Summary

Adds explicit CPU fp32 and fp16 recipes for valentinafevu/yolos-fashionpedia, a YOLOS object-detection checkpoint for apparel and accessory detections. This is an Effort L0 / Outcome L0 recipe-only contribution because current main already exports the model without a recipe; the recipes record verified precision coverage and the pinned Fashionpedia evaluation contract. Full required-tuple coverage reached the committed highest Goal with L3 PASS.

Model metadata

1. What the model does — A YOLOS vision-transformer checkpoint fine-tuned on Fashionpedia that accepts a fashion image and predicts up to 100 objects as 46 apparel/accessory categories with normalized bounding boxes. Evidence: checkpoint model card and config at revision 0cfdb77c01cfbf0838b940ecff9c572f84c9d4f4, YolosForObjectDetection source, and ONNX outputs logits [1,100,47] and pred_boxes [1,100,4]. Confidence: verified.
2. Primary user stories — A user supplies a fashion image to obtain apparel/accessory category detections and bounding boxes for cataloging, visual search, or fashion-image analysis. Evidence: the checkpoint model card describes a Fashionpedia fine-tuned object detector, and the pinned config contains the Fashionpedia label map. Confidence: mapped.
3. Supported tasks — object-detection across the checkpoint, Transformers, Optimum ONNX, and WinML surfaces. Evidence: Hub pipeline_tag, YolosForObjectDetection architecture, Optimum registration probe, and WinML inspect/config on current main. Confidence: verified.
4. Model architecture — YOLOS-Small object detector: 16×16 patch projection into a 384-wide token sequence, 100 learned detection tokens, twelve 6-head transformer encoder layers with 1536-wide GELU MLPs, and separate class and box heads.

YolosForObjectDetection
├── YOLOS vision transformer (YolosModel)
│   ├── Patch projection (Conv2d, 3→384, kernel/stride 16)
│   ├── CLS token + 100 detection tokens + position embeddings
│   ├── Transformer encoder layer × 12
│   │   ├── 6-head self-attention (hidden size 384)
│   │   ├── Feed-forward (384→1536→384, GELU)
│   │   └── Residual + LayerNorm
│   └── Final LayerNorm
├── Class head (384→47: 46 labels + no-object)
└── Box MLP head (384→384→384→4, sigmoid)

• Source/confidence: pinned checkpoint dimensions, Transformers YolosForObjectDetection source, and built ONNX scopes/I/O (mapped).

Validation and support evidence

Baseline

• Replacement charter revision 2 pins current main at 26a4a82b638f2169b5a2f75758abb31a60b1774e; WinML 0.2.0; checkpoint revision 0cfdb77c01cfbf0838b940ecff9c572f84c9d4f4.
• Starting auto-config: PASS with task object-detection, AutoModelForObjectDetection, input [1,3,512,864], outputs logits and pred_boxes, requested opset 17, and current-default export.dynamo=true.
• Recipe-free build: PASS in 72.65718199999999 s; Dynamo could not lower requested opset 17 and retained actual opset 18; the optimized artifact has 479 nodes, with 197/225 modules recovered and 512/512 pre-optimization exported nodes tagged. Public I/O is pixel_values [1,3,512,864] float32 to logits [1,100,47] and pred_boxes [1,100,4].
• Baseline CPU fp32 perf: PASS in 24.2293859 s; mean 390.159 ms, p50 385.305 ms, throughput 2.56 samples/s, RSS total delta 413.77 MB, local/shared VRAM total delta 0.0/0.0 MB.
• Baseline pinned Fashionpedia eval: PASS in 97.80928689999999 s on 10 deterministic, unshuffled, streaming images from the val split of detection-datasets/fashionpedia revision 80845435ce686b8a9dbf70a05452fbfb8e09cdd7; map=0.008183383382856846, map_50=0.03340508043766022, and map_75=0.0009040034492500126.
• Optimum's yolos registration exposes feature-extraction and object-detection both before and after WinML registration, with no WinML-added task (VENDOR-ONLY). The current-main recipe-free Dynamo build/perf/eval raises the baseline Goal floor to L3.
• Runtime measurements used a host-only fail-closed catalog workaround after the auto-installed OpenVINO 1.8.15.0 plugin failed because openvino.dll was missing. CPUExecutionProvider behavior and task semantics were unchanged.

Goal

• Committed axes: Effort L0 / Goal ceiling L3 / Outcome L0.
• Charter revision 2 superseded revision 1 after main moved from 5deebd422e95f28fe8fd912ee50ce874710187b3 to 26a4a82b638f2169b5a2f75758abb31a60b1774e and Dynamo became the default exporter. Effort, Goal ceiling, Outcome, and the CPU fp32/fp16 precision plan did not change; the baseline Goal floor changed from L0 to L3, and both recipes were refreshed from the dynamo=true starting config.
• Success definition: both required CPU fp32/fp16 tuples freshly pass L0 build/structure/precision, L1 perf, independent pinned named-input PyTorch parity at L2, and pinned immutable Fashionpedia task evaluation at L3.

Outcome

• Shipped tier: L0; highest Goal verdict: L3 PASS; coverage: full; deferred tuples: none; no ceiling downgrade.
• Shipped recipes: examples/recipes/valentinafevu_yolos-fashionpedia/cpu/cpu/object-detection_fp32_config.json and examples/recipes/valentinafevu_yolos-fashionpedia/cpu/cpu/object-detection_fp16_config.json. No source, test, skill, or recipe README files changed.
• Final evidence is bound to PR head 587dc4940ccc55052f92cfcf7471c483abbfbbee, matching the pushed remote head. Both artifacts preserve named I/O, pass deterministic PyTorch parity, and complete the pinned Fashionpedia evaluation with equal metrics.
• Learner findings retained in the head-bound handoff: current-default Dynamo removes the old legacy-exporter requirement for this checkpoint; CPU fp16 halves artifact size but is slower and uses more RSS than fp32 on this tested host while preserving task output; and static QNN partial support identifies a portability target, not runtime non-CPU support. No generalized model-knowledge or methodology finding was appended, so no Lane A change was created.
• Quality: 8/8 PASS at the bound head — license insertion, Ruff, and mypy (409 source files) passed; test groups passed with 1441 passed, 45 skipped, 1374 passed, 6 skipped, 1 xfailed, 547 passed, 16 skipped, 1 xfailed, 2381 passed, 8 skipped, 2 deselected, 1 warning, and 895 passed, 1 skipped, 1 deselected.

Per-EP/device/precision results — including perf and eval data

Goal ladder

Perf

L2 named-output parity

Eval

Dataset for both rows: detection-datasets/fashionpedia at revision 80845435ce686b8a9dbf70a05452fbfb8e09cdd7, no subset, split val, 10 deterministic unshuffled streaming samples.

Delta

• Both shipped recipes retain the replacement-charter starting config's /export/dynamo = true; the old /export/dynamo = false legacy-exporter pin is removed.
• Relative to the current-main Dynamo auto-config, both recipes add /eval from missing to the pinned object-detection contract: detection-datasets/fashionpedia revision 80845435ce686b8a9dbf70a05452fbfb8e09cdd7, split val, 10 samples, shuffle=false, seed 42, streaming=true, input image, annotations objects, bbox key bbox, category key category, absolute xywh boxes.
• The fp32 recipe has no other JSON-pointer delta.
• The fp16 recipe also changes /quant from null to {mode: fp16, samples: 10, calibration_method: minmax, weight_type: uint8, activation_type: uint8, per_channel: false, symmetric: false, weight_symmetric: null, activation_symmetric: null, save_calibration: false, distribution: uniform, seed: null, calibration_load_path: null, calibration_save_path: null, op_types_to_quantize: null, nodes_to_exclude: null, fp16_keep_io_types: true, fp16_op_block_list: null}.
• There are no code paths, symbols, or class-wide behavior changes. Export, loader, optimization, and compile fields are identical to the replacement-charter auto-config; only checkpoint-specific /eval declarations and the explicit fp16 /quant precision delta remain. Reducibility is consistent with the charter, recipe-free acceptance is PASS on current main, and examples/recipes/README.md remains untouched.

Analyze summary — component level and op level

Static rule analysis is ANALYZE-PARTIAL-SUCCESS for both artifacts: winml analyze intentionally exits 1 when any requested EP is not fully supported, but each output is complete for all 11 requested EP/device pairs. This is static compatibility analysis, not runtime execution.

Component-level summary

Op-level summary

Rule-less CUDA GPU, MIGraphX GPU, DML GPU, CPU, and VitisAI NPU rows classify every operator type as unknown. Each artifact also reports 12 informational SUBGRAPH/GeluPattern instances.

Reproduce commands

$OUT = Join-Path $PWD 'yolos-fashionpedia-validation'
New-Item -ItemType Directory -Force $OUT | Out-Null
$MODEL = python -c "from huggingface_hub import snapshot_download; print(snapshot_download('valentinafevu/yolos-fashionpedia', revision='0cfdb77c01cfbf0838b940ecff9c572f84c9d4f4'))"
$env:WINMLCLI_RULES_DIR = '<populated-rules-directory>'
winml build -m $MODEL -c examples/recipes/valentinafevu_yolos-fashionpedia/cpu/cpu/object-detection_fp32_config.json -o $OUT/fp32 --ep cpu --device cpu --precision fp32 --no-analyze --no-compile --rebuild --no-color
winml build -m $MODEL -c examples/recipes/valentinafevu_yolos-fashionpedia/cpu/cpu/object-detection_fp16_config.json -o $OUT/fp16 --ep cpu --device cpu --precision fp16 --no-analyze --no-compile --rebuild --no-color
winml perf -m $OUT/fp32/model.onnx --ep cpu --device cpu --precision fp32 --iterations 10 --warmup 2 --memory --output $OUT/perf-fp32.json --overwrite --format json --no-color
winml perf -m $OUT/fp16/model.onnx --ep cpu --device cpu --precision fp16 --iterations 10 --warmup 2 --memory --output $OUT/perf-fp16.json --overwrite --format json --no-color
winml eval --model $OUT/fp32/model.onnx --model-id valentinafevu/yolos-fashionpedia --dataset detection-datasets/fashionpedia --dataset-revision 80845435ce686b8a9dbf70a05452fbfb8e09cdd7 --task object-detection --device cpu --ep cpu --samples 10 --split val --no-shuffle --streaming --column input_column=image --column annotation_column=objects --column bbox_key=bbox --column category_key=category --column box_format=xywh --column box_coords=absolute --output $OUT/eval-fp32.json --overwrite --format json --no-color
winml eval --model $OUT/fp16/model.onnx --model-id valentinafevu/yolos-fashionpedia --dataset detection-datasets/fashionpedia --dataset-revision 80845435ce686b8a9dbf70a05452fbfb8e09cdd7 --task object-detection --device cpu --ep cpu --samples 10 --split val --no-shuffle --streaming --column input_column=image --column annotation_column=objects --column bbox_key=bbox --column category_key=category --column box_format=xywh --column box_coords=absolute --output $OUT/eval-fp16.json --overwrite --format json --no-color
winml analyze --model $OUT/fp32/model.onnx --ep all --device all --htp-metadata $OUT/fp32/export_htp_metadata.json --output $OUT/analyze-fp32.json --overwrite --format json --no-color
winml analyze --model $OUT/fp16/model.onnx --ep all --device all --htp-metadata $OUT/fp16/export_htp_metadata.json --output $OUT/analyze-fp16.json --overwrite --format json --no-color
Remove-Item Env:WINMLCLI_RULES_DIR
uv run pre-commit run insert-license --all-files
uv run ruff check src/ tests/
uv run mypy -p winml.modelkit
uv run pytest tests/unit/analyze --tb=short --no-cov -m 'not e2e and not npu and not gpu'
uv run pytest tests/unit/models tests/unit/loader tests/unit/datasets tests/unit/export --tb=short --no-cov -m 'not e2e and not npu and not gpu'
uv run pytest tests/unit/optim --tb=short --no-cov -m 'not e2e and not npu and not gpu'
uv run pytest tests/unit/commands tests/unit/config tests/unit/build tests/unit/compiler tests/unit/session tests/unit/eval --tb=short --no-cov -m 'not e2e and not npu and not gpu'
uv run pytest tests/unit/core tests/unit/onnx tests/unit/cache tests/unit/utils tests/unit/test_helpers tests/unit/sysinfo tests/unit/inspect tests/unit/optracing tests/regression tests/cli --tb=short --no-cov -m 'not e2e and not npu and not gpu'