recipe(stanfordmimi/synthpose-vitpose-base-hf): add verified CPU fp32/fp16 recipes

[codykk]

Summary

Add verified CPU fp32 and fp16 keypoint-detection recipes for stanfordmimi/synthpose-vitpose-base-hf, a ViTPose base model trained on synthetic pose data for human keypoint estimation. 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 for CPUExecutionProvider. Highest Goal verdict: L1 PASS (L3 CLI-BLOCKED: winml eval metrics computation failed for keypoint-detection).

Model metadata

1. What the model does — ViTPose is a vision-transformer-based model for human pose estimation that predicts body keypoint heatmaps from cropped person images. This checkpoint is trained on synthetic pose data (SynthPose). Evidence: HF checkpoint config, VitPoseForPoseEstimation architecture. Confidence: verified.
2. Primary user stories — A user supplies a cropped person image to obtain keypoint heatmaps for human pose estimation. Evidence: checkpoint model card, task registration. Confidence: verified.
3. Supported tasks — keypoint-detection across checkpoint, Transformers, and WinML surfaces. Evidence: winml inspect on current main, Optimum probe (WINML-ONLY). Confidence: verified.
4. Model architecture — ViTPose base model for pose estimation.

VitPoseForPoseEstimation
├── ViT Backbone
│   ├── Patch embedding (Conv2d 3→768, patch 16×16, input 256×192)
│   ├── Position embeddings
│   └── Transformer layer × 12 (self-attention + MLP)
└── Keypoint head → heatmaps [B, num_keypoints, H, W]

• Source/confidence: Transformers VitPoseForPoseEstimation source, winml inspect I/O shapes (mapped).

Validation and support evidence

Baseline

• Current main at 3ea6df73d5438040d21167b502c251d3e451fb2b.
• Starting auto-config (winml config): PASS with task keypoint-detection, VitPoseForPoseEstimation, input pixel_values [1,3,256,192], output heatmaps, opset 17.
• Recipe-free build: PASS in 24.2s; final artifact 343.5 MB.
• Baseline CPU fp32 perf: PASS; mean 57.44 ms, P50 57.07 ms, throughput 17.41 samples/s, RAM model load +360 MB, inference +27 MB.
• Baseline eval: CLI-BLOCKED — winml eval ran inference but failed at metrics computation for task keypoint-detection.
• Optimum probe: keypoint-detection in added_by_winml only (WINML-ONLY).

Goal

• Committed axes: Effort L0 / Goal ceiling L1 / Outcome L0.
• L0 is inherited from baseline (build passes on main). Ceiling is L1 (perf) because L3 is CLI-BLOCKED (metrics computation failure) and L2 is CLI-pending.

Outcome

• Shipped tier: L0; highest Goal verdict: L1 PASS; coverage: full (all required CPU tuples passed); deferred tuples: none.
• Shipped recipes:
  • examples/recipes/stanfordmimi_synthpose-vitpose-base-hf/cpu/cpu/keypoint-detection_fp32_config.json
  • examples/recipes/stanfordmimi_synthpose-vitpose-base-hf/cpu/cpu/keypoint-detection_fp16_config.json
• No source code changes.

Per-EP/device/precision results

Goal ladder

Tier	CPUExecutionProvider / cpu / fp32	CPUExecutionProvider / cpu / fp16
L0 — build	PASS — 22.5s, 343.5 MB	PASS — 21.6s, 171.8 MB
L1 — perf	PASS — P50 58.39 ms	PASS — P50 70.35 ms
L3 — eval	CLI-BLOCKED	CLI-BLOCKED

Perf

EP / Device	Precision	Verdict	Mean	P50	Throughput	RAM model load Δ	RAM inference Δ
CPUExecutionProvider / cpu	fp32	PASS	56.46 ms	58.39 ms	17.71 samples/s	+359 MB	+27 MB
CPUExecutionProvider / cpu	fp16	PASS	70.38 ms	70.35 ms	14.21 samples/s	+368 MB	+33 MB

Eval

EP / Device	Precision	Verdict	Evidence
CPUExecutionProvider / cpu	fp32	CLI-BLOCKED	winml eval inference completed but metrics computation failed for keypoint-detection
CPUExecutionProvider / cpu	fp16	CLI-BLOCKED	Same as above

Delta

• fp32 recipe: identical to winml config auto-config output — filed for verified cpu/cpu coverage.
• fp16 recipe: identical to winml config -p fp16 auto-config output — filed for verified cpu/cpu coverage.

Reproduce

# fp32
winml build -m stanfordmimi/synthpose-vitpose-base-hf \
  -c examples/recipes/stanfordmimi_synthpose-vitpose-base-hf/cpu/cpu/keypoint-detection_fp32_config.json \
  --ep cpu --device cpu -o out/fp32
winml perf -m out/fp32/model.onnx --ep cpu --device cpu

# fp16
winml build -m stanfordmimi/synthpose-vitpose-base-hf \
  -c examples/recipes/stanfordmimi_synthpose-vitpose-base-hf/cpu/cpu/keypoint-detection_fp16_config.json \
  -p fp16 --ep cpu --device cpu -o out/fp16
winml perf -m out/fp16/model.onnx --ep cpu --device cpu