recipe(ms-marco-MiniLM-L12-v2): add CPU fp32 and fp16 recipes

[codykk]

Add verified CPU recipes (fp32 + fp16) for cross-encoder/ms-marco-MiniLM-L12-v2, a BERT-based cross-encoder fine-tuned on MS MARCO for passage reranking. Effort L0, Goal L1 (build + perf), Outcome L0 (recipe only). Complements existing L4-v2 and L6-v2 recipes in the same cross-encoder family.

Model metadata

What the model does: Cross-encoder reranker that scores query-passage relevance for information retrieval, fine-tuned from MiniLM-L12 on the MS MARCO passage ranking dataset.

Primary user stories:

• User supplies a query-passage pair to obtain a relevance score for passage reranking in search pipelines.

Supported tasks:

• text-classification — Optimum vendor-supported (BertForSequenceClassification); winml resolves via TasksManager.

Model architecture:

BertForSequenceClassification
├── Embeddings (vocab 30522 x 384)
├── Encoder stack x 12
│   ├── Self-attention (12 heads, 384 dim)
│   ├── Feed-forward (384 → 1536 → 384, GELU)
│   └── Residual + LayerNorm
└── Classification head (384 → 1: relevance logit, Tanh)

• Source/confidence: pinned checkpoint config.json (architectures: ["BertForSequenceClassification"], model_type: "bert") (verified).

Validation and support evidence

Baseline

• main commit: 5deebd42 | winml: 0.2.0
• Optimum probe: VENDOR-ONLY — text-classification registered by Optimum for bert; winml adds nothing.
• Build: ✅ PASS (auto-config, no recipe) — AutoModelForSequenceClassification / text-classification
• Perf: ✅ PASS — CPU avg 68.10 ms, 14.68 samples/sec
• Goal floor: L1 (build + perf inherited from main)

Goal

• Effort: L0 — recipe only, no source edits.
• Goal ceiling: L1 — baseline already demonstrates L0 + L1.
• Outcome: L0 — recipe + report.

Outcome

Highest reached: L1 PASS. Coverage: full. Deferred tuples: none.

Shipped recipes:

• examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/text-classification_fp32_config.json
• examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/text-classification_fp16_config.json

Per-EP/device/precision results

Tier	EP / Device	Precision	Verdict	Mean	p50	Throughput	RAM Δ
L0	CPUExecutionProvider / cpu	fp32	PASS	—	—	—	—
L1	CPUExecutionProvider / cpu	fp32	PASS	57.86 ms	57.67 ms	17.55 samples/sec	+178.4 MB
L0	CPUExecutionProvider / cpu	fp16	PASS	—	—	—	—
L1	CPUExecutionProvider / cpu	fp16	PASS	—	—	8.43 samples/sec	—

Delta

Recipe delta vs winml config: identical — filed for verified cpu/cpu coverage. Production recipe README unchanged.

Analyze summary — component level and op level

Static rule analysis completed; this is compatibility analysis, not runtime execution.

Component-level summary

Artifact	Architecture coverage	Mapping	Actionable EP findings
fp32	embeddings (3x Gather); 12x encoder attention/FFN (Gemm, MatMul, Softmax, Gelu, LayerNorm); classification heas (after autoconf: gelu_fusion, matmul_add_fusion)	None actionable — CPU is rule-less

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	388 ops / 16 types	Reshape 120; Gemm 74; Transpose 48; Add 38; LayerNormalization 25; Mul 24; MatMul 24	CPU: all 16 types unknown (rule-less EP)

Rule-less EPs (CPU): all operator types classified as unknown — no runtime check rules available.

Reproduce commands

# fp32
winml build -m cross-encoder/ms-marco-MiniLM-L12-v2 `
  -c examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/text-classification_fp32_config.json `
  -o $OUT --ep cpu --device cpu
winml perf -m $OUT/model.onnx --ep cpu --device cpu

# fp16
winml build -m cross-encoder/ms-marco-MiniLM-L12-v2 `
  -c examples/recipes/cross-encoder_ms-marco-MiniLM-L12-v2/cpu/cpu/text-classification_fp16_config.json `
  -o $OUT_FP16 --ep cpu --device cpu -p fp16
winml perf -m $OUT_FP16/model.onnx --ep cpu --device cpu