool path
Follow-up to the Bypass Permissions feature. Addresses the review findings:
- Anthropic /v1/messages 500: declare bypass_permissions on
AnthropicMessagesRequest so tool requests that omit the field default to
False instead of raising AttributeError (extra='allow' does not set absent
attributes).
- /proc parent-env leak: stripping the child env did not stop a same-uid
bypassed child from reading /proc/<parent>/environ to recover the
tool-executing process's unfiltered secrets. Clear PR_SET_DUMPABLE on that
process before the first bypass exec so its /proc entries become root-owned.
Hardening is fail-closed: if prctl is denied, bypass execution is refused
rather than run with the parent environ still readable. Mitigation, not a
full boundary; documented in the code.
- Broker/capability vars: strip SSH_AUTH_SOCK, SSH_AGENT_PID, GPG_AGENT_INFO,
GNUPGHOME, KUBECONFIG, DOCKER_HOST so a bypassed tool cannot use the
operator's live agents.
- Credential-bearing URL values: drop any env var whose value embeds URL
userinfo (scheme://user:pass@ and token-only scheme://token@) regardless of
the variable name. Benign proxy/index URLs without credentials are kept, so
proxy-only and internal-index setups still work in bypass mode.
- Windows temp isolation: repoint TEMP and TMP (not just TMPDIR) at the
per-session sandbox dir.
- Frontend: stop persisting bypassPermissions; a reload now starts with the
sandbox/confirmation bypass off and requires re-accepting the warning dialog.
Adds regression tests for each finding in test_bypass_permissions.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strip cred-location env vars (HF_HOME etc.) in Bypass Permissions
Repointing HOME did not stop SDKs auto-reading cached creds via vars that
point at the real home/cache/config: HF_HOME (startup always sets it; token
lives under $HF_HOME/token), HF/XDG cache roots, NETRC/BOTO_CONFIG/
PIP_CONFIG_FILE, and Windows HOMEDRIVE/HOMEPATH. Drop those, and repoint
USERPROFILE/APPDATA/LOCALAPPDATA at the per-session workdir. Adds regression
tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: lock in bypass HF token resolution with an end-to-end test
The drop-based fix relies on the whole HF_HOME/XDG fallback chain being
removed so huggingface_hub resolves under the repointed HOME. Add a test
that sets HF_HOME and XDG_CACHE_HOME at a real cache and asserts the
resolved token path lands under the workdir, not the operator's cache.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strip npm _auth, MYSQL_PWD, and BASH_ENV from bypass env
Three more credential vectors dodged the bypass scrubber: NPM_CONFIG__AUTH
(npm _auth, base64 so no URL userinfo and no AUTH marker), MYSQL_PWD (markers
use PASSWD, not PWD, since PWD is the cwd var), and BASH_ENV (bash -c sources
it for non-interactive shells, so a startup file can re-export stripped
secrets). Add an AUTH marker, the exact MYSQL_PWD name, and drop BASH_ENV plus
PGPASSFILE. Adds regression tests incl. an end-to-end check that a bypass
terminal call does not source BASH_ENV.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: extend bypass env scrubber and enforce confirm precedence in loops
From a parallel review pass over the bypass changes:
- Drop more credential-location vars in _build_bypass_env: npm/yarn/git/cargo/
rclone config pointers (NPM_CONFIG_USERCONFIG, NPM_CONFIG_GLOBALCONFIG,
YARN_RC_FILENAME, GIT_CONFIG_GLOBAL, GIT_CONFIG_SYSTEM, CARGO_HOME,
RCLONE_CONFIG) and the GIT_ASKPASS/SSH_ASKPASS auth helpers.
- Enforce confirm_tool_calls AND NOT bypass_permissions inside the safetensors
and GGUF tool loops, not just at the route, so a direct internal caller
passing both flags never prompts.
- Soften the toggle hint: environment secrets are stripped, but bypassed code
can still read files and credentials on the machine (no overclaim that keys
stay hidden).
Adds regression tests for the new names and the loop-level precedence.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add GGUF loop test for bypass-over-confirm precedence
The safetensors loop precedence is covered behaviorally; the GGUF loop needs a
live llama-server so add an AST guard asserting its _needs_confirm gate
references both confirm_tool_calls and bypass_permissions, matching the other
llama_cpp source-inspection tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add red Bypass Permissions badge in the composer
When Bypass Permissions is on, show a persistent red pill in the composer
tool-pill row (like the Search/Code pills), matching Claude Code's always-
visible bypass indicator. Clicking it turns bypass off, mirroring the other
composer toggles. Enabling still goes through the settings toggle + warning
dialog. Adds a data-variant=danger style for the destructive-colored pill.
* Studio: show Bypass Permissions badge in the Thread composer too
The empty-state and active Thread render their own composer (thread.tsx),
not shared-composer, so the badge only appeared in the split layout. Mirror
the red dismissible pill in ComposerAction so it shows in every composer.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep the Bypass Permissions badge visible when the composer is collapsed
The Thread composer only renders the pill row when expanded, so the active-mode
badge vanished on the default (collapsed) empty state. Render it before the
expand gate (it returns null when bypass is off) so the red indicator always
shows while bypass is on.
* Studio: make the Bypass Permissions confirm button a solid red button
The destructive button variant is a subtle 10% tint that read as bare red text
next to the outlined Cancel. Force the solid destructive fill (the variant's
class loses to the tint through AlertDialogAction's Slot merge, so use the !
override the codebase already uses for this case) and shorten the label to
'I understand' so it fits the small dialog's two-column footer.
* Studio: add Bypass Permissions to the composer + More menu
Adds a 'Bypass Permissions' entry to the composer plus-menu (under More by
default) in both composers, so it can be toggled without opening Run settings.
Enabling routes through the same danger warning dialog; disabling is immediate.
A shared BypassPermissionsMenuItem keeps the two composers in sync.
* Studio: harden bypass env scrubber for IMDS opt-out and connection strings
Two gaps in the Bypass Permissions secret scrubber:
- The broad AWS_ prefix also dropped AWS_EC2_METADATA_DISABLED, a non-secret
opt-out. Removing it re-opens the IMDS instance-role credential path that the
operator explicitly disabled, so a bypassed boto/AWS-CLI call could recover
cloud creds. Keep that flag (and AWS_EC2_METADATA_V1_DISABLED) via a keep-list
while still stripping the real AWS credential vars.
- Azure App Service connection strings (SQLCONNSTR_/CUSTOMCONNSTR_/...,
WEBSITE_CONTENTAZUREFILECONNECTIONSTRING) and values like Password=/AccountKey=
/SharedAccessKey= slipped past the name and URL-only value classifiers. Add
CONNSTR/CONNECTIONSTRING name markers and a connection-string value matcher.
* Studio: let Bypass Permissions suppress the confirm-tool-calls guards
The confirm-vs-bypass precedence (confirm and not bypass) was applied at the
loop call sites but not at the earlier request guards, so a client sending
confirm_tool_calls + bypass_permissions together was rejected (stream=true
required / unsupported for external or Anthropic tools) before the precedence
took effect. Gate all four confirm guards on not bypass_permissions so both
flags together proceed with the gate suppressed, matching the documented rule.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* feat: add Anthropic-compatible thinking parameter
Add `thinking` parameter using Anthropic's format ({type: 'disabled'} /
{type: 'enabled'}) alongside the existing `enable_thinking` boolean for
backward compatibility.
The new parameter is mapped internally to `enable_thinking` at the route
layer so all downstream templates and backends continue to work unchanged.
Changes:
- Add ThinkingConfig model and `thinking` field to ChatCompletionRequest
- Add mapping logic in routes: thinking.type -> enable_thinking
- Add `thinking` field to frontend TypeScript types
- Update frontend request building to send thinking parameter
- Add tests for new thinking parameter
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: move thinking→enable_thinking mapping to model_validator
The Gemini review correctly identified that the route-level mapping
bypasses normalization for external provider requests. Moving the
mapping into a @model_validator on ChatCompletionRequest ensures it
runs during Pydantic validation regardless of routing path.
* Document ThinkingConfig scope and thinking validation behavior
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Studio: serve DiffusionGemma GGUFs with the on-device visual decoder
* Studio: render the DiffusionGemma denoising canvas live in chat with honest stats
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden DiffusionGemma runner resolution (Windows .exe, build/bin lookup, clear stale audio flag, safe PYTHONPATH, Linux-only pdeathsig)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(studio/responses): forward chat_template_kwargs enable_thinking to chat request
The /v1/responses translation in _build_chat_request dropped
chat_template_kwargs (e.g. {"enable_thinking": true}) sent via the
Responses extra-body, so reasoning control was silently ignored.
Lift enable_thinking onto the typed ChatCompletionRequest field,
mirroring openai_chat_completions, so both the non-streaming and
streaming Responses pass-through paths honor it.
Fixes#6198
Signed-off-by: Tai An <antai12232931@outlook.com>
* Fix/adjust Responses reasoning for PR #6202
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix/adjust reasoning none for PR #6202
* Fix/adjust structured reasoning for PR #6202
* Fix/adjust responses reasoning review findings for PR #6202
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix/adjust responses reasoning follow-ups for PR #6202
* Fix/adjust think parsing gate for PR #6202
---------
Signed-off-by: Tai An <antai12232931@outlook.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: Add Tensor-Parallel llama.cpp support
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Tensor-Parallel fallback and GPU selection
* Studio: reconcile split-mode extras and harden tensor-split planning
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: reconcile split-mode extras in backend duplicate-load guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: preserve inherited non-tensor split modes on reload
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor cancellation in tensor fallback, preserve tensor mode on rollback, and don't raise an explicit small context
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: reconcile split-mode in reload check and strip it on tensor downgrade
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strip --tensor-split alongside --split-mode so inherited ratios don't override the tensor planner
An inherited or stale --tensor-split in llama_extra_args was appended after
Studio's computed --tensor-split and won last in llama.cpp, re-introducing the
asymmetric-GPU OOM tensor mode is meant to prevent. Group -ts/--tensor-split
into the split-mode shadow set so it is stripped on inherit and on the layer
fallback; parse_split_mode_override still keys on the mode value only.
* Drop quantized KV for the tensor attempt and report native max context
Tensor mode aborts on a quantized KV cache, so a user with q8_0/q4_1 etc. who
enabled Tensor Parallelism silently fell back to layer split. Clear the cache
type (and strip inherited/explicit --cache-type) for the tensor attempt only;
the layer fallback re-runs with tensor off and keeps the user's choice.
Also report max_available_ctx from the native context, not an explicit small
-c, so the context slider no longer warns too early in tensor mode.
* Reconcile inherited split-mode extras in the already-loaded check
When a same-model load omitted llama_extra_args, the tensor comparison resolved
the raw (None) request and treated an inherited --split-mode tensor server as a
mismatch, forcing a needless reload. Compare using the stored extras stripped
the same way the reload strips them.
* Pass tensor_parallel through compare-mode loads
The generalized compare path loaded each GGUF without tensor_parallel, so
compare ran layer split even with the toggle on and left the settings sheet
stale. Send the toggle and hydrate the loaded state from the response, matching
the main chat and recipe load paths.
* Add --tensor-parallel flag to unsloth studio run
The headless one-liner could only reach tensor mode by passing --split-mode
tensor as a raw llama.cpp extra. Add a first-class --tensor-parallel/
--no-tensor-parallel option that sets the tensor_parallel field on the
/api/inference/load payload, forwarded through the studio-venv re-exec like the
other polarity flags. Matches the web UI toggle and the API field.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* Studio: Add inline confirmation (Allow/Always allow/Deny) for tool calls
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix race in tool-call confirmation gate
* Studio: gate built-in tool calls and harden the confirmation handshake
The Allow / Always allow / Deny controls only lived in the fallback tool
card, but the built-in tools (web search, python, terminal, code
execution, image generation) render with their own components and so
never showed the buttons. Those calls paused after tool_start with no way
to approve them, hanging until the 1 hour timeout. Only MCP tools, which
use the fallback renderer, actually worked.
Render the controls for every tool card by wrapping each registered tool
component (and the fallback) in thread.tsx with a shared
ToolConfirmationControls, so the gate applies uniformly.
Also make the handshake robust:
- The gate keys on a per-call approval_id minted by the backend and
echoed in tool_start, instead of session_id alone, so a stale or
concurrent confirmation can no longer resolve the wrong call.
- The approval slot is registered before tool_start is yielded, closing
the race where a fast click or an auto "Always allow" could reach the
backend before the waiter existed.
- The frontend resolves with the same session id the request was sent
with (plus the approval_id), fixing the new-thread mismatch where the
confirmation targeted a different session than the blocked stream.
- The confirm endpoint returns {resolved}; the UI keeps the buttons and
shows a retry hint until the backend confirms a match, instead of
hiding them on a failed or mistargeted post.
- The gate runs after the disabled-tool and duplicate-call checks, so a
call that will not execute is not put up for approval. A denied call is
still excluded from duplicate detection, so re-issuing and approving it
works.
- "Always allow" is scoped per session to match the backend gate.
Add backend tests for the approval registry, the SSE no-deadlock
handshake, and the loop integration (allow, deny, disabled, duplicate,
re-issue after deny).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Move "Confirm tool calls" to the Tools section
* Studio: Keep tool group open while a tool call awaits confirmation
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix tool confirmation session scope for PR #5869
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix confirmation follow-ups for PR #5869
* Apply pre-commit formatting for PR #5869
* Fix confirmation cleanup for PR #5869
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden confirmation lookups for PR #5869
* Studio: make the tool-call confirmation decision immutable
resolve_tool_decision accepted a second confirmation for the same approval_id
and overwrote slot["decision"] in the window before the waiter reads it and
pops the slot, so a duplicate or out-of-order POST could flip an Allow to Deny
(and returned a misleading resolved:true). Reject once the slot's event is
already set so the first decision wins. Adds a regression test.
* Fix/adjust tool confirmations for PR #5869
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
* Studio: report the real llama-server context window and add an opt-in overflow policy for OpenAI-compatible serving
A community report showed OpenCode failing tool calls every few minutes
against Studio's OpenAI-compatible API while the same GGUF was stable on
LM Studio. Root cause: Studio advertises the requested context length, but
llama-server can allocate less (memory-fit step on small GPUs, --parallel
slot split), so clients budget against a window that does not exist. Their
generations truncate mid tool call at the real wall (finish_reason=length
with cut JSON arguments) and eventually the prompt itself exceeds the real
window, returning a 400 that agentic clients treat as non-retryable.
Changes:
- After llama-server health, read default_generation_settings.n_ctx from
/props and adopt it whenever it is below Studio's computed context, with
a warning. The load response, status route, UI value, and the passthrough
max_tokens ceiling all become honest automatically.
- Expose context_length and max_context_length on /v1/models so clients can
budget against the enforced window.
- Accept empty role=tool content (commands with no output are routine in
agentic loops; OpenAI and llama-server both accept it) instead of a 400.
- Add context_overflow=truncate_middle (per request, or server-wide via
UNSLOTH_CONTEXT_OVERFLOW=truncate_middle): on exceed_context_size_error
the passthrough drops whole middle turn-groups (system prompt, first turn,
and recent turns kept; tool calls stay paired with their results), clips
oversized contents middle-out when group-dropping is not enough, clamps
max_tokens to the generation headroom, and retries. Default stays 'error'
with code=context_length_exceeded so clients running their own compaction
keep full control.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: allocate the requested context for real (kv-unified, fit-ctx floor)
Two launch-flag gaps caused the advertised vs allocated divergence at the
source:
- llama-server enables --kv-unified only when the slot count is auto; Studio
always passes --parallel N, which silently splits -c into per-slot windows
of -c/N. Pass --kv-unified when N > 1 so a single request can use the full
advertised window (same total KV memory, shared pool).
- with --fit on the fit step may set ctx as low as 4096; pass
--fit-ctx <requested> for explicit requests so fit offloads or fails into
the existing --fit off retry instead of silently shrinking the window.
Both flags are gated on --help capability probing so older builds keep the
current behavior, where the /props readback remains the backstop. Verified
live: -c 98304 --parallel 4 now serves per-slot n_ctx 98304 (was 24576),
48k-token requests pass through the passthrough, and the readback warning no
longer fires.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: surface the llama.cpp update affordance when MTP is disabled
When a model asks for MTP (auto on an MTP model, or forced mtp / mtp+ngram)
but it gets disabled, the load already degrades gracefully and serves without
speculative decoding. Until now the UI gave no hint why, or that an update
would fix it.
Record why MTP was dropped on the backend (spec_fallback_reason): the probe
found no mtp token (binary_no_mtp), the spawn aborted with an outdated-arch /
context-build error such as a prebuilt that predates the Gemma drafter
(binary_outdated), or the current build could not run it, e.g. a CUDA kernel
limit (runtime_error). Expose it in the inference status. In the chat
Speculative Decoding section, show a short note and, for the two update-fixable
reasons, an inline Update llama.cpp button that reuses the existing update flow.
A runtime_error gets the note without an update push, since a newer build may
not fix it.
Backend tests cover the reason being set / cleared. Frontend typechecks.
* Address review: tighten the update hint to genuinely outdated binaries
Reserve binary_outdated (which surfaces the Update llama.cpp affordance) for an
unknown-architecture abort, which proves the prebuilt predates the model;
classify the generic memory/context build failures as runtime_error, where an
update may not help. Frontend: only append the "Update llama.cpp to enable it"
sentence when an update is actually available, so the text never points at an
action the UI is not offering.
* Studio: sync detected model capabilities into models[] after load
The chat composer gates audio upload on activeModel.hasAudioInput, but
/api/models/list omits audio fields for default and active-GGUF entries
and the single chat load path never wrote the load response's
capability flags back into the store. Audio-capable models such as the
Gemma 4 GGUFs therefore never unlocked audio input in the main chat,
while the compare composer (which does sync) worked.
Add syncModelCapabilities and call it after a successful load and after
the status fetch in refresh, so the flags also survive F5 and are not
clobbered by stale catalog data.
* Studio: merge audio upload into the Add photos & files picker
Remove the separate Upload audio row from the composer plus menu and
register an AudioAttachmentAdapter in the shared attachment pipeline,
so the standard picker and drag-drop accept wav, mp3, m4a, ogg, flac
and webm directly. Gating matches images: the picker always lists
audio and models without audio input get a toast at add() time. The
50MB limit is kept and the file shows as a normal attachment chip.
On send the adapter emits an audio content part on the attachment and
findLatestUserAudioBase64 now also scans attachment content, so the
request still carries audio_base64 exactly as before.
* Studio: extract AudioAttachmentAdapter into its own module
Move the adapter out of runtime-provider.tsx so it is importable in
isolation, export the audio send-path and capability-sync helpers for
tests, and guard attachment id generation for non-secure contexts
(crypto.randomUUID is undefined over plain HTTP on a LAN, matching the
existing guard in startCompare).
* Studio: do not claim .webm by extension in the audio adapter
A video/webm file would match the .webm extension entry and route to
the audio adapter. Real audio webm (MediaRecorder output) always
reports the audio/webm MIME, so matching webm by MIME only keeps video
files out while keeping recorded audio working.
* Studio: only send audio from the newest user message
audio_base64 switches the backend onto the audio generation path
(generate_whisper_response ignores chat messages entirely and
generate_audio_input_response bypasses the normal streaming path), so
replaying audio from an older turn hijacked text-only follow-ups:
Whisper would retranscribe the stale clip instead of erroring cleanly,
and audio VLMs lost tools and streaming. Stop the scan at the newest
user message, matching the consumed-on-send semantics of the legacy
pendingAudio path. Regenerating the audio turn itself still resends
its audio since it is the newest user message in that run.
Also guard extractAudioPartBase64 against null parts in deserialized
history content.
* Studio: forward audio input to llama-server for GGUF models (#6096)
* Studio: forward audio input to llama-server for GGUF models
* Studio: harden GGUF audio input handling (multi-format decode, size cap)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: carry GGUF audio in the message list so it works with tools
* Studio: bound decoded audio length and make the soundfile decoder optional
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle audio attachment edge cases
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate audio file picker by loaded model capability (#6142)
* Gate audio attachments by loaded model
* Use conditional spread for audio attachment adapter
* Preserve audio fallback while filtering picker
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: oobabooga <oobabooga4@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Studio: support separate-file MTP GGUF drafters (Gemma 4)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix review findings for separate-file MTP drafters
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: pair local MTP drafters by name and include them in reload dedup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: manage --model-draft in extras and reject MTP/ copies as models
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Adds a self-contained RAG stack to Studio: knowledge bases with chunked indexing, hybrid (dense + lexical) retrieval, and an automatic first-pass context inject into chat. Embeddings run through a local llama-server GGUF backend (default unsloth/bge-small-en-v1.5-GGUF) with a sentence-transformers fallback. The chat tool loop gains a search_knowledge_base tool, a per-turn re-search cap, and source citation, layered on top of the shared ToolLoopController.
* Studio: fix OpenAI- and Anthropic-compatible API spec compliance
* Studio: fix API spec-compliance gaps on passthrough and streaming paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: carry context_length_exceeded through the OpenAI passthrough error path
* Studio: count tool-schema tokens in the Anthropic server-tool stream, and small stream-handling guards
* Studio: guard message_delta usage against None and normalize developer role before proxying
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor max_completion_tokens on the external-provider proxy path
* Studio: forward llama-server cached_tokens into OpenAI prompt_tokens_details
* Studio: sanitize messages in count_tokens to match the /v1/messages prompt
* Studio: report max_tokens for truncated tool calls and guard null usage in metadata events
* Studio: drop the request-id middleware (headers aren't declared in either spec)
* Studio: include the required request_id field in Anthropic error bodies
* Studio: honor max_completion_tokens on the audio (TTS / audio-input) paths
* Studio: add the _effective_max_tokens helper and route all max-token sites through it
* Studio: align API compatibility edge cases
* Studio: clarify multi-choice chat support
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: clarify logprobs chat support
* Studio: opt the local chat UI into the streaming usage chunk so the context bar and tok/s repopulate
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: forward seed to llama-server, and fix Anthropic server-tool stop_reason, tool_result id correlation, and parallel-tool execution cap
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: align OpenAI chat completion spec edge cases
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: align backend API compatibility tests
* Studio: honor tool caps and internal stream usage
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: coerce nullable stream usage counts
* Studio: preserve system prompts with developer messages
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Daniel Han