recipe(lt-wikidata-comp-multi): add verified CPU fp32/fp16 recipes

[codykk]

Add verified CPU recipes (fp32 + fp16) for dell-research-harvard/lt-wikidata-comp-multi. This XLM-RoBERTa checkpoint produces multilingual embeddings for Wikidata entity comparison. Effort L0, Goal L1 (build + perf), Outcome L0 (recipe only). Both precisions build and perf successfully on CPU.

Model metadata

What the model does: Produces dense embeddings from multilingual text for Wikidata entity comparison and linking tasks, fine-tuned from XLM-RoBERTa-large on a multilingual Wikidata entity comparison dataset.

Primary user stories:

• User supplies multilingual entity descriptions to obtain dense embeddings for entity comparison, deduplication, and linking across Wikidata knowledge bases.

Supported tasks:

• feature-extraction — Optimum vendor-supported (XLMRobertaModel); winml resolves via TasksManager with generic inference class.

Model architecture:

XLMRobertaModel
├── Embeddings (vocab 250002 x 1024)
├── Encoder stack x 24
│   ├── Self-attention (16 heads, 1024 dim)
│   ├── Feed-forward (1024 → 4096 → 1024, GELU)
│   └── Residual + LayerNorm
└── [No task head — feature-extraction outputs last_hidden_state]

• Source/confidence: pinned checkpoint config.json (architectures: ["XLMRobertaModel"], model_type: "xlm-roberta") and Optimum vendor Wav2Vec2OnnxConfig export config (verified).

Validation and support evidence

Baseline

• main commit: 5deebd42 | winml: 0.2.0
• Optimum probe: VENDOR-ONLY — feature-extraction registered by Optimum for xlm-roberta; winml adds nothing.
• Build: ✅ PASS (auto-config, no recipe) — AutoModel / feature-extraction
• Perf: ✅ PASS — CPU avg 120.74 ms, 8.28 samples/sec
• Goal floor: L1 (build + perf inherited from main)

Goal

• Effort: L0 — recipe only, no source edits; Optimum vendor fully covers xlm-roberta feature-extraction.
• Goal ceiling: L1 — baseline already demonstrates L0 + L1; L2 (numeric vs PyTorch) not built into CLI; L3 not applset for feature-extraction).
• Outcome: L0 — recipe + report.

Outcome

Highest reached: L1 PASS (build + perf on both precisions). Coverage: full (all target tuples passed). Deferred tuples: none.

Shipped recipes:

• examples/recipes/dell-research-harvard_lt-wikidata-comp-multi/cpu/cpu/feature-extraction_fp32_config.json
• examples/recipes/dell-research-harvard_lt-wikidata-comp-multi/cpu/cpu/feature-extraction_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	120.74 ms	120.09 ms	8.28 samples/sec	+445.2 MB
L0	CPUExecutionProvider / cpu	fp16	PASS	—	—	—	—
L1	CPUExecutionProvider / cpu	fp16	PASS	—	—	4.65 samples/sec	—

Delta

Recipe delta vs winml config (auto-config baseline): identical — filed for verified cpu/cpu coverage. The production recipe README (examples/recipes/README.md) remains untouched.

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; 24x encoder attention/FFN (Gemm, MatMul, Softmax, Gelu, LayerNorm)	392 total ops (after autoconf: gelu_fusion, matmul_add_fusion)	None actionable — CPU is rule-less

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	392 ops / 18 types	Reshape 120; Gemm 72; Transpose 48; Add 39; Mul 25; LayerNormalization 25; MatMul 24	CP-less EP)

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

Reproduce commands

# fp32
winml build -m dell-research-harvard/lt-wikidata-comp-multi `
  -c examples/recipes/dell-research-harvard_lt-wikidata-comp-multi/cpu/cpu/feature-extraction_fp32_config.json `
  -o $OUT --ep cpu --device cpu
winml perf -m $OUT/model.onnx --ep cpu --device cpu

# fp16
winml build -m dell-research-harvard/lt-wikidata-comp-multi `
  -c examples/recipes/dell-research-harvard_lt-wikidata-comp-multi/cpu/cpu/feature-extraction_fp16_config.json `
  -o $OUT_FP16 --ep cpu --device cpu -p fp16
winml perf -m $OUT_FP16/model.onnx --ep cpu --device cpu