recipe(nli-MiniLM2-L6-H768): add CPU text classification configs

[fuliucansheng]

Summary

Adds CPU recipe coverage for cross-encoder/nli-MiniLM2-L6-H768, an NLI cross-encoder exported through WinML's existing RoBERTa sequence-classification path. This is an L0 recipe-only contribution: current main already builds and runs the model, and this PR records verified CPU fp32/fp16 coverage with checked-in recipes.

Model metadata

What the model does: cross-encoder/nli-MiniLM2-L6-H768 classifies sentence pairs into NLI labels (contradiction, entailment, neutral). WinML resolves it as text-classification with RobertaForSequenceClassification.

Primary user stories:

Story	Confidence
A user supplies a premise/hypothesis pair to obtain NLI logits for entailment-style classification.	verified from checkpoint config labels and architecture

Supported tasks:

Task	Surface	Confidence
text-classification	Transformers / Optimum ONNX / WinML export surface	verified
NLI / sentence-pair classification	checkpoint label semantics	verified

Model architecture:

RobertaForSequenceClassification
├── RobertaModel
│   ├── Embeddings (vocab 50265, hidden 768)
│   └── Encoder stack x6
│       ├── Self-attention (12 heads)
│       └── Feed-forward + GELU + residual/layernorm
└── RobertaClassificationHead -> logits [B, 3]

Source/confidence: pinned checkpoint config resolves model_type: roberta, architectures: [RobertaForSequenceClassification], hidden size 768, 6 layers, 12 attention heads, and 3 NLI labels.

Validation and support evidence

Baseline

Baseline was run from current origin/main at 2c378c7dc213e0a69afec87d3d7c5015eb4e570d with winml, version 0.2.0.

Probe	Result
winml inspect	PASS: loader AutoModelForSequenceClassification, exporter RobertaIOConfig, WinML class WinMLModelForSequenceClassification; inputs input_ids and attention_mask int32 [1, 512]; output logits.
Optimum coverage	VENDOR-ONLY for roberta: vendor tasks include text-classification; WinML adds no extra task override.
Recipe-free build	PASS: build complete in 27.9s; final artifact emitted; 368 exported ONNX nodes.
Recipe-free perf	PASS on CPU fp32: avg 167.59 ms, p50 189.91 ms, throughput 5.97 samples/sec, RAM total +282.3 MB.
L2 compare	PASS on CPU fp32: cosine mean 0.9999999999995716, max absolute diff max 2.2649765014648438e-06.
L3 task metric	BLOCKED: default text-classification eval uses GLUE/MRPC labels and fails label alignment for label equivalent; this checkpoint's labels are contradiction, entailment, and neutral.

Baseline winml config -p fp32 --ep cpu --device cpu and winml config -p fp16 --ep cpu --device cpu are the sources for the checked-in recipes.

Goal

Effort: L0. Goal ceiling: L2. Outcome: L0 recipe-only.

Success means the checked-in CPU fp32/fp16 recipes build structurally valid ONNX artifacts, run winml perf on CPU, and match the PyTorch reference numerically through WinML compare mode. L3 is explicitly blocked by default dataset label semantics for this NLI checkpoint.

Outcome

Highest verified goal: L2 PASS on CPU for both fp32 and fp16. Coverage is full for the targeted EP set (cpu/cpu/fp32, cpu/cpu/fp16); no accelerator tuples are claimed.

Shipped recipes:

Recipe
examples/recipes/cross-encoder_nli-MiniLM2-L6-H768/cpu/cpu/text-classification_fp32_config.json
examples/recipes/cross-encoder_nli-MiniLM2-L6-H768/cpu/cpu/text-classification_fp16_config.json

Methodology friction observed: none. Model finding: RoBERTa NLI cross-encoders export cleanly through text-classification; default MRPC eval is not semantically compatible without an NLI-specific dataset/label mapping.

Per-EP/device/precision results

Tier	EP / Device	Precision	Verdict	Evidence
L0 build	CPUExecutionProvider / cpu	fp32	PASS	Recipe build complete in 22.0s; optimized model is loadable with inputs input_ids, attention_mask int32 [1, 512], output logits [1, 3], 210 nodes, external data colocated with model.onnx.
L0 build	CPUExecutionProvider / cpu	fp16	PASS	Recipe build complete in 25.3s with --precision fp16; final artifact is about 165.1 MB and reports fp16 at perf time, with 211 nodes.
L1 perf	CPUExecutionProvider / cpu	fp32	PASS	Avg 538.38 ms, p50 528.47 ms, throughput 1.86 samples/sec, RAM total +277.7 MB.
L1 perf	CPUExecutionProvider / cpu	fp16	PASS	Avg 587.20 ms, p50 669.09 ms, throughput 1.70 samples/sec, RAM total +296.9 MB.
L2 compare	CPUExecutionProvider / cpu	fp32	PASS	winml eval --mode compare: cosine mean 0.9999999999995716, max absolute diff max 2.2649765014648438e-06, SQNR mean 117.705 dB.
L2 compare	CPUExecutionProvider / cpu	fp16	PASS	winml eval --mode compare: cosine mean 0.9999997277161565, max absolute diff max 0.0017644912004470825, SQNR mean 63.299 dB.
L3 eval	CPUExecutionProvider / cpu	fp32	CLI-BLOCKED	Default text-classification eval dataset nyu-mll/glue/mrpc fails label alignment for label equivalent; checkpoint labels are NLI labels.
L3 eval	CPUExecutionProvider / cpu	fp16	CLI-BLOCKED	Same label-alignment blocker as fp32; no task-metric value is reported.

Delta

Both recipes are generated from current winml config output with explicit CPU targeting. The fp32 recipe preserves quant: null; the fp16 recipe preserves the generated fp16 quantization block and model metadata. No source code or tests are changed, and examples/recipes/README.md is untouched.

Reducibility: recipe-only is appropriate because auto-config already resolves task, loader, exporter, and WinML inference class correctly; this PR is a per-checkpoint verified CPU coverage declaration rather than a class-wide code fix.

Analyze summary

Static rule analysis completed with partial-success exit semantics: QNN/OpenVINO rows were complete, while VitisAI has no rule data and is all-unknown. This is compatibility analysis, not runtime execution.

Component-level summary

Artifact	Architecture coverage	Mapping	Actionable EP findings
fp32	embeddings; 6x encoder attention/FFN; sequence-classification head	368 exported nodes tagged during HTP export; 210 optimized ONNX ops analyzed	QNN/OpenVINO NPU: all analyzed op types supported; VitisAI: no rule data.
fp16	embeddings; 6x encoder attention/FFN; sequence-classification head	368 exported nodes tagged during HTP export; 211 optimized ONNX ops analyzed	QNN/OpenVINO NPU: all analyzed op types supported; VitisAI: no rule data.

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	210 ops / 19 types	Reshape 60; Gemm 38; Transpose 24; Add 21; LayerNormalization 13; Mul 13; MatMul 12	QNN/OpenVINO: 210 supported, 0 partial, 0 unsupported, 0 unknown; VitisAI: 210 unknown.
fp16	211 ops / 19 types	Reshape 60; Gemm 38; Transpose 24; Add 21; LayerNormalization 13; Mul 13; MatMul 12	QNN/OpenVINO: 211 supported, 0 partial, 0 unsupported, 0 unknown; VitisAI: 211 unknown.

Reproduce commands

$MODEL = "cross-encoder/nli-MiniLM2-L6-H768"
$OUT = "temp/repro_cross_encoder_nli_MiniLM2_L6_H768"

uv run winml inspect -m $MODEL --format json
uv run winml build -c examples/recipes/cross-encoder_nli-MiniLM2-L6-H768/cpu/cpu/text-classification_fp32_config.json -m $MODEL -o "$OUT/fp32" --ep cpu --device cpu --no-compile --rebuild
uv run winml build -c examples/recipes/cross-encoder_nli-MiniLM2-L6-H768/cpu/cpu/text-classification_fp16_config.json -m $MODEL -o "$OUT/fp16" --ep cpu --device cpu --precision fp16 --no-compile --rebuild
uv run winml perf -m "$OUT/fp32/model.onnx" --ep cpu --device cpu
uv run winml perf -m "$OUT/fp16/model.onnx" --ep cpu --device cpu
uv run winml eval -m "$OUT/fp32/model.onnx" --model-id $MODEL --task text-classification --mode compare --ep cpu --device cpu --samples 20
uv run winml eval -m "$OUT/fp16/model.onnx" --model-id $MODEL --task text-classification --mode compare --ep cpu --device cpu --samples 20
uv run winml analyze --model "$OUT/fp32/model.onnx" --ep all --output "$OUT/fp32/analyze_all.json"
uv run winml analyze --model "$OUT/fp16/model.onnx" --ep all --output "$OUT/fp16/analyze_all.json"