recipe(owlvit): add base patch16 fp16 recipe

[kujin66]

Adds a CPU-validated FP16 recipe for google/owlvit-base-patch16 zero-shot object detection. The model already exports on current origin/main, but the auto-config/baseline artifact assigns ONNX input names to the wrong shapes; this recipe fixes the input order to match OwlViTForObjectDetection.forward(input_ids, pixel_values, attention_mask) and adds a verified FP16 artifact path.

1. Recipe path(s)

   • examples/recipes/google_owlvit-base-patch16/zero-shot-object-detection_fp16_config.json

2. README row

   • Not added. examples/recipes/README.md currently says the table lists model/task tuples that pass fp16 eval on all 10 EP/device buckets; this PR has CPU build/perf evidence only, so adding a row would over-claim catalog coverage.

3. Build output dir

   • Baseline: temp/baseline_google_owlvit_base_patch16
   • Recipe: temp/verify_google_owlvit_base_patch16_fp16_fixed
   • Perf: temp/owlvit_fp16_perf_cpu.json

4. Build log

   • Baseline command: uv run winml build -m google/owlvit-base-patch16 -o temp/baseline_google_owlvit_base_patch16 --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
   • Baseline result: Build complete in 126.9s; final artifact temp/baseline_google_owlvit_base_patch16/model.onnx; baseline HEAD 38767add6f91c7b10b6394fae3af6f437e02effd; winml-cli package version 0.2.0.
   • Baseline structural issue: pixel_values [1,16], input_ids [1,3,768,768], attention_mask [1,16].
   • Recipe command: uv run winml build -c examples/recipes/google_owlvit-base-patch16/zero-shot-object-detection_fp16_config.json -m google/owlvit-base-patch16 -o temp/verify_google_owlvit_base_patch16_fp16_fixed --ep cpu --device cpu --precision fp16 --no-analyze --no-compile --rebuild
   • Recipe result: Build complete in 90.4s; final artifact temp/verify_google_owlvit_base_patch16_fp16_fixed/model.onnx; FP16 stage 14.5s; quantized artifact 291.1 MB.

5. Appended findings

   • Skill knowledge appended outside this model PR: model_knowledge/owlvit.json, finding owlvit-001.
   • Key finding: OwlViT recipes must order input_tensors as input_ids, pixel_values, attention_mask; the inspect/autoconfig order can build but produces semantic ONNX input-name/shape swaps.
   • No skill_meta finding added; existing reviewer guidance already covers semantically wrong recipe I/O and FP16 initializer validation.

6. Optimum-coverage probe

   • model_type: owlvit
   • architectures: ['OwlViTForObjectDetection']
   • vendor: ['feature-extraction', 'zero-shot-object-detection']
   • after_winml: ['feature-extraction', 'zero-shot-object-detection']
   • added_by_winml: []
   • Verdict: VENDOR-ONLY
   • winml inspect resolves task zero-shot-object-detection, loader AutoModelForZeroShotObjectDetection, exporter OwlViTOnnxConfig, WinML class WinMLModelForGenericTask.

7. Claimed (Effort, Goal, Outcome)

   • Effort: L0 recipe-only, with an existing OWLv2 zero-shot-object-detection template.
   • Goal ceiling: L1 CPU.
   • Outcome: L0 model PR plus skill-side model_knowledge finding.
   • Target EPs: ['cpu'].

8. Goal-ladder verdict table

   • L0 PASS: recipe JSON parses; build completed in 90.4s; ONNX structural validation: inputs input_ids [1,16], pixel_values [1,3,768,768], attention_mask [1,16]; outputs logits [1,2304,1], pred_boxes [1,2304,4], text_embeds [1,1,512], image_embeds [1,48,48,768]; FP16 validation found 419 FLOAT16 initializers and 0 FLOAT initializers; model.onnx.data is 304,811,008 bytes vs baseline 609,623,040 bytes.
   • L1 PASS: uv run winml perf -m temp/verify_google_owlvit_base_patch16_fp16_fixed/model.onnx --device cpu --ep cpu --format json -o temp/owlvit_fp16_perf_cpu.json --overwrite wrote JSON with CPUExecutionProvider, 100 iterations, mean 2566.361 ms, p50 2550.853 ms, throughput 0.39 samples/s, RSS total delta 1978.02 MB. Available providers snapshot: ['DmlExecutionProvider', 'CPUExecutionProvider'].

9. Methodology-evolution declaration

   • No methodology friction requiring a new _meta rule was observed. The run did surface a model-family gotcha, captured in model_knowledge/owlvit.json (owlvit-001), and the existing reviewer checklist already requires semantic recipe sanity plus emitted FP16 validation.

[kujin66]

APPROVE

Reviewer checks performed from PR branch kujin66/add-owlvit-base-patch16-recipe at e0726398:

• PR exists and is draft: recipe(owlvit): add base patch16 fp16 recipe #1155, base main, head kujin66/add-owlvit-base-patch16-recipe.
• Diff scope matches L0 recipe-only claim: git diff --name-only origin/main...HEAD shows only examples/recipes/google_owlvit-base-patch16/zero-shot-object-detection_fp16_config.json.
• PR body contains the required 9 hand-off items: recipe path, README row decision, build dirs/logs, findings, Optimum probe, claimed tiers, Goal ladder, methodology declaration.
• Baseline gate cited current origin/main 38767add6f91c7b10b6394fae3af6f437e02effd; worktree HEAD before commit matched that commit.
• Recipe schema parses with uv run python -m json.tool examples/recipes/google_owlvit-base-patch16/zero-shot-object-detection_fp16_config.json.
• Structural validation of rebuilt artifact temp/verify_google_owlvit_base_patch16_fp16_fixed/model.onnx: inputs are input_ids [1,16], pixel_values [1,3,768,768], attention_mask [1,16]; outputs are logits [1,2304,1], pred_boxes [1,2304,4], text_embeds [1,1,512], image_embeds [1,48,48,768].
• FP16 validation: 419 FLOAT16 initializers, 0 FLOAT initializers.
• Goal L1 evidence: temp/owlvit_fp16_perf_cpu.json contains CPUExecutionProvider perf, mean 2566.361 ms, p50 2550.853 ms, throughput 0.39 samples/s, RSS total delta 1978.02 MB.
• Knowledge capture exists outside this model PR: model_knowledge/owlvit.json / owlvit-001, JSON-validated locally.

Coverage annotation: full for charter target EP cpu; non-target EPs are not claimed.