Feat/sam3 onnx hub support (issue #324)#582
Conversation
There was a problem hiding this comment.
Pull request overview
Adds support for Hub-hosted pre-exported ONNX inputs (<org>/<repo>/<path>.onnx) to enable SAM 3 consumption via Scenario D, and fixes pre-quantized (QDQ + QOperator) detection/routing so already-quantized models skip ORT optimization/quantization stages that can crash on unsupported integer kernels.
Changes:
- Introduces
loader/onnx_hub.pyto recognize and download Hub-hosted ONNX files (plus best-effort.onnx_datasidecars), and wires resolution into multiple entry points (CLI + inference/model loading). - Expands
is_quantized_onnxto detect QOperator quantized models and adds shared quant-op constants. - Updates build pipelines to honor “pre-quantized” routing (skip optimize/quantize) and adds regression/unit/integration tests for SAM 3 and Hub-ONNX refs.
Reviewed changes
Copilot reviewed 26 out of 26 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
| tests/unit/onnx/test_detection.py | New tests for quantized (QDQ + QOperator) and compiled ONNX detection. |
| tests/unit/loader/test_onnx_hub.py | New unit tests for Hub-ONNX path detection/splitting/downloading behavior. |
| tests/unit/inference/test_engine.py | Ensures Hub-ONNX refs are resolved before inference routing logic. |
| tests/unit/commands/test_perf_cli.py | Verifies winml perf resolves Hub-ONNX refs before ONNX path validation. |
| tests/unit/commands/test_hub_onnx_ref.py | New CLI tests for wmk config/wmk build with Hub-hosted ONNX refs. |
| tests/unit/commands/test_eval.py | Adds Hub-ONNX resolution coverage in eval model-path resolution. |
| tests/unit/build/test_onnx.py | Regression tests ensuring pre-quantized ONNX skips optimize/quantize stages. |
| tests/unit/build/test_hf.py | Regression tests ensuring pre-quantized exported ONNX skips optimize/quantize stages. |
| tests/integration/test_sam3_e2e.py | End-to-end integration tests for SAM3 decoder + encoder via Hub-hosted ONNX artifacts. |
| src/winml/modelkit/onnx/detection.py | Updates quantized detection to use unified QDQ+QOperator op set. |
| src/winml/modelkit/models/auto.py | Resolves Hub-ONNX refs before fast-path ONNX/HF routing in from_pretrained. |
| src/winml/modelkit/loader/onnx_hub.py | New implementation for Hub-hosted ONNX ref resolution via hf_hub_download. |
| src/winml/modelkit/loader/init.py | Re-exports Hub-ONNX helper APIs. |
| src/winml/modelkit/inference/engine.py | Normalizes Hub-ONNX refs to local paths in load and load_schema_only. |
| src/winml/modelkit/data/hub_models.json | Adds SAM3 encoder/decoder catalog entries using Hub-ONNX refs. |
| src/winml/modelkit/compiler/utils.py | Adds QOperator + union quantization op-type constants. |
| src/winml/modelkit/compiler/init.py | Exposes new quantization op-type constants from compiler package. |
| src/winml/modelkit/commands/perf.py | Resolves Hub-ONNX refs prior to ONNX path checks. |
| src/winml/modelkit/commands/inspect.py | Treats Hub-ONNX refs as ONNX inputs for consistent “not supported” messaging. |
| src/winml/modelkit/commands/eval.py | Resolves Hub-ONNX refs before validating ONNX file existence. |
| src/winml/modelkit/commands/config.py | Resolves Hub-ONNX refs before dispatching config generation path. |
| src/winml/modelkit/commands/build.py | Resolves Hub-ONNX refs and adds skip-optimize plumbing for ONNX pipeline. |
| src/winml/modelkit/build/onnx.py | Ensures pre-quantized models skip ORT optimize and don’t crash on QOperator ops. |
| src/winml/modelkit/build/hf.py | Same as above for HF-exported ONNX artifacts that are already quantized. |
| src/winml/modelkit/build/common.py | Adds skip_optimize to the shared optimize/analyze loop. |
| README.md | Documents Hub-hosted ONNX input form and supported commands. |
Comments suppressed due to low confidence (1)
src/winml/modelkit/commands/build.py:1162
--no-optimizesetsextra_kwargs["skip_optimize"], but_build_onnx_pipeline()never reads it. As a result, the CLI flag has no effect unlessis_quantized_onnx()happens to detect the model as pre-quantized. Consider consumingextra_kwargs.pop("skip_optimize", False)and combining it withis_pre_quantized(and settingmax_iters=0when skipping) so users can force skipping the optimize/autoconf stage for problematic ONNX inputs.
from ..onnx import copy_onnx_model, is_quantized_onnx
max_iters: int = extra_kwargs.pop("hack_max_optim_iterations", 3)
# ── Validate + setup ─────────────────────────────────────────
if not onnx_path.exists():
raise FileNotFoundError(f"ONNX file not found: {onnx_path}")
try:
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| # ``ConvInteger``. Skip the optimize pass and the autoconf re-optim | ||
| # loop; analyze still runs lint-only. |
| logger.info( | ||
| "Pre-quantized model detected (QDQ nodes present). " | ||
| "Pre-quantized model detected (QDQ or QOperator nodes present). " | ||
| "Skipping optimize + quantize, running analyze-only." | ||
| ) | ||
| stages_skipped.append("optimize") | ||
| # Optimize+analyze only, no autoconf re-optimization | ||
| # Analyze-only: skip ORT-based graph optimization (no kernel for | ||
| # QOperator ops like ConvInteger on the host EP), no autoconf loop. | ||
| current_path, _, analyze_iters, analyze_unsupported, analyze_details = ( |
| if is_pre_quantized: | ||
| logger.info( | ||
| "Pre-quantized model detected (QDQ nodes present). " | ||
| "Pre-quantized model detected (QDQ or QOperator nodes present). " | ||
| "Skipping optimize + quantize, running analyze-only." | ||
| ) | ||
| stages_skipped.append("optimize") | ||
| # Optimize+analyze only, no autoconf re-optimization | ||
| # Analyze-only: skip ORT-based graph optimization (no kernel for | ||
| # QOperator ops like ConvInteger on the host EP), no autoconf loop. | ||
| current_path, _, analyze_iterations, analyze_unsupported_nodes, analyze_details = ( |
| ep: Target execution provider for the analyzer. | ||
| device: Target device for the analyzer. | ||
| max_optim_iterations: Maximum autoconf re-optimization rounds. | ||
| 0 means optimize+analyze only (no autoconf re-optimization). | ||
| skip_optimize: When True, skip the initial ``optimize_onnx`` call and | ||
| just copy the input model to ``optimized_path``. Used for | ||
| pre-quantized models (QDQ or QOperator format) where ORT-based |
DingmaomaoBJTU
left a comment
There was a problem hiding this comment.
(supersedes the earlier review — consolidated with design feedback)
Code-level issues
See inline comments below.
One additional nit not inlined (existing code, not touched by this PR): _run_quantize_stage at line 952 of commands/build.py still says "(QDQ nodes already present)". Since this PR broadens detection to cover QOperator, the message should read "(QDQ or QOperator nodes already present)" for consistency with the log messages updated in build/onnx.py and build/hf.py.
Design feedback
Three structural concerns — the bug fixes (QOperator detection, skip_optimize) are solid, but the way the Hub-ONNX input form is introduced creates maintenance surface that should be addressed before or shortly after merge.
1. "Hub-hosted ONNX" is not a distinct input type — it is a download step
org/repo/path/file.onnx -> hf_hub_download() -> /local/cache/file.onnx
After the download, this is just a local .onnx file. The downstream pipeline does not (and should not) care that it once lived on the Hub. Yet the PR threads detection logic (is_hf_onnx_path / maybe_resolve_hf_onnx_path) through 7+ command entry points, giving it the status of a persistent input type. This inflates both the concept count and the code surface.
A cleaner model is: resolve once at a single entry point, return a local path, downstream is unaware.
2. Model input identification needs a single resolver — is_hub_model already exists
hub_utils.py already has is_hub_model() with comprehensive local-path rejection (checks Path.exists(), ./, ../, ~/, Windows drive letters). The new is_hf_onnx_path reimplements only Path.exists(), missing the other cases.
The codebase now has three parallel detection mechanisms with no shared logic:
| Input form | Detector | Local-path rejection |
|---|---|---|
HF model ID (org/model) |
is_hub_model() |
Full (exists, ./, ../, ~/, Win drive) |
| Local ONNX file | scattered path.suffix == ".onnx" checks |
N/A |
Hub-hosted ONNX (org/repo/path.onnx) |
is_hf_onnx_path() (new) |
Partial (only Path.exists()) |
Suggested direction: a single resolve_model_input() that classifies and resolves in one place, reusing is_hub_model's rejection logic. Each command calls it once; adding a fourth input form means changing one function, not 7+.
def resolve_model_input(value: str) -> ModelInput:
# 1. Local-path rejection (reuse is_hub_model logic)
# 2. If org/repo/path.onnx -> download -> return as local_onnx
# 3. If org/model -> return as hf_id
# 4. If local .onnx exists -> return as local_onnx
# No persistent "hub_onnx" type needed3. Pre-quantized ONNX handling should be decided once, not scattered across three pipelines
The detect-and-skip logic lives independently in three places with inconsistent behavior:
| Location | Detects via | Skips optimize | Skips quantize |
|---|---|---|---|
build/onnx.py |
is_quantized_onnx() |
skip_optimize=True |
Explicit |
build/hf.py |
is_quantized_onnx() |
skip_optimize=True |
Explicit |
commands/build.py |
is_quantized_onnx() |
skip_optimize=True |
Relies on user config having quant: null |
Plus _run_quantize_stage has its own is_quantized_onnx() guard (line 951) — a fourth detection point.
The CLI pipeline is the odd one out: if a user provides a config with quant settings for a pre-quantized model, it will attempt to re-quantize. The library functions protect against this; the CLI does not.
Suggested direction: detect once at pipeline entry, stamp the result onto WinMLBuildConfig (e.g. config.skip_optimize = True; config.quant = None), and have all downstream stages read config. This also eliminates redundant is_quantized_onnx calls on the same model.
4. No discovery mechanism for Hub ONNX files
The current UX requires the user to already know the exact file path inside a Hub repo (onnx/vision_encoder_int8.onnx), the right variant (fp32 vs int8, encoder vs decoder), and the task. Compare with the HF model ID flow where architecture, task, and export are all auto-detected.
The only discovery path is the static hub_models.json catalog, which is manually curated. If this is an intentional V1 scope limitation, it should be documented as such. Otherwise, consider accepting a repo ID (onnx-community/sam3-tracker-ONNX) and listing available .onnx files, or reading repo metadata to auto-configure task and roles.
| # is downloaded once and treated as a local .onnx path thereafter. | ||
| from ..loader import maybe_resolve_hf_onnx_path | ||
|
|
||
| model_path = maybe_resolve_hf_onnx_path(str(model_path)) or str(model_path) |
There was a problem hiding this comment.
Dead code: or str(model_path) never triggers. maybe_resolve_hf_onnx_path(str(x)) always returns a non-None string when given a non-None string — it only returns None when the input is None. Since str(model_path) is never None, the or branch is unreachable.
Every other call site in this PR uses the simpler pattern:
model_id = maybe_resolve_hf_onnx_path(model_id)Suggestion: drop the or to match the other sites, or add a comment explaining the defensive intent. Same applies to load_schema_only below (line 383).
| **onnx_kwargs, | ||
| **config.optim, | ||
| ) | ||
| # 1. Optimize (or skip for pre-quantized models) |
There was a problem hiding this comment.
skip_optimize=True + max_optim_iterations > 0 is not guarded. All current callers pair skip_optimize=True with max_optim_iterations=0, so this is not a live bug. However, the function contract allows a caller to pass both skip_optimize=True and max_optim_iterations=3, in which case the autoconf loop would discover flags and call optimize_onnx on a pre-quantized model — the exact crash this fix prevents.
Consider making the invariant self-enforcing:
if skip_optimize:
max_optim_iterations = 0 # re-optimize would crash on pre-quantized models| # Union of all quantization op types (QDQ + QOperator). Use this for | ||
| # "is the model already quantized?" detection regardless of which format | ||
| # the producer used. | ||
| QUANTIZATION_OP_TYPES: frozenset[str] = QDQ_OP_TYPES | QOPERATOR_OP_TYPES |
There was a problem hiding this comment.
Consider adding DynamicQuantizeLinear (and DynamicQuantizeMatMul). DynamicQuantizeLinear is a standard ONNX op (opset 11+) used in dynamic quantization. Models quantized via onnxruntime.quantization with QuantType.QUInt8 in dynamic mode contain this op instead of static QDQ pairs or QOperator fused ops.
Without it, a dynamically-quantized model would not be detected by is_quantized_onnx and would be routed through optimize + quantize. If this is an intentional exclusion (SAM 3 uses static QOperator), a comment noting the limitation would help.
| # Hub-hosted ONNX (e.g. ``onnx-community/sam3-tracker-ONNX/onnx/...``) | ||
| # is downloaded once and treated as a local .onnx path thereafter. | ||
| from ..loader import is_hf_onnx_path, resolve_hf_onnx_path | ||
|
|
There was a problem hiding this comment.
role=path sub-model entries are not wired to the Hub resolver. The role=path branch above (around line 434) validates Path(path).exists() for each sub-model entry but does not call is_hf_onnx_path / resolve_hf_onnx_path. A Hub ref like image-encoder=onnx-community/sam3-tracker-ONNX/onnx/vision_encoder_int8.onnx would fail with "ONNX file not found".
Not blocking for current SAM 3 usage (single-model), but worth a TODO if multi-role eval on Hub-hosted models is planned.
Questions on validation and evaluationThe bug fixes are well-tested (synthetic ONNX models, mock-based CLI tests, regression assertions) — nice work on the silent-skip audit and the QOperator coverage. A few questions on the SAM 3 model validation side — would be great if you could share any additional results you have. What the PR verifies (solid foundation)
These are all great and necessary. A few additional data points would help reviewers feel confident about the end-to-end quality: Would love to see
Task inconsistency between catalog and testsSmall thing I noticed — the encoder's task differs between
The encoder only produces image embeddings, so SuggestionIf some of the above are blocked (NPU availability, no |
SAM 3 (facebook/sam3) requires transformers>=5, but optimum-onnx pins
transformers<4.58, so the standard HF + Optimum export route SAM 2 uses
is blocked. This change wires SAM 3 in through the existing pre-exported
ONNX (Scenario D) pipeline by recognizing path-style Hub references
('org/repo/path/to/file.onnx') and downloading the file once via
huggingface_hub.
Changes:
- New src/winml/modelkit/loader/onnx_hub.py: is_hf_onnx_path,
resolve_hf_onnx_path. Mirrors the is_xxx/resolve_xxx pair pattern
used by is_compiled_onnx/is_quantized_onnx.
- Wire into wmk config, wmk build, and WinMLAutoModel.from_pretrained
with the same 2-line 'if is_hf_onnx_path(x): x = str(resolve_hf_onnx_path(x))'
pattern.
- Add 2 sam3_tracker entries to hub_models.json so 'wmk hub --model-type
sam3_tracker' lists them.
- Tests: 12 unit tests for the resolver, 2 CLI plumbing tests, and 3
end-to-end integration tests (slow/network/integration).
The existing build_onnx_model pipeline runs unchanged on the resolved
local path: the int8 ONNX is auto-detected as quantized via
is_quantized_onnx, the quantization stage is skipped, and the artifact
flows through Optimize -> Analyze<->Optimize -> Compile -> Finalize.
consumer. Also fix two latent bugs in the build pipeline that any QOperator-quantized model would have hit. Background ---------- Issue #324 asks for SAM 2-style native HuggingFace export support for ``facebook/sam3`` (Sam3*IOConfig, Sam3ModelPatcher, etc.). That path is blocked by an upstream constraint: ``optimum-onnx`` pins ``transformers<4.58``, but ``facebook/sam3`` requires ``transformers>=5`` (the ``Sam3Model`` class only exists there). Resolving the pin would need either an upstream optimum-onnx PR or vendoring SAM 3 patcher code that bypasses optimum entirely. Instead, this PR introduces a generic "Hub-hosted ONNX file" input form and lets SAM 3 ride on the existing pre-exported-ONNX (Scenario D) pipeline that already worked for any local ``.onnx`` file. The infrastructure is reusable for any future model with similar version constraints (Whisper / Phi / RWKV / etc. all ship pre-exported ONNX repos on the Hub today). What's added ------------ 1. Hub-hosted ONNX URI resolver - ``loader/onnx_hub.py``: ``is_hf_onnx_path()``, ``resolve_hf_onnx_path()``, ``maybe_resolve_hf_onnx_path()`` - Recognizes inputs of the form ``<org>/<repo>/<path/to/file>.onnx``, downloads via ``huggingface_hub.hf_hub_download``, returns the local cache path. Falls through unchanged for HF model IDs / local paths / ``None``. - Best-effort ``.onnx_data`` sidecar fetch for >2 GiB models. ``EntryNotFoundError`` is expected (inlined weights); ``OSError`` surfaces as a WARNING (disk/permission/network problems should not be silently dropped — the model would later fail to load with a confusing error). 2. CLI wiring (every command that accepts a model identifier) - ``wmk config`` / ``wmk build``: resolve at the top of the command - ``wmk inspect``: friendly "ONNX inspection not yet supported" error for Hub-ONNX refs (matches local .onnx UX) - ``wmk run`` / ``wmk serve``: ``InferenceEngine.load()`` and ``load_schema_only()`` resolve before routing - ``wmk perf``: resolve before the ``Path(model_id).suffix == '.onnx' and exists()`` check (otherwise Hub refs are mistaken for missing local files and rejected with FileNotFoundError) - ``wmk eval``: ``_resolve_model_path`` resolves before the local existence check - ``WinMLAutoModel.from_pretrained``: resolves before HF/ONNX dispatch - Stage-tool commands (``analyze``/``optimize``/``quantize``/ ``compile``/``export``) intentionally NOT wired — they take ``click.Path(exists=True)`` and operate on local files only. 3. SAM 3 catalog entries (``data/hub_models.json``) - Two entries for ``onnx-community/sam3-tracker-ONNX``: the vision encoder and the prompt-encoder + mask-decoder. Note: was already present in the base branch — this PR does not modify it. 4. Integration tests (``tests/integration/test_sam3_e2e.py``) - 4 decoder tests + 2 encoder tests, marked ``@slow @network @integration`` - Asserts: Hub URI resolves, quantization detected, build produces ``model.onnx``, autoconf produces an ``optimization_config``, and for the encoder: pre-quantized round-trip preserves the ``ConvInteger`` / ``MatMulInteger`` ops byte-identically. - Skips narrowed to ``HfHubHTTPError`` / ``OSError`` only — real bugs in the build/analyze pipeline will surface as test failures rather than green skips. Bugs fixed (would affect any QOperator-quantized model, not just SAM 3) --------------------------------------------------------------------- A. ``is_quantized_onnx`` only detected QDQ format (``QuantizeLinear`` / ``DequantizeLinear``). The SAM 3 vision encoder uses ``QuantFormat.QOperator`` (no QDQ pairs, just integer ops: ``ConvInteger``, ``MatMulInteger``, ``QLinear*``). Previously misclassified as not quantized → routed through the optimize + quantize stages → tried to re-quantize an already-int8 model. Fix: ``compiler/utils.py`` adds ``QOPERATOR_OP_TYPES`` and ``QUANTIZATION_OP_TYPES = QDQ ∪ QOperator``. ``onnx/detection.py`` uses the union. B. The ``is_pre_quantized`` branches in ``build_onnx_model``, ``build_hf_model``, and the CLI's ``_build_onnx_pipeline`` logged "skipping optimize" but still invoked ``optimize_onnx`` → ``ort_graph`` → loaded the model into an ORT session. For QOperator models on hosts without a CPU ``ConvInteger`` kernel (e.g. ``onnxruntime-windowsml`` 1.23.x), this crashes the build stage with ``NOT_IMPLEMENTED``. Fix: ``build/common.py::run_optimize_analyze_loop`` gains a real ``skip_optimize: bool`` knob that bypasses ``optimize_onnx`` and the autoconf re-optim loop, just copying the input as the "optimized" artifact. All three pre-quantized branches now pass ``skip_optimize=True``. The downstream behavior (skip quantize + skip compile when configured) is unchanged. Verification ------------ - ``onnx.checker.check_model(full_check=True)`` passes on built artifacts - Built decoder produces NUMERICALLY IDENTICAL outputs to input decoder (``max|built - input| = 0.0`` across all 3 outputs) — pre-quantized round-trip is a true pass-through, not just structurally similar - Encoder runtime feasibility on CPU is identical to input encoder (both fail on CPU because of upstream ORT ``ConvInteger`` kernel gap; encoder requires NPU EP — unchanged from input) - Decoder real inference produces sane SAM-shaped outputs: ``iou_scores ∈ [0, 1]``, ``pred_masks`` logits span both signs, ``object_score_logits`` non-degenerate Test count ---------- - 4518 unit tests pass (+12 new regression tests across: ``test_onnx_hub.py``, ``test_detection.py``, ``test_eval.py``, ``test_perf_cli.py``, ``test_engine.py``) - 6 integration tests pass (live HF download, ~30s) - Ruff check + format clean on all 24 changed files Silent-skip audit (per SAM 2 review feedback) --------------------------------------------- Removed ``except Exception: pytest.skip(...)`` patterns from SAM 3 integration tests — they were swallowing real bugs (including the ``ConvInteger`` regression fixed in this PR). All skips now narrowed to ``HfHubHTTPError`` / ``OSError`` (network) or specific runtime exceptions; ``RuntimeError`` from ``build_onnx_model`` and ``analyze_onnx`` now fails loudly. Removed unnecessary ``pytest.importorskip("huggingface_hub")`` (it's a hard transitive dep). Sidecar download ``OSError`` now logs WARNING instead of DEBUG. Known limitations (not addressed in this PR) -------------------------------------------- - SAM 3 encoder requires NPU EP (QNN / OpenVINO / VitisAI) because ``onnxruntime-windowsml`` ships no CPU kernel for ``ConvInteger(10)``. This is true for both the input and built artifact — our build preserves runtime behavior exactly. Decoder uses ``MatMulInteger`` and runs on either CPU or NPU. - Catalog entries for SAM 3 have ``quantization: null`` so ``wmk perf`` falls back to default random-input shapes that violate the SAM 3 decoder's internal reshape constraints. Populating ``quantization.input_tensors`` with proper shape hints (the pattern every other catalog entry follows) is the recommended fix; out of scope for this PR.
df2f38a to
6e47c07
Compare
6e47c07 to
4464ce6
Compare
| # ``normalize_model_arg`` returns ``str | None`` per its signature; | ||
| # the ``or model`` keeps the narrowed ``str`` type for downstream use. | ||
| hf_model: str = cli_utils.normalize_model_arg(model) or model | ||
| model = hf_model |
4464ce6 to
5bf6fb7
Compare
…on evaluator Follow-up to the SAM 3 / Hub-hosted ONNX commits, addressing inline review feedback from @DingmaomaoBJTU, Copilot review, and CodeQL. Refactor: single model-input classifier and resolver ---------------------------------------------------- New module ``winml.modelkit.utils.model_input`` is the single entry point for classifying a ``-m/--model`` value (one of ``hub_onnx``, ``hf_id``, ``local_onnx``, ``build_dir``, ``invalid``) and resolving it (download for ``hub_onnx``, pass-through otherwise). Replaces the previous scattered detection across ``loader/onnx_hub`` (``is_hf_onnx_path``, ``maybe_resolve_hf_onnx_path``), ``utils/cli`` (``is_onnx_file_path``), and ad-hoc ``Path(value).suffix == ".onnx"`` checks in commands. Callers updated: ``commands/build``, ``config``, ``eval``, ``inspect``, ``perf``; ``inference/engine`` (``load`` and ``load_schema_only``); ``models/auto.WinMLAutoModel.from_pretrained``. Dead code removed ----------------- * ``resolve_model_input(...).local_path or str(model_path)`` -- the ``or`` branch was unreachable; ``local_path`` is set for every resolvable input. * Case-sensitive ``.onnx`` check in ``loader/onnx_hub`` (now consistent with case-insensitive ``.lower()`` check in callers). Build pipeline -------------- * ``ensure_pre_quantized_stamped`` is now the single defensive detection point for pre-quantized models in library entry points; the unified CLI path stamps the config up front, library entry points only run detection when needed. * ``run_optimize_analyze_loop`` enforces ``max_optim_iterations = 0`` whenever ``skip_optimize=True``; ORT lacks kernels for ``ConvInteger`` / ``MatMulInteger`` on host EPs, so re-optimize would crash for the same reason the initial optimize is skipped. * Skip-log message updated to ``"QDQ or QOperator nodes present"`` to match ``is_quantized_onnx`` accepting both quantization formats. * ``compiler/utils.QUANTIZATION_OP_TYPES`` extended with ``DynamicQuantizeLinear`` / ``DynamicQuantizeMatMul``; exported via ``__all__`` to satisfy CodeQL unused-global warning. Mask-generation evaluator (response to "would love to see eval") ---------------------------------------------------------------- * ``winml.modelkit.eval.mask_generation_evaluator`` drives encoder + decoder ORT sessions for SAM-family promptable mask generation. Profile-dispatch design supports SAM 3 (1008 input, mean=std=0.5, direct resize) and SAM 2.1 (1024 input, ImageNet mean/std, longest-side-pad). Verified preprocessing is byte-correct against ``onnx-community/sam3-tracker-ONNX/preprocessor_config.json``. * ``winml.modelkit.datasets.mask_generation`` -- ``MaskGenerationDataset`` wraps ``mattmdjaga/human_parsing_dataset`` with binary foreground/background masks. * ``winml.modelkit.eval.metrics.binary_segmentation`` -- mIoU + Dice on binary masks. * Composite SAM 3 entry in ``models_with_acc.json``. Reference scripts ----------------- * ``scripts/sam3_reference_check.py`` -- spot-check against the published reference ``iou_scores`` from the model card. * ``scripts/mask_generation_eval.py`` -- generic harness for any SAM-family ONNX pair with ``--preset`` (sam2 / sam3) and ``--dataset`` (human_parsing / coco). * ``scripts/sam3_smoke_eval.py`` -- back-compat shim that delegates to the generic harness with the SAM 3 preset. Test cleanups (CodeQL + Copilot) -------------------------------- * ``tests/integration/test_sam3_e2e.py`` fixtures: explicit ``raise`` after ``pytest.skip`` to make control flow obvious. * ``tests/unit/onnx/test_detection.py``: consolidate ``import onnx`` / ``from onnx import ...`` into a single import form. * ``tests/unit/core/test_onnx_utils.py``: expected keys updated for new ``input_symbolic_shapes`` field in ``get_io_config`` output. Other ----- * ``.gitignore``: ignore stray ``<UUID>.data`` ORT external-data sidecars at repo root, ``quantized_info.csv`` Quark side-effect file, and ``scripts/_*.py`` / ``scripts/_*.json`` private debug scripts.
5bf6fb7 to
a0e9c7c
Compare
Thanks for the detailed review and for calling this out. This has been addressed in the latest push, but via a small flow refactor rather than just a string edit: _run_quantize_stage no longer carries the old “(QDQ nodes already present)” wording path. Pre-quantized detection is now decided earlier and stamped on config (skip_optimize=True, quant=None), so _run_quantize_stage exits on config.quant is None for both QDQ and QOperator cases. For consistency, the stage-level logs in build/onnx.py and build/hf.py now use “pre-quantized model” wording (covering QDQ + QOperator). If you prefer an explicit QDQ/QOperator mention in the commands/build.py skip path as well, I can add that as a follow-up log line too. |
Thanks for the detailed consolidated feedback. This is very helpful. On the quantize-stage wording nit, I agree on the consistency goal. In the current branch, the old skip-message path you referenced is no longer the active behavior. The quantize stage now exits based on config state (quant is None) for pre-quantized inputs, and the pre-quantized wording is handled in the build pipelines. See src/winml/modelkit/commands/build.py, src/winml/modelkit/build/onnx.py, and src/winml/modelkit/build/hf.py. On design point 1 (Hub-hosted ONNX as download step): I agree with what you suggested. The current implementation is now centered on single-point normalization and resolution, where entry points resolve to local paths and downstream pipeline stages operate on local ONNX paths. See src/winml/modelkit/utils/cli.py and src/winml/modelkit/utils/model_input.py. |
This is addressed in the latest revision by introducing a single resolver path and routing command entry points through it. We now use src/winml/modelkit/utils/model_input.py as the classification and resolution center, and src/winml/modelkit/utils/cli.py normalize_model_arg delegates to resolve_model_input so callers do not have to implement their own Hub-vs-local detection. Also, local-path rejection is no longer limited to Path.exists-style checks; model_input reuses shared local-path logic via _is_local_path from src/winml/modelkit/utils/hub_utils.py, which covers the ./, ../, ~/, and Windows-drive cases you listed. On the hub_onnx type: agreed on the principle that downstream should operate on local ONNX after resolution. In the current shape, hub_onnx is an internal classification state used before resolution; after resolve_model_input, downstream paths consume local_path (for example src/winml/modelkit/inference/engine.py and src/winml/modelkit/models/auto.py). |
This is addressed in the latest revision. We now decide pre-quantized status once and stamp it onto config, then downstream stages read config state instead of re-detecting: A shared stamping helper sets skip_optimize true and clears quant when the input is pre-quantized (or when skip_optimize is forced). |
You are right that the current Hub ONNX UX is still artifact-path driven: users must provide a fully qualified file path, and we do not yet support repo-only input or automatic task and role inference from repo metadata. So the limitation you described is real. What this PR does improve is the failure and guidance path: when a Hub ONNX path is incorrect, we now surface available ONNX files from that repo in the error hint, so users can recover without guessing blindly. That is a usability improvement, but it is not full discovery. Below is the status as of now:
I will track the full discovery work as follow-up so this does not remain an implicit V1 constraint. |
I ran an informal offline mask-generation eval locally and can share a baseline: Model: onnx-community/sam3-tracker-ONNX (int8 encoder and decoder) mIoU: 0.9494 |
Thanks for pointing this. This support has been added in this PR. The eval path now includes a dedicated mask-generation evaluator and binary-segmentation metrics: Task routing for mask-generation is wired in src/winml/modelkit/eval/evaluate.py. |
Yes, I do have preliminary perf numbers, but they were collected via manual ORT benchmarking scripts rather than wmk perf because of the current shape-hints limitation you mentioned. What I measured so far: Scope: SAM 3 encoder ONNX (pixel_values shape [1,3,1008,1008]), not full end-to-end tracker pipeline. |
I validated against the published onnx-community/sam3-tracker-ONNX reference example (same image and prompt) and confirmed the pipeline is producing close outputs, but I did not run a direct facebook/sam3 PyTorch-to-ONNX parity comparison in this change set. For the reference spot-check, the IoU scores were: Reference: [0.931315, 0.037516, 0.512856] |
Thanks for raising these, In the current branch I have addressed the following:
|
This PR’s functional pieces are done and passing (Hub-ONNX resolution, pre-quantized QDQ/QOperator handling, and test coverage), but I can’t say we have stable end-to-end NPU perf yet in this environment. When I reproduced the suggested flow (winml.modelkit sys + winml.modelkit perf -m microsoft/resnet-50 --device npu --monitor --verbose), sys detected AMD NPU hardware, but perf failed because EP setup did not complete (MIGraphXExecutionProvider download/install failed, resulting EPs were CPU + DML only, then ValueError: Device 'npu' requested but no compatible EP is available). I do have script-based measurements from a different path: SAM3 encoder run (VitisAI + CPU fallback) showed min 32172 ms, p50 33406 ms, mean 33621.8 ms, max 35109 ms; CPU baseline for the same encoder was min 21734 ms, p50 23532 ms, mean 25097.0 ms, max 34219 ms; and ResNet50 script run (VitisAI + CPU fallback) averaged 35.87 ms (min 33.08 ms). We have preliminary mixed-provider numbers from custom scripts, but not a clean reproducible winml perf NPU benchmark in this setup; given that, I propose we merge this PR for correctness and move full end-to-end NPU validation/benchmarking to a dedicated follow-up PR. |
DingmaomaoBJTU
left a comment
There was a problem hiding this comment.
PR Review: SAM3 ONNX Hub Support (#324)
This PR is well-structured with a clean ModelInput classifier/resolver design, robust sidecar handling, and excellent error enrichment for missing files. Several issues need attention before merge.
See inline comments for details.
| "model_id": "onnx-community/sam3-tracker-ONNX/onnx/vision_encoder_int8.onnx", | ||
| "task": "image-feature-extraction", | ||
| "model_type": "sam3_tracker", | ||
| "supported_eps": {}, |
There was a problem hiding this comment.
Bug: supported_eps: {} makes these entries invisible to EP-based catalog lookups.
catalog.py line 112 computes {normalize_ep_name(k) for k in (m.get('supported_eps') or {})} — an empty dict yields an empty set, so wmk list --ep cpu (or any EP filter) will never surface these models. They should list at minimum CPU support:
json "supported_eps": {"cpu": ["CPU"]}
All other models in this file have populated supported_eps.
There was a problem hiding this comment.
You're right, empty supported_eps: {} made them invisible to catalog filters. I've corrected both SAM3 entries to separate model_id (org/repo) from onnx_path (file path) and added "supported_eps": {"cpu": ["CPU"]}.
Now wmk list --ep cpu will surface them correctly.
There was a problem hiding this comment.
The fix doesn't appear to be reflected in the current diff yet — supported_eps: {} is still empty for both SAM 3 entries in hub_models.json. Could you double-check that the change was pushed?
There was a problem hiding this comment.
Thanks for the catch! The fix is now committed. Both SAM3 entries in hub_models.json now have "supported_eps": {"cpu": ["CPU"]} properly set. The models will surface correctly in catalog filters.
DingmaomaoBJTU
left a comment
There was a problem hiding this comment.
These SAM 3-specific scripts (sam3_smoke_eval.py, sam3_reference_check.py, sam3_decoder_shapes.json, mask_generation_eval.py) are placed directly under scripts/, which will get harder to navigate as more models are onboarded.
Would it be worth moving them into scripts/models/sam3/? That would set a clean precedent for future model-specific scripts (e.g. scripts/models/florence2/, scripts/models/phi4/) and keep the top-level scripts/ uncluttered.
| }, | ||
| "size_mb": 157.2 | ||
| }, | ||
| { |
There was a problem hiding this comment.
Note: hub_models.json currently lists only built-in models that have been validated across all 10 EPs with passing perf and accuracy benchmarks. SAM 3 entries may be updated here once it's been tested on all target platforms.
There was a problem hiding this comment.
Got it, SAM 3 Hub model entries are currently included for functional availability. Full validation and performance benchmarking across all 10 execution providers (EPs) is planned as a follow-up PR.
The current implementation follows the established pattern in hub_models.json and is ready for initial integration and testing.
| from ..quant import quantize_onnx | ||
| from ..utils.console import StageLive | ||
|
|
||
| if config.quant is None: |
There was a problem hiding this comment.
Bug: pre-quantized models may be double-quantized when a config file is loaded via -c.
This PR removes the is_quantized_onnx(current_path) guard that previously sat between the config.quant is None check and the actual quantize call, replacing it with the invariant that config.quant is always None for pre-quantized models.
That invariant holds for the auto-generated config path (generate_onnx_build_config + �nsure_pre_quantized_stamped), but not for the file-loaded config path (-c config.json). When a user passes a hand-authored or previously exported config that has a quant section, _build_onnx_pipeline calls neither generate_onnx_build_config nor �nsure_pre_quantized_stamped, so config.quant is never cleared. The function then falls through to the �lif config.quant is not None branch and attempts to quantize an already-quantized model.
Suggested fix: either call �nsure_pre_quantized_stamped(config, onnx_path) in _build_onnx_pipeline before the quantize stage (mirroring �uild_onnx_model), or add a cheap guard here: if config.skip_optimize: return current_path.
There was a problem hiding this comment.
This is addressed by adding the guard at the quantize stage.
This prevents re-quantization of pre-quantized models loaded via -c config.json, even if the config file contains a quant section. The skip_optimize flag is properly set by ensure_pre_quantized_stamped() for the auto-generated path and remains the source of truth for the file-loaded path.
| if i == 0: | ||
| sym_name = sym[i] if i < len(sym) else None | ||
| if isinstance(sym_name, str) and sym_name in overrides: | ||
| resolved.append(int(overrides[sym_name])) |
There was a problem hiding this comment.
Bug: TypeError crash when a shape_config key matches both an input name and a symbolic dim name.
The Form 1 check at the top of _resolve_shape guards against non-list values entering the per-input-name branch (isinstance(overrides[input_name], (list, tuple))), but the symmetric guard is absent here at the symbolic dim branch.
Scenario: shape_config = {"image_embeddings": [1, 256, 64, 64]} where "image_embeddings" is both an input name and a dim_param on a different input. Form 1 correctly returns the shape for the direct-name input. But for any other input whose sym[i] == "image_embeddings", this line executes int([1, 256, 64, 64]) and raises TypeError: int() argument must be a string, a bytes-like object or a real number, not 'list' with no useful diagnostic for the user.
Suggested fix: add a type guard before the int() cast:
python if isinstance(sym_name, str) and sym_name in overrides and isinstance(overrides[sym_name], (int, float, str)): resolved.append(int(overrides[sym_name]))
or wrap in ry/except TypeError with a clear error message.
There was a problem hiding this comment.
Added the type guard to prevent TypeError when a shape_config key matches both an input name and a symbolic dim name.
This ensures only scalar values (int, float, str) are accepted at the symbolic dimension level, matching the guard already in place for direct input-name lookups. If a user mistakenly passes a list value for a dimension key, it's safely skipped rather than crashing.
| composite ``role=path`` model dict (currently only | ||
| ``mask-generation``), returns ``None`` -- the evaluator reads | ||
| ``config.model_path`` directly. Going through ``WinMLAutoModel``'s | ||
| composite registry would require registering the model type (e.g., |
There was a problem hiding this comment.
Could you please share more details why you cannot use WinMLComposite model's from_onnx method to load the multiple models?
In our design principle, we want to use the model class to encapsulate the ort session and the model inference. Therefore, the evaluate can take a model class and run it.
If the today's composite model cannot fulfill the requirement, you can create another model class for sam3 which can manage the session. So the evaluator code just need to run the model and focus on metric calculation logic.
There was a problem hiding this comment.
I chose the direct session management approach for this PR because SAM3's evaluation pipeline requires custom orchestration (encoder to mask decoder with specific data prep) that doesn't fit neatly into a general purpose composite model class. It also keeps the evaluator focused on metrics and makes testing easier.
I agree that the proper 'SAM3Composite' class that encapsulates encoder/decoder sessions is the right long-term approach and aligns with your design principle. But building that requires designing a reusable interface, integrating with 'WinMLAutoModel''s registry, and testing across all EPs, that's substantial enough work to be added to this already exhaustive PR, I plan to add it as a separate PR and merge this one cleanly.
So the plan is that this PR delivers Hub support with pragmatic choices, and the next PR will refactor to a proper composite model architecture. This keeps each PR focused and ensures the architectural work gets proper attention.
| sess_opts = ort.SessionOptions() | ||
| sess_opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL | ||
|
|
||
| enc = ort.InferenceSession( |
There was a problem hiding this comment.
We typically encapsulate the model inference session inside the model class's forward method. If needed, you can create a model class for the sam3 (or for the mask generation task in general, which is better)
There was a problem hiding this comment.
Agreed, encapsulating session management inside a model class's forward method is the right approach. As mentioned above, this PR uses direct session management, but the follow-up PR will create a proper MaskGenerationModel (or SAM3Composite) class that handles encoder/decoder session orchestration, allowing the evaluator to simply call the model's forward method and focus on metrics.
| ``mask-generation``), returns ``None`` -- the evaluator reads | ||
| ``config.model_path`` directly. Going through ``WinMLAutoModel``'s | ||
| composite registry would require registering the model type (e.g., | ||
| SAM 3), which is a heavier follow-up; this bypass lets standalone |
There was a problem hiding this comment.
The comment notes that winml eval with composite
ole=path inputs (encoder + decoder) isn't fully wired up yet for mask-generation — the evaluator bypasses WinMLAutoModel's composite registry and reads config.model_path directly. Is there a follow-up PR planned to close this gap so winml eval -m image-encoder=... -m prompt-decoder=... --task mask-generation works end-to-end through the standard eval path?
There was a problem hiding this comment.
Yes, right! the gap is part of the broader composite model refactoring planned for the follow-up PR.
Once the proper SAM3Composite (or MaskGenerationModel) class is registered in WinMLAutoModel's composite registry, the full eval flow will work end-to-end:
winml eval -m image-encoder=... -m prompt-decoder=... --task mask-generation
Instead of bypassing the registry, the evaluator will consume the model class through the standard path. That PR will wire everything together so composite models are treated uniformly across build, perf, and eval commands.
| # ``get_available_providers()`` but not in ``get_ep_devices()``. Use the | ||
| # legacy ``SessionOptions.add_provider`` path for those. ``add_provider`` | ||
| # takes ``dict[str, str]`` so coerce values. | ||
| if ep_name in ort.get_available_providers(): |
There was a problem hiding this comment.
This legacy fallback for VitisAIExecutionProvider (onnxruntime-vitisai 1.23.x) adds AMD-specific branching to the general-purpose session layer. Since we're not planning to support legacy AMD, would it make sense to drop this block entirely, along with _LEGACY_EP_DEVICE_FALLBACK in sysinfo/device.py and the for_vitisai factory in compiler/configs.py?
There was a problem hiding this comment.
The legacy VitisAI fallback and add_provider path have been removed. winml.py now uses the modern add_provider_for_devices API uniformly for all EPs, including VitisAI, with no special-case branching.
| _WIN_DRIVE_RE = re.compile(r"^[A-Za-z]:[\\/]") | ||
|
|
||
|
|
||
| def _is_local_path(value: str | None) -> bool: |
There was a problem hiding this comment.
_is_local_path is a general filesystem heuristic with no Hub-specific logic, but it lives in hub_utils.py and is already imported across module boundaries by model_input.py. Would it make more sense to move it to a more generic util module (e.g. cli.py or a new path_utils.py) so the dependency direction is clearer — hub_utils imports from utils, not the other way round?
There was a problem hiding this comment.
The generic path heuristic has been moved to src/winml/modelkit/utils/path_utils.py, and both hub_utils.py and model_input.py now import it from there. hub_utils.py only keeps a temporary _is_local_path alias for backward compatibility with existing callers.
| @@ -211,10 +211,24 @@ def for_openvino(cls, device: str | None = None) -> WinMLCompileConfig: | |||
|
|
|||
| @classmethod | |||
| def for_vitisai(cls, device: str | None = None) -> WinMLCompileConfig: | |||
There was a problem hiding this comment.
This Vitis AI-specific factory adds EP-specific logic (RYZEN_AI_INSTALLATION_PATH, xclbin path construction, xlnx_enable_py3_round) into what should be a generic compile config layer. Since we're not planning to support legacy AMD, would it make sense to drop this entirely, along with _LEGACY_EP_DEVICE_FALLBACK in sysinfo/device.py and the legacy add_provider fallback in winml.py?
There was a problem hiding this comment.
The VitisAI-specific compile factory and legacy AMD fallback paths have been removed. compiler/configs.py now keeps the compile layer generic, and winml.py no longer relies on the legacy add_provider path.
| # Models exported through ``onnxruntime.quantization`` with | ||
| # ``QuantFormat.QOperator`` (or sourced from Hub repos like | ||
| # ``onnx-community/sam3-tracker-ONNX``) use this format. | ||
| QOPERATOR_OP_TYPES: frozenset[str] = frozenset( |
There was a problem hiding this comment.
QOPERATOR_OP_TYPES and DYNAMIC_QUANT_OP_TYPES are pure ONNX op-type knowledge with no compiler-specific logic, but they live in compiler/utils.py. The only consumer is is_quantized_onnx() in core/onnx_utils.py, which means core now depends on compiler — the wrong direction. Would it make more sense to define these constants (and the existing QDQ_OP_TYPES) directly in core/onnx_utils.py, or a new core/quantization.py, so the dependency flows the right way?
There was a problem hiding this comment.
The quantization op-type constants have been moved to core/quantization.py, and compiler/utils.py now only re-exports them for compatibility. This keeps the dependency direction correct and avoids core depending on compiler.
| """ | ||
|
|
||
| DEFAULT_DATASET = "mattmdjaga/human_parsing_dataset" | ||
| DEFAULT_SPLIT = "train" |
There was a problem hiding this comment.
Nit: other dataset implementations (e.g. ImageSegmentationDataset) also default to the train split, so this is consistent with existing practice. Presumably these datasets only ship a train split, or the intent is smoke-test / functional verification rather than strict accuracy benchmarking. If so, a short comment here would help clarify the intent.
There was a problem hiding this comment.
The class now includes a short comment explaining that DEFAULT_SPLIT = "train" is intentional because this dataset only ships a train split and the evaluator is meant for smoke-test / functional verification rather than strict accuracy benchmarking. See src/winml/modelkit/datasets/mask_generation.py.
| # exposed via the new ``get_ep_devices()``/AutoEP machinery. Mapped to the | ||
| # canonical device they target so they can still be selected via the legacy | ||
| # ``SessionOptions.add_provider`` code path. | ||
| _LEGACY_EP_DEVICE_FALLBACK: dict[EPName, str] = { |
There was a problem hiding this comment.
Since we're not planning to support legacy AMD (VitisAI on older onnxruntime-vitisai 1.23.x), would it make sense to drop _LEGACY_EP_DEVICE_FALLBACK and the corresponding fallback in winml.py entirely? Keeping it adds AMD-specific branching to the general session layer with no active consumer.
There was a problem hiding this comment.
Already addressed. The legacy fallback map has been removed, and winml.py no longer has a legacy add_provider branch for VitisAI. The session layer now relies on the standard get_ep_devices() / add_provider_for_devices() path only.
Issue:
|
| # of treating the path as a HuggingFace repo id (which would try to | ||
| # load the .onnx file as a JSON config and crash). | ||
| if cli_utils.is_onnx_file_path(model_id): | ||
| config_or_configs = generate_build_config( |
| device=device, | ||
| ) | ||
| else: | ||
| config_or_configs = generate_build_config( |
All SAM3-specific scripts have been moved to scripts/models/sam3/, including: mask_generation_eval.py |
--shape-config is now allowed for ONNX files and is applied during dummy input generation, which lets SAM3 ONNX benchmarks override dynamic dims that the positional defaults can’t satisfy. |
DingmaomaoBJTU
left a comment
There was a problem hiding this comment.
Re-review: SAM3 ONNX Hub Support (Issue #324)
This re-review checks whether previously-raised issues were addressed in code, and surfaces additional findings. All 6 previously-raised issues remain unresolved in code (author replied in comments but made no corresponding commits). New issues are also noted below.
Issues not fixed since initial review
hub_models.json:supported_eps: {}is still empty for both SAM3 entries (catalog EP filtering broken)hub_models.json:model_idstill uses full ONNX path instead oforg/reposchemaonnx_hub.py: hint URL is still hardcoded to/tree/mainregardless ofrevisioneval.py: Hub download failures in_resolve_model_pathare still unhandledmask_generation_evaluator.py: bare dict key access on hardcoded embedding names is still presentmask_generation_evaluator.py: textprompt_modestill silently fails per-sample
New findings
perf.py: Same unhandled exception pattern aseval.pyfornormalize_model_argbuild.py(_build_onnx_pipeline): readsconfig.skip_optimizebut never callsensure_pre_quantized_stamped, allowing double-quantization when-c config.jsonwithquant != Noneis passed with a pre-quantized ONNX model
| }, | ||
| { | ||
| "model_id": "onnx-community/sam3-tracker-ONNX/onnx/vision_encoder_int8.onnx", | ||
| "task": "image-feature-extraction", |
There was a problem hiding this comment.
[Bug — not fixed] model_id is set to the full ONNX file path (org/repo/path/file.onnx) rather than the org/repo form that every other entry in this file uses. This breaks api.model_info(model_id) calls and catalog display. The ONNX-specific path should be stored in a separate onnx_path field (matching build/hf.py which reads model_info.get('onnx_path')), and model_id should be onnx-community/sam3-tracker-ONNX.
| "task": "image-feature-extraction", | ||
| "model_type": "sam3_tracker", | ||
| "supported_eps": {}, | ||
| "size_mb": 504.2 |
There was a problem hiding this comment.
[Bug — not fixed] supported_eps is an empty object {}. Catalog lookups use {normalize_ep_name(k) for k in (m.get('supported_eps') or {})}, so an empty dict makes this entry invisible to all EP-based filters (wmk list --ep cpu, etc.). Every other model entry in this file has populated supported_eps. Minimum: {"cpu": ["CPU"]}.
| "task": "mask-generation", | ||
| "model_type": "sam3_tracker", | ||
| "supported_eps": {}, | ||
| "size_mb": 9.6 |
There was a problem hiding this comment.
[Bug — not fixed] Same issue: supported_eps: {} for the second SAM3 entry. Both entries are invisible to EP catalog filters.
| return ( | ||
| f"Could not list available .onnx files in '{repo_id}' " | ||
| f"(see https://huggingface.co/{repo_id}/tree/main)." | ||
| ) |
There was a problem hiding this comment.
[Warning — not fixed] The error hint URL is hardcoded to /tree/main regardless of the revision parameter. If a user pins a specific revision (e.g., onnx-community/sam3-tracker-ONNX@abc1234/onnx/file.onnx), the hint points to main which may not contain the file. Fix:
f"(see https://huggingface.co/{repo_id}/tree/{revision or 'main'})."Same issue applies to the fallback hint two lines below.
| param_hint="-m/--model", | ||
| ) | ||
| path = cli_utils.normalize_model_arg(path) or path | ||
| if not Path(path).exists(): |
There was a problem hiding this comment.
[Warning — not fixed] cli_utils.normalize_model_arg() can raise FileNotFoundError, RepositoryNotFoundError, or OSError when Hub download fails, but there is no try/except wrapping these calls in _resolve_model_path. The user will see a raw Python traceback. Contrast with build.py line 587 which wraps the same call inside a try/except Exception with actionable error messages. Add a try/except (FileNotFoundError, OSError) as e: raise click.ClickException(str(e)) guard around both normalize_model_arg calls here.
| # Hub-hosted ONNX (e.g. ``onnx-community/sam3-tracker-ONNX/onnx/...``) | ||
| # is downloaded once and treated as a local .onnx path thereafter. | ||
| value = cli_utils.normalize_model_arg(value) or value | ||
| if not Path(value).exists(): |
There was a problem hiding this comment.
[Warning — not fixed] Second normalize_model_arg call (for multi-model eval paths) is also missing exception handling. Same fix needed as the call above.
| # ``normalize_model_arg`` returns ``str | None`` per its signature; | ||
| # the ``or model`` keeps the narrowed ``str`` type for downstream use. | ||
| hf_model: str = cli_utils.normalize_model_arg(model) or model | ||
| model = hf_model |
There was a problem hiding this comment.
[Warning — new] normalize_model_arg at this line has no surrounding try/except. If the Hub download fails (network error, repo not found, gated repo), the exception propagates as an unhandled traceback. The perf command should wrap this call (and/or the entire body of the perf() function) with appropriate exception handling, as build() does.
| "input_labels": labels, | ||
| "input_boxes": box, | ||
| "image_embeddings.0": emb["image_embeddings.0"], | ||
| "image_embeddings.1": emb["image_embeddings.1"], |
There was a problem hiding this comment.
[Bug — not fixed] Bare dict key access on hardcoded SAM3 output names. If the encoder produces differently-named outputs (e.g., a future version or a SAM3 variant that uses image_embedding_0 instead of image_embeddings.0), this raises a raw KeyError with no context. Use .get() with a clear error, or validate encoder output names against these keys at session creation time:
for key in ("image_embeddings.0", "image_embeddings.1", "image_embeddings.2"):
if key not in emb:
raise ValueError(f"Encoder output '{key}' not found. Got: {list(emb.keys())}")| self.config = config | ||
| mapping = config.dataset.columns_mapping or {} | ||
| self._prompt_mode: str = mapping.get("prompt_mode", "bbox") | ||
| self._enc_sess, self._dec_sess = self._load_sessions() |
There was a problem hiding this comment.
[Info — not fixed] prompt_mode='text' is accepted here and stored, but _build_decoder_inputs raises ValueError for any mode that is not 'bbox' or 'point'. This means text mode will silently fail per-sample (the except Exception in compute() swallows the error), yielding mIoU=0.0, num_samples=0 with no top-level error message. Either reject unsupported modes in __init__ with a clear ValueError, or document that only 'bbox' and 'point' are supported.
| # (here and inside ``build_onnx_model``) read the flag instead of | ||
| # re-running ``is_quantized_onnx`` on the same file. | ||
| is_pre_quantized = config.skip_optimize | ||
|
|
There was a problem hiding this comment.
[Bug — new] _build_onnx_pipeline reads config.skip_optimize to detect pre-quantized models but never calls ensure_pre_quantized_stamped() (which is defined in build/common.py and also clears config.quant). If a user runs wmk build -m pre_quantized_int8.onnx -c config.json and config.json has quant != null, the model will be double-quantized: the optimize stage runs on a pre-quantized model (since config.skip_optimize defaults to False), then the quantize stage also runs (since config.quant was never cleared). Fix: call ensure_pre_quantized_stamped(current_path, config) immediately after copying the model to output_dir (around line 1511), before either the optimize or quantize stages.
| SAM3_PROFILE = MaskGenProfile( | ||
| name="sam3", | ||
| encoder_ref="onnx-community/sam3-tracker-ONNX/onnx/vision_encoder_int8.onnx", | ||
| decoder_ref="onnx-community/sam3-tracker-ONNX/onnx/prompt_encoder_mask_decoder_int8.onnx", |
There was a problem hiding this comment.
[Bug — DML precision failure, tested] The int8 preset is hardcoded for both CPU and DML (--ep dml), but DML does not produce correct masks for this model. Tested on Qualcomm Adreno X1-85 (Snapdragon X Elite):
| EP | Model | mIoU | Dice | enc s/img |
|---|---|---|---|---|
| CPU | vision_encoder_int8.onnx |
0.9175 ✅ | 0.9559 | ~96s |
| DML | vision_encoder_int8.onnx |
0.0000 ❌ | 0.0000 | ~50s |
| DML | vision_encoder_fp16.onnx |
0.0000 ❌ | 0.0000 | 7.1s |
The int8 failure is because DmlExecutionProvider does not support QLinearConv/ConvInteger operators — they silently produce wrong outputs instead of erroring. The fp16 failure is a separate numerical precision issue (decoder pred_masks are all-negative after DML inference, yielding empty masks after sigmoid + threshold), despite pIoU≈0.84 internally.
Fix options:
- Switch the default DML preset to
fp32(vision_encoder.onnx) — DML supports fp32 natively, and the fp16 encoder is 13× faster than CPU already shows DML GPU scheduling works. - Add a DML-specific profile or a runtime warning when
--ep dmlis used with thesam3preset.
For now, the sam3 preset should document that only --ep cpu is validated, or the script should warn when DML is selected with an int8/fp16 model.
Adds SAM 3 support and introduces a generic Hub-hosted ONNX input form (
<org>/<repo>/<path>.onnx) that downloads pre-exported ONNX files from HuggingFace viahuggingface_hub. SAM 3 is the first consumer.The standard SAM 2-style export route is blocked:
optimum-onnxpinstransformers<4.58, butfacebook/sam3requirestransformers>=5. This PR uses the pre-exportedonnx-community/sam3-tracker-ONNXartifacts via the existing Scenario D pipeline.Changes
loader/onnx_hub.py)wmk config,wmk build,wmk inspect,wmk run,wmk serve,wmk perf,wmk evalis_quantized_onnxto detect QOperator format (ConvInteger,MatMulInteger,QLinear*) — was QDQ-onlyis_pre_quantizedbuild branch to truly skip the optimize stage (previously crashed on QOperator models)Tests
Limitations
ConvIntegerkernel inonnxruntime-windowsml); decoder runs on CPU + NPU.