recipe(cross-encoder/ms-marco-MiniLM-L4-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-L4-v2 (BertForSequenceClassification, 4 layers, 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-L4-v2/cpu/cpu/text-classification_fp32_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/cpu/cpu/text-classification_fp16_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/dml/gpu/text-classification_fp32_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/dml/gpu/text-classification_fp16_config.json

2. README row

N/A — README not modified in this PR.

3. Build output dir

• temp/minilm4_cpu_fp32/
• temp/minilm4_cpu_fp16/
• temp/minilm4_dml_fp32/
• temp/minilm4_dml_fp16/

4. Build log

• cpu/cpu fp32: ✅ Build complete in 16.2s (Export 3.8s, Optimize 12.0s)
• cpu/cpu fp16: ✅ Build complete in 16.9s (Export 3.5s, Optimize 12.2s, FP16 0.7s)
• dml/gpu fp32: ✅ Build complete in 16.5s (Export 3.7s, Optimize 12.3s)
• dml/gpu fp16: ✅ Build complete in 17.3s (Export 3.8s, Optimize 12.5s, FP16 0.7s)

5. Appended findings

N/A — recipe-only L0 contribution.

6. Optimum-coverage probe

bert architecture fully supported by Optimum and winml. winml inspect confirms all components at "Default" status.

7. Claimed (Effort, Goal, Outcome)

Axis	Tier
Effort	L0 (recipe-only)
Goal	L1 (build + perf)
Outcome	L0 (recipe + report)

8. Goal-ladder verdict table

Tier	Verdict	Evidence
L0 (build)	PASS	All 4 configs build successfully
L1 (perf)	PASS	All 4 configs produce valid latency/throughput (see item 10)
L2	N/A	Goal ceiling = L1
L3	N/A	Goal ceiling = L1

9. Methodology-evolution declaration

No methodology friction observed.

10. Perf & eval data

EP / Device	Precision	Verdict	Throughput	RAM Δ	VRAM Δ (local)
OpenVINOExecutionProvider / cpu	fp32	PASS	82.69 samples/s	+159.30 MB	—
OpenVINOExecutionProvider / cpu	fp16	PASS	85.10 samples/s	+160.10 MB	—
DmlExecutionProvider / gpu	fp32	PASS	81.25 samples/s	+283.00 MB	+171.30 MB
DmlExecutionProvider / gpu	fp16	PASS	31.42 samples/s	+243.20 MB	+101.00 MB
QNNExecutionProvider / npu	*	HOST-BLOCKED	—	—	No NPU hardware

11. Component / op-level data

• Architecture: 4 hidden layers, 12 attention heads, hidden size 384
• Fusion patterns applied: gelu_fusion, matmul_add_fusion

12. Reproducible commands

# Build (cpu/cpu fp32)
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/cpu/cpu/text-classification_fp32_config.json -m cross-encoder/ms-marco-MiniLM-L4-v2 -o temp/minilm4_cpu_fp32

# Build (cpu/cpu fp16)
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/cpu/cpu/text-classification_fp16_config.json -m cross-encoder/ms-marco-MiniLM-L4-v2 -o temp/minilm4_cpu_fp16 --precision fp16

# Build (dml/gpu fp32)
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/dml/gpu/text-classification_fp32_config.json -m cross-encoder/ms-marco-MiniLM-L4-v2 -o temp/minilm4_dml_fp32 --ep dml --device gpu

# Build (dml/gpu fp16)
winml build -c examples/recipes/cross-encoder_ms-marco-MiniLM-L4-v2/dml/gpu/text-classification_fp16_config.json -m cross-encoder/ms-marco-MiniLM-L4-v2 -o temp/minilm4_dml_fp16 --ep dml --device gpu --precision fp16

# Perf
winml perf -m temp/minilm4_cpu_fp32/model.onnx --device cpu --iterations 20
winml perf -m temp/minilm4_cpu_fp16/model.onnx --device cpu --iterations 20
winml perf -m temp/minilm4_dml_fp32/model.onnx --device gpu --ep dml --iterations 20
winml perf -m temp/minilm4_dml_fp16/model.onnx --device gpu --ep dml --iterations 20