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.
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.
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).
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.
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).
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.
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.
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.
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).
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
