recipe(deberta): add cross-encoder/nli-deberta-v3-base (text-classification, NLI)

[DingmaomaoBJTU]

Summary

Adds a curated CPU float recipe for cross-encoder/nli-deberta-v3-base — a DeBERTa-v3 cross-encoder for natural-language inference (3-way sequence classification: contradiction / entailment / neutral; config model_type = deberta-v2). Ships fp32 + fp16 variants under cpu/cpu/ (both quant: null; no CPU quantized variant per repo convention).

This is a recipe-only (L0★) contribution: Optimum already covers deberta-v2 text-classification natively, so main builds this model with zero source changes. The delta this PR adds is the curated CPU reference recipe plus an L2 numeric-parity proof that the trained 3-way NLI head is preserved end-to-end (which a plain build does not demonstrate). Claimed tiers: Effort L0★ · Goal ceiling L2 · Outcome L0.

---

1. Recipe path(s)

• examples/recipes/cross-encoder_nli-deberta-v3-base/cpu/cpu/text-classification_fp32_config.json
• examples/recipes/cross-encoder_nli-deberta-v3-base/cpu/cpu/text-classification_fp16_config.json

Both are byte-identical (quant: null float bucket; on CPU both realize as fp32, fp16 materializes on GPU/NPU). opset 17, batch 1, inputs input_ids[1,512] + attention_mask[1,512] (int32), output logits. No token_type_ids — deberta-v2 has type_vocab_size = 0.

2. README row

None. Recipe-only, CPU-only — deliberately not added to the "Total: N (model, task) tuples that pass fp16 eval on all 10 (EP, device) buckets" table, which would be a factual overclaim for a CPU-only recipe (consistent with the reviewed outcome on #1084 / #1112).

3. Build output dir

temp/nli_deberta_v3/ (scratch, gitignored) — model.onnx + model.onnx.data (786 MB fp32).

4. Build log

✅ Build complete — recipe-driven winml build exit 0 (269.5 s). ONNX IR 8, opset 17, external data co-located, output logits[1,3].

5. Appended findings

model_knowledge/deberta.json → deberta-001 (new family). Lane A (skill repo); not part of this model PR's diff.

6. Optimum-coverage probe

deberta-v2 text-classification is VENDOR-ONLY (added_by_winml = []). Vendor onnx tasks: [feature-extraction, fill-mask, multiple-choice, question-answering, text-classification, token-classification]. No winml-added exporter exists or is needed.

7. Claimed (Effort, Goal, Outcome)

L0★ / L2 / L0. Baseline: winml build -m cross-encoder/nli-deberta-v3-base on main (origin/main = 74342698, winml 0.2.0) already PASSES build-only. Contribution = curated CPU float recipe (fp32 + fp16) + L2 trained-head-preservation proof.

8. Goal-ladder verdict table

Tier	Verdict	Evidence
L0 (build)	PASS	recipe-driven build exit 0 (269.5 s); ONNX opset 17, inputs input_ids[1,512]+attention_mask[1,512], output logits[1,3], external data co-located
L1 (perf)	PASS	CPU/fp32 avg 5951 ms, throughput 0.17 samples/s, RAM Δ +618 MB (slow but deterministic; ContextPooler pools fixed position 0, so random-dummy perf is safe)
L2 (numeric vs PyTorch)	PASS	4 real NLI pairs: cosine 1.000000, max-abs 1.9e-6, argmax agreement 4/4 — trained 3-way NLI head + disentangled attention export losslessly

Ceiling L2 reached; no downgrade.

9. Methodology-evolution declaration

No methodology friction this cycle. deberta is a new model family (new model_knowledge/deberta.json), but the existing pipeline handled it without a skill_meta change.

10. Perf & eval data

EP / Device	Precision	Verdict	Mean	p50	Throughput	RAM Δ	Task metric
CPUExecutionProvider / cpu	fp32	PASS	5951 ms	—	0.17 samples/s	+618 MB	N/A (L3 not marched — L2 ceiling)

winml eval not run (Goal ceiling L2). NLI label order for this checkpoint is {0:contradiction, 1:entailment, 2:neutral}, which differs from the GLUE/MNLI dataset order — an eval would need a label remap, so no eval block is shipped (mirrors facebook/bart-large-mnli).

11. Component / op-level data

winml analyze --ep all: 568 total operators, 17 unique. Per-EP op classification: QNN NPU 17/17 supported, OpenVINO NPU 17/17 supported (includes the disentangled-attention GatherElements ops); VitisAI all-unknown (no rule data — analyze exit 1, expected, not a functional failure). Artifact: temp/nli_deberta_v3/analyze_all.json.

12. Reproducible commands

# baseline (main already builds this model, no recipe)
winml build -m cross-encoder/nli-deberta-v3-base -o temp\nli_deberta_v3_baseline

# recipe-driven build (this PR)
winml build -m cross-encoder/nli-deberta-v3-base `
  -c examples\recipes\cross-encoder_nli-deberta-v3-base\cpu\cpu\text-classification_fp32_config.json `
  -o temp\nli_deberta_v3

# L1 perf (CPU)
winml perf -m temp\nli_deberta_v3\model.onnx --iterations 20 --warmup 5 --no-analyze

# op coverage
winml analyze -m temp\nli_deberta_v3\model.onnx --ep all -o temp\nli_deberta_v3\analyze_all.json

# L2 parity vs PyTorch: temp\nli_deberta_v3_l2.py (4 real NLI pairs, pad to 512, int32 inputs, drop token_type_ids)
python temp\nli_deberta_v3_l2.py

[DingmaomaoBJTU]

VERDICT: APPROVE

coverage: partial
not_yet_tested_on: [gpu, npu, full-fp16-eval-matrix]
reachable_verified: CPUExecutionProvider / cpu (fp32 artifact from quant:null CPU recipe)
deferred: GPU/NPU/full fp16 eval matrix are host/charter-deferred for this recipe-only CPU contribution.

Independent evidence gathered:

• Diff scope: git diff --name-only origin/main...df29999a is exactly the two recipe JSON files under examples/recipes/cross-encoder_nli-deberta-v3-base/cpu/cpu/; no source/README/test changes.
• Recipe validation: both JSON files parse as WinMLBuildConfig; SHA256 byte-identical (59e337a7a81a37b3c8c1572b84d1f6006d9f3bcf497bf4656cd533affc13d968); quant: null; opset 17; inputs are input_ids and attention_mask int32 [1,512]; no token_type_ids; output logits.
• Artifact L0: temp\nli_deberta_v3\model.onnx loads in ONNX/ORT; IR 8, opset 17, inputs input_ids/attention_mask, output logits[1,3]; external data model.onnx.data is co-located (824,199,168 bytes).
• L2 parity: inspected temp\nli_deberta_v3_l2.py (loads HF PyTorch model and ORT ONNX separately; no hard-coded pass/ONNX-vs-ONNX comparison) and reran it. Result: min cosine 1.000000, max_abs 1.907349e-06, argmax agreement 4/4 on real NLI pairs.
• Perf sanity: short CPU run with --device cpu --ep cpu --iterations 5 --warmup 2 --no-analyze produced avg 1060.84 ms, throughput 0.94 samples/s, RAM total +627.8 MB (same seconds-per-inference/RAM order as the PR's longer CPU log: 5951 ms, 0.17 samples/s, +618 MB). Running the PR's no-device command on this host auto-selected QNN/NPU, so I used explicit CPU flags for the CPU claim.
• PR body audit: all 12 hand-off sections present; README-row absence is explicitly justified; baseline honestly states main already builds this; label-order caveat is present; no eval block is shipped.
• Optimum probe rerun with import optimum.exporters.onnx.model_configs: deberta-v2 vendor tasks include text-classification; added_by_winml=[], matching L0★/recipe-only. PR checks shown by gh pr checks are passing.

No blocking discrepancies found.