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).
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).
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 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.
* Studio: honor stream=false on the GGUF agentic tool path (#6570)
* Studio: dedup the #6570 non-streaming tool tests and cover cached_tokens
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: cover the cached_tokens metadata fix and clarify the drain comment (#6570)
* Studio: align the GGUF tool drain naming and tighten its comment (#6570)
---------
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>
* Studio: cap GGUF context to unified memory on Apple Silicon
* Studio: tighten Apple ctx-cap comments and drop the overstated MLX-sync claim
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: reserve flat MTP fraction and floor sparse-KV ctx in the Apple unified-memory cap
The Apple Silicon GGUF context cap mirrored the discrete-GPU auto-fit branch but
missed two protections the discrete path already applies:
- It passed the full unified-memory budget with budget_frac=1.0 without first
reserving the flat MTP fraction the discrete path takes off via _pin_fraction.
With an MTP draft whose KV cannot be byte-sized (e.g. Qwen3.6-MTP, #6529), the
cap filled the whole budget and left nothing for the draft, so unified memory
could still over-commit. Reserve _flat_mtp_reserve up front; this is a no-op
when MTP is not engaged.
- It required _can_estimate_kv(), so a GGUF with sparse KV metadata skipped the
cap entirely and launched at full native context. Mirror the discrete
file-size-only fallback and floor the auto context to 4096 when the cache
cannot be sized.
Adds regression tests for both paths.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten comments in the Apple unified-memory context cap
Condense the verbose comment blocks in the Apple budget helper, the no-GPU
Metal branch, and the context-fit tests. Comments only, no code change
(verified with ast-based comment_tools check); suite still green.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Studio: lazy-import matplotlib so the server starts when the wheel is blocked
matplotlib.pyplot was imported at the top of core/training/training.py, on the
server boot path. When matplotlib's native extension fails to load (e.g. an
unsigned wheel blocked by Windows Smart App Control), that import crashed the
whole Studio server at startup instead of just disabling loss plots.
Move it into a lazy _load_pyplot() helper called from _create_loss_plot, using
the headless Agg backend, and return None when matplotlib is unavailable so
plotting degrades gracefully. The plot return was already Optional, so callers
need no changes. Keep the type-only import under TYPE_CHECKING and quote the
annotations.
Fixes#6588
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: pin matplotlib==3.11.0
Pin matplotlib to the current latest so a new unsigned release does not
reintroduce the Smart App Control block on Windows. Belt-and-suspenders on
top of the lazy import. Pinned in both studio.txt and extras.txt.
* Pin matplotlib to 3.10.9 so Studio still installs on Python 3.10
matplotlib 3.11.0 requires Python >=3.11, so the pin had no installable wheel on
Python 3.10 (still supported) and pip install failed there. 3.10.9 is the latest
3.10.x (requires-python >=3.10) and covers Python 3.10 through 3.13. Also tighten
the lazy-import docstrings.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.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>
* studio: report the true reasoning duration and fix the Stop button for thinking models
For a local GGUF the "Thought for N" label was timed entirely on the client by a
brittle edge-detector, so an always-think model (Qwen3 MTP) that buffers its whole
reasoning and flushes it in one chunk showed "1 second" instead of the real
minute-plus. The client cannot time reasoning it receives atomically, so make the
timing backend-authoritative.
Backend: generate_chat_completion_with_tools measures wall-clock reasoning and
emits a Studio reasoning_summary event (duration_ms) at the moment reasoning ends
-- the first answer token, or end-of-stream for a reasoning-only reply -- for both
the tool-detection pass and the final-answer pass. Timing resets per tool
iteration so the final answer's thinking time wins on the client (which takes the
latest reasoning_summary). routes/inference.py forwards the event in the GGUF tool
stream.
Frontend: parse the reasoning_summary SSE into a _reasoningDurationMs chunk and
use it as the authoritative reasoning duration (last write wins), clamped to >= 0
and guarded to a finite number so a malformed or proxied chunk cannot produce a
NaN label; the persisted value wins for the final "Thought for N" label, with the
previous live timer kept only as a fallback when no metadata arrives.
* [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>
