Summary

Recipe-only contribution (Effort L0). Adds fp32 and fp16 recipe configs for
google/owlv2-base-patch16-ensemble (Owlv2ForObjectDetection, task zero-shot-object-detection)
on two verified EP/device combinations: CPU and DML (GPU). Goal L1 (perf) PASS on all 4 configs.

1. Recipe path(s)

• examples/recipes/google_owlv2-base-patch16-ensemble/cpu/cpu/zero-shot-object-detection_fp32_config.json
• examples/recipes/google_owlv2-base-patch16-ensemble/cpu/cpu/zero-shot-object-detection_fp16_config.json
• examples/recipes/google_owlv2-base-patch16-ensemble/dml/gpu/zero-shot-object-detection_fp32_config.json
• examples/recipes/google_owlv2-base-patch16-ensemble/dml/gpu/zero-shot-object-detection_fp16_config.json

2. README row

N/A — README not modified in this PR. The model has not yet passed fp16 eval on all 10 (EP, device) buckets; only cpu/cpu and dml/gpu are verified.

3. Build output dir

• temp/owlv2_cpu_fp32/ (cpu/cpu fp32)
• temp/owlv2_cpu_fp16/ (cpu/cpu fp16)
• temp/owlv2_dml_fp32/ (dml/gpu fp32)
• temp/owlv2_dml_fp16/ (dml/gpu fp16)

4. Build log

All 4 configs completed successfully:

• cpu/cpu fp32: ✅ Build complete in 46.4s (Export 24.0s, Optimize 21.7s)
• cpu/cpu fp16: ✅ Build complete in 51.6s (Export 23.3s, Optimize 22.6s, FP16 5.1s)
• dml/gpu fp32: ✅ Build complete in 47.2s (Export 23.9s, Optimize 22.6s)
• dml/gpu fp16: ✅ Build complete in 46.2s (Export 23.2s, Optimize 22.3s)

Note: Einsum op in class_head triggers OpUnsupportedError during optimize analysis but does not block the build.

5. Appended findings

N/A — no model_knowledge/ or skill_meta/ entries added (recipe-only L0 contribution; skill repo is separate from this working repo).

6. Optimum-coverage probe

N/A — owlv2 architecture is already supported by HuggingFace transformers. No custom OnnxConfig was needed.

7. Claimed (Effort, Goal, Outcome)

8. Goal-ladder verdict table

9. Methodology-evolution declaration

No methodology friction observed during this contribution.

10. Perf & eval data

Model size: fp32 = 585.55 MB (export); fp16 = 293.0 MB (quantized, 50% size reduction).

11. Component / op-level data

From winml analyze (post-optimization):

• Total operators: 1025 (after fusion)
• Unique operator types: 28
• Top operators: Reshape (411), Gemm (150), Transpose (121), Add (79), Mul (77), LayerNormalization (52), MatMul (49), Sigmoid (25), Softmax (24)
• Notable ops: Einsum (1, triggers OpUnsupportedError during analysis but does not block build), ArgMax (1), Gelu (2)
• Fusion patterns applied: gelu_fusion, matmul_add_fusion (autoconf converged in 2 iterations, 1007 → 1025 nodes)
• Artifact: temp/owlv2_cpu_fp32/analyze_result.json

12. Reproducible commands

# Build (cpu/cpu fp32)
winml build -c examples/recipes/google_owlv2-base-patch16-ensemble/cpu/cpu/zero-shot-object-detection_fp32_config.json -m google/owlv2-base-patch16-ensemble -o temp/owlv2_cpu_fp32

# Build (cpu/cpu fp16) — --precision fp16 required to override auto-precision when --device is passed
winml build -c examples/recipes/google_owlv2-base-patch16-ensemble/cpu/cpu/zero-shot-object-detection_fp16_config.json -m google/owlv2-base-patch16-ensemble -o temp/owlv2_cpu_fp16 --precision fp16

# Build (dml/gpu fp32)
winml build -c examples/recipes/google_owlv2-base-patch16-ensemble/dml/gpu/zero-shot-object-detection_fp32_config.json -m google/owlv2-base-patch16-ensemble -o temp/owlv2_dml_fp32 --ep dml --device gpu

# Build (dml/gpu fp16) — --precision fp16 required to override auto-precision when --device is passed
winml build -c examples/recipes/google_owlv2-base-patch16-ensemble/dml/gpu/zero-shot-object-detection_fp16_config.json -m google/owlv2-base-patch16-ensemble -o temp/owlv2_dml_fp16 --ep dml --device gpu --precision fp16

# Perf (cpu fp32)
winml perf -m temp/owlv2_cpu_fp32/model.onnx --device cpu --iterations 5

# Perf (cpu fp16)
winml perf -m temp/owlv2_cpu_fp16/model.onnx --device cpu --iterations 5

# Perf (dml/gpu fp32)
winml perf -m temp/owlv2_dml_fp32/model.onnx --device gpu --ep dml --iterations 5

# Perf (dml/gpu fp16)
winml perf -m temp/owlv2_dml_fp16/model.onnx --device gpu --ep dml --iterations 5