recipe(google/owlv2-base-patch16-ensemble): add zero-shot-object-detection fp16, w8a8 and w8a16 recipes

[codykk]

Add fp16, w8a8, and w8a16 quantized recipes for google/owlv2-base-patch16-ensemble (zero-shot object detection, 155M params). Effort L0★ — first zero-shot-object-detection recipe in the repo, no source edits. Goal ceiling L1 reached: all three recipes build and the w8a16 variant perfs on CPU. Quantization delivers up to 3.9x model size reduction (585.6 MB fp32 → 148.4 MB w8a8).

---

12-Item Verification

1. Recipe path(s)

• examples/recipes/google_owlv2-base-patch16-ensemble/zero-shot-object-detection_fp16_config.json
• examples/recipes/google_owlv2-base-patch16-ensemble/zero-shot-object-detection_w8a16_config.json
• examples/recipes/google_owlv2-base-patch16-ensemble/zero-shot-object-detection_w8a8_config.json

2. README row

Added to examples/recipes/README.md:

| google/owlv2-base-patch16-ensemble | zero-shot-object-detection |

3. Build output dir

temp/owlv2_fp16/   (fp16)
temp/owlv2_build/  (w8a16)
temp/owlv2_w8a8/   (w8a8)

4. Build log

Recipe	Status	Time	Artifact Size
fp16	✅ Build complete	59.3s	585.6 MB
w8a16	✅ Build complete	119.7s	175.2 MB
w8a8	✅ Build complete	125.4s	148.4 MB

All artifacts: 1423 ONNX nodes, opset 17, IR version 8.

5. Appended findings

N/A — no model_knowledge/owlv2.json existed prior; no methodology friction observed that warrants a new finding. The model exported cleanly via the existing vendor OwlV2OnnxConfig registered by winml.

6. Optimum-coverage probe

{'vendor': [], 'after_winml': ['feature-extraction', 'zero-shot-object-detection'], 'added_by_winml': ['feature-extraction', 'zero-shot-object-detection']}

• Verdict: WINML-ONLY — no vendor Optimum coverage; winml registers OwlV2OnnxConfig for both tasks.
• Cross-check: model_type = "owlv2", architectures = ["Owlv2ForObjectDetection"], is_encoder_decoder = False.

7. Claimed (Effort, Goal, Outcome)

Axis	Tier	Justification
Effort	L0★	First zero-shot-object-detection recipe; no source edits; OnnxConfig already registered by winml
Goal	L1 (ceiling)	Build + perf verified; baseline already passes on main so L0 is inherited
Outcome	L0	Recipe JSON × 3 + README row

8. Goal-ladder verdict table

Tier	Verdict	Evidence
L0 (build)	PASS	All three recipes (fp16/w8a16/w8a8) build successfully; ✅ Build complete in 59.3s / 119.7s / 125.4s
L1 (perf)	PASS	CPU w8a16: mean 3865.88ms, P50 3859.67ms, 0.26 samples/s, RAM +2836.8 MB
L2 (numeric)	N/A	Not in ceiling
L3 (eval)	N/A	Not in ceiling

Baseline build on main (defa3f63): PASS — L0 is inherited from baseline; the contribution's delta is quantization.

9. Methodology-evolution declaration

No methodology friction observed. No new _meta-NNN finding needed.

10. Perf & eval data

EP / Device	Precision	Verdict	Mean	P50	Throughput	RAM Δ	VRAM Δ
CPUExecutionProvider / cpu	w8a16	PASS	3865.88 ms	3859.67 ms	0.26 samples/s	+2836.8 MB	—

Available providers on test host: ['DmlExecutionProvider', 'CPUExecutionProvider']

DmlExecutionProvider was not in charter.target_eps (CPU-only target), so no HOST-BLOCKED row.

11. Component / op-level data

N/A — winml analyze runtime check rules not populated on this host (source-tree build, no parquet rules downloaded). The build-embedded analyze_result.json only covers CPU (no rules → all ops unknown). See _meta-013.

Note from build: Einsum op in /class_head/Einsum flagged as unsupported during optimize (3 occurrences), but optimization and quantization completed successfully around it.

12. Reproducible commands

# Environment
winml --version
git rev-parse HEAD  # defa3f63ffdc5c4cd3c252a4b3f60fcc9939f953 (origin/main)
python -c "import onnxruntime as ort; print(ort.get_available_providers())"
# ['DmlExecutionProvider', 'CPUExecutionProvider']

# Baseline (no recipe, main branch)
winml build -m google/owlv2-base-patch16-ensemble -o temp/baseline_owlv2/ --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild

# Build fp16
winml build -c examples/recipes/google_owlv2-base-patch16-ensemble/zero-shot-object-detection_fp16_config.json -m google/owlv2-base-patch16-ensemble -o temp/owlv2_fp16/ --rebuild

# Build w8a16
winml build -c examples/recipes/google_owlv2-base-patch16-ensemble/zero-shot-object-detection_w8a16_config.json -m google/owlv2-base-patch16-ensemble -o temp/owlv2_build/ --rebuild

# Build w8a8
winml build -c examples/recipes/google_owlv2-base-patch16-ensemble/zero-shot-object-detection_w8a8_config.json -m google/owlv2-base-patch16-ensemble -o temp/owlv2_w8a8/ --rebuild

# Perf (w8a16 on CPU)
winml perf -m temp/owlv2_build/model.onnx --device cpu --ep cpu