recipe(mmarco-mMiniLMv2-L12-H384-v1): add CPU text classification configs

[fuliucansheng]

Summary

Adds CPU recipe coverage for cross-encoder/mmarco-mMiniLMv2-L12-H384-v1, a multilingual MMARCO cross-encoder reranker exported through WinML's existing XLM-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/mmarco-mMiniLMv2-L12-H384-v1 scores query/passage text pairs for multilingual retrieval reranking. The checkpoint exposes a single-logit XLMRobertaForSequenceClassification head, so WinML exports it through text-classification even though its semantic use is text ranking.

Primary user stories:

Story	Confidence
A user supplies a query and candidate passages to obtain ranking scores for retrieve-and-rerank search.	verified from checkpoint task/model card metadata and config

Supported tasks:

Task	Surface	Confidence
text-ranking	checkpoint/model-card semantic task	verified
text-classification	Transformers / Optimum ONNX / WinML export surface	verified

Model architecture:

XLMRobertaForSequenceClassification
├── XLMRobertaModel
│   ├── Embeddings (vocab 250002, hidden 384)
│   └── Encoder stack x12
│       ├── Self-attention (12 heads)
│       └── Feed-forward + GELU + residual/layernorm
└── XLMRobertaClassificationHead -> logits [B, 1]

Source/confidence: pinned checkpoint config resolves model_type: xlm-roberta, architectures: [XLMRobertaForSequenceClassification], hidden size 384, 12 layers, 12 attention heads, and one output label.

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 XLMRobertaIOConfig, WinML class WinMLModelForSequenceClassification; inputs input_ids and attention_mask int32 [1, 512]; output logits.
Optimum coverage	VENDOR-ONLY for xlm-roberta: vendor tasks include text-classification; WinML adds no extra task override.
Recipe-free build	PASS: build complete in 49.1s; final artifact emitted; 686 exported ONNX nodes, optimized to 444 nodes in the no-analyze baseline build.
Recipe-free perf	PASS on CPU fp32: avg 43.06 ms, p50 40.80 ms, throughput 23.22 samples/sec, RAM total +177.4 MB.
L2 compare	PASS on CPU fp32: cosine mean 1.0, max absolute diff max 1.71661376953125e-05.
L3 task metric	BLOCKED: default text-classification eval uses GLUE/MRPC label alignment and fails on label equivalent; this checkpoint is a single-score reranker, not a MRPC class-label model.

Baseline winml config is the source for the checked-in recipe shape, task, loader, and optimization defaults.

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 dataset/task semantics for this reranking 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_mmarco-mMiniLMv2-L12-H384-v1/cpu/cpu/text-classification_fp32_config.json
examples/recipes/cross-encoder_mmarco-mMiniLMv2-L12-H384-v1/cpu/cpu/text-classification_fp16_config.json

Methodology friction observed: none. Model finding: XLM-RoBERTa single-logit cross-encoder rerankers export cleanly through text-classification; task-metric eval needs a text-ranking dataset/metric rather than the default GLUE/MRPC text-classification path.

Per-EP/device/precision results

Tier	EP / Device	Precision	Verdict	Evidence
L0 build	CPUExecutionProvider / cpu	fp32	PASS	Recipe build complete in 61.9s; optimized model is loadable with inputs input_ids, attention_mask int32 [1, 512], output logits [1, 1], 396 nodes, external data colocated with model.onnx.
L0 build	CPUExecutionProvider / cpu	fp16	PASS	Recipe build complete in 55.1s with --precision fp16; final artifact is about 235.8 MB and reports fp16 at perf time, with 397 nodes.
L1 perf	CPUExecutionProvider / cpu	fp32	PASS	Avg 540.46 ms, p50 553.04 ms, throughput 1.85 samples/sec, RAM total +173.6 MB.
L1 perf	CPUExecutionProvider / cpu	fp16	PASS	Avg 667.73 ms, p50 730.24 ms, throughput 1.50 samples/sec, RAM total +193.9 MB.
L2 compare	CPUExecutionProvider / cpu	fp32	PASS	winml eval --mode compare: cosine mean 1.0, max absolute diff max 1.71661376953125e-05, SQNR mean 114.598 dB.
L2 compare	CPUExecutionProvider / cpu	fp16	PASS	winml eval --mode compare: cosine mean 1.0, max absolute diff max 0.013412952423095703, SQNR mean 58.651 dB.
L3 eval	CPUExecutionProvider / cpu	fp32	CLI-BLOCKED	Default text-classification eval dataset nyu-mll/glue/mrpc fails label alignment for label equivalent; checkpoint emits a single reranking score.
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 for this checkpoint. The fp32 recipe is identical to auto-config for loader/export/optim fields. The fp16 recipe is identical to winml config -p fp16, preserving 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; 12x encoder attention/FFN; sequence-classification head	686 exported nodes tagged during HTP export; 396 optimized ONNX ops analyzed	QNN/OpenVINO NPU: all analyzed op types supported; VitisAI: no rule data.
fp16	embeddings; 12x encoder attention/FFN; sequence-classification head	686 exported nodes tagged during HTP export; 397 optimized ONNX ops analyzed	QNN/OpenVINO NPU: one fp16 Cast op unknown, no unsupported ops; VitisAI: no rule data.

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	396 ops / 19 types	Reshape 120; Gemm 74; Transpose 48; Add 39; LayerNormalization 25; Mul 25; MatMul 24	QNN/OpenVINO: 396 supported, 0 partial, 0 unsupported, 0 unknown; VitisAI: 396 unknown.
fp16	397 ops / 19 types	Reshape 120; Gemm 74; Transpose 48; Add 39; LayerNormalization 25; Mul 25; MatMul 24	QNN/OpenVINO: 396 supported, 0 partial, 0 unsupported, 1 unknown (Cast); VitisAI: 397 unknown.

Reproduce commands

$MODEL = "cross-encoder/mmarco-mMiniLMv2-L12-H384-v1"
$OUT = "temp/repro_cross_encoder_mmarco_mMiniLMv2_L12_H384_v1"

uv run winml inspect -m $MODEL --format json
uv run winml build -c examples/recipes/cross-encoder_mmarco-mMiniLMv2-L12-H384-v1/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_mmarco-mMiniLMv2-L12-H384-v1/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"