recipe(cross-encoder/ms-marco-MiniLM-L6-v2): add text-classification fp32/fp16 recipes

[codykk]

Recipe-only contribution (Effort L0). Adds fp32 and fp16 recipe configs for
cross-encoder/ms-marco-MiniLM-L6-v2 (BertForSequenceClassification, task text-classification)
on two verified EP/device combinations: CPU and DML (GPU). Goal L1 (perf) PASS on all 4 configs.

---

1. Recipe path(s)

• examples/recipes/cross-encoder_ms-marco-MiniLM-L6-v2/cpu/cpu/text-classification_fp32_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L6-v2/cpu/cpu/text-classification_fp16_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L6-v2/dml/gpu/text-classification_fp32_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L6-v2/dml/gpu/text-classification_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/minilm_cpu_fp32/ (cpu/cpu fp32)
• temp/minilm_cpu_fp16/ (cpu/cpu fp16)
• temp/minilm_dml_fp32/ (dml/gpu fp32)
• temp/minilm_dml_fp16/ (dml/gpu fp16)

4. Build log

All 4 configs completed successfully:

• cpu/cpu fp32: ✅ Build complete in 20.1s (Export 4.4s, Optimize 15.2s)
• cpu/cpu fp16: ✅ Build complete in 21.3s (Export 4.3s, Optimize 15.6s, FP16 0.8s)
• dml/gpu fp32: ✅ Build complete in 19.9s (Export 4.3s, Optimize 15.0s)
• dml/gpu fp16: ✅ Build complete in 20.6s (Export 4.3s, Optimize 15.0s, FP16 0.8s)

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

bert architecture is fully supported by Optimum's BertOnnxConfig and winml's BertIOConfig. No custom OnnxConfig needed. winml inspect confirms all components at "Default" status.

7. Claimed (Effort, Goal, Outcome)

Axis	Tier
Effort	L0 (recipe-only, no per-architecture code)
Goal	L1 (build + perf)
Outcome	L0 (recipe + report)

8. Goal-ladder verdict table

Tier	Verdict	Evidence
L0 (build)	PASS	All 4 configs (cpu/cpu fp32, cpu/cpu fp16, dml/gpu fp32, dml/gpu fp16) build successfully with winml build
L1 (perf)	PASS	All 4 configs produce valid latency/throughput numbers via winml perf (see item 10)
L2 (numeric vs PyTorch)	N/A	Not attempted (Goal ceiling = L1)
L3 (task metric)	N/A	Not attempted (Goal ceiling = L1)

9. Methodology-evolution declaration

No methodology friction observed during this contribution.

10. Perf & eval data

EP / Device	Precision	Verdict	Mean	p50	p90	Throughput	RAM Δ	VRAM Δ (local)
OpenVINOExecutionProvider / cpu	fp32	PASS	21.110 ms	21.980 ms	24.710 ms	47.38 samples/s	+195.70 MB	—
OpenVINOExecutionProvider / cpu	fp16	PASS	21.440 ms	21.850 ms	25.760 ms	46.64 samples/s	+196.00 MB	—
DmlExecutionProvider / gpu	fp32	PASS	16.350 ms	15.920 ms	19.070 ms	61.16 samples/s	+314.00 MB	+184.80 MB
DmlExecutionProvider / gpu	fp16	PASS	47.800 ms	46.470 ms	51.710 ms	20.92 samples/s	+262.80 MB	+107.80 MB
QNNExecutionProvider / npu	*	HOST-BLOCKED	—	—	—	—	—	No NPU hardware available on test host

Model size: fp32 = 86.7 MB; fp16 = 43.4 MB (50% size reduction).

Note: DML fp16 is slower than fp32 on this model (47.8ms vs 16.4ms). This is a known characteristic of small models where fp16 conversion overhead on GPU outweighs the compute savings.

11. Component / op-level data

From winml analyze (post-optimization):

• Total operators: 202 (after fusion)
• Fusion patterns applied: gelu_fusion, matmul_add_fusion (autoconf converged in 2 iterations)
• Model architecture: 6 hidden layers, 12 attention heads, hidden size 384
• Artifact: temp/minilm_cpu_fp32/analyze_result.json

12. Reproducible commands

# Build (cpu/cpu fp32)
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L6-v2/cpu/cpu/text-classification_fp32_config.json -m cross-encoder/ms-marco-MiniLM-L6-v2 -o temp/minilm_cpu_fp32

# Build (cpu/cpu fp16) — --precision fp16 required to override auto-precision when --device is passed
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L6-v2/cpu/cpu/text-classification_fp16_config.json -m cross-encoder/ms-marco-MiniLM-L6-v2 -o temp/minilm_cpu_fp16 --precision fp16

# Build (dml/gpu fp32)
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L6-v2/dml/gpu/text-classification_fp32_config.json -m cross-encoder/ms-marco-MiniLM-L6-v2 -o temp/minilm_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/cross-encoder_ms-marco-MiniLM-L6-v2/dml/gpu/text-classification_fp16_config.json -m cross-encoder/ms-marco-MiniLM-L6-v2 -o temp/minilm_dml_fp16 --ep dml --device gpu --precision fp16

# Perf (cpu fp32)
winml perf -m temp/minilm_cpu_fp32/model.onnx --device cpu --iterations 20

# Perf (cpu fp16)
winml perf -m temp/minilm_cpu_fp16/model.onnx --device cpu --iterations 20

# Perf (dml/gpu fp32)
winml perf -m temp/minilm_dml_fp32/model.onnx --device gpu --ep dml --iterations 20

# Perf (dml/gpu fp16)
winml perf -m temp/minilm_dml_fp16/model.onnx --device gpu --ep dml --iterations 20