deberta + deberta-v2 / text-classification + zero-shot-classification: all models pass wmk perf

Summary

DeBERTa (v1 xlarge) fails with ONNX protobuf size limit. Multiple mDeBERTa-v3 models (deberta-v2) hit compilation timeouts. This blocks the entire DeBERTa family across text-classification and zero-shot-classification tasks.

Eval Results (2026-03-11)

4 fail, 2 pass — larger models hit TIMEOUT.

Root Cause

1. deberta-xlarge (deberta model_type): Large model ONNX exceeds 2GB protobuf limit — same as xlm-roberta issue (#429). Fix: external data format.
2. mDeBERTa-v3-large (deberta-v2 model_type): Small deberta-v2 models pass (nli-deberta-v3-small ✓), but large variants timeout during compilation or export. DeBERTa-v3 uses a disentangled attention mechanism with complex graph structure that slows compilation on QNN.

Current State

• No deberta.py in modelkit/models/hf/
• cross-encoder/nli-deberta-v3-small passes — confirms deberta-v2 export path works for small models
• Timeout for large deberta-v2 models suggests compilation (QNN graph compose) is the bottleneck

Desired State

All 4 failing DeBERTa models pass wmk perf.

Acceptance Criteria

• microsoft/deberta-xlarge-mnli passes (fix ONNX size via external data format)
• MoritzLaurer/deberta-v3-large-zeroshot-v2.0 passes
• MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli passes
• MoritzLaurer/mDeBERTa-v3-base-mnli-xnli passes
• Fix is universal — no hardcoded model names (CLAUDE.md Cardinal Rule #1)
• uv run pytest tests/ passes (CLAUDE.md Cardinal Rule #3)

Technical Notes

• For deberta-xlarge ONNX size: same fix as xlm-roberta/qwen — enable external data format
• For deberta-v2 timeouts: investigate if compilation timeout is in ONNX export phase or QNN compilation phase. Consider raising the timeout threshold, or adding a DeBERTa-specific optimization config that reduces graph complexity (e.g., attention head fusion)
• mDeBERTa-v3-base-mnli-xnli ran for 3753s (>1h) — indicates a loop/hang, not just slow compilation

Related Files

• modelkit/models/hf/bert.py — pattern for model config
• modelkit/export/config.py — export config with potential timeout/size settings
• eval_results/2026-03-11/models/microsoft__deberta-xlarge-mnli__text-classification/result.json

Start with modelkit/models/hf/bert.py and modelkit/export/config.py, then inspect eval_results/2026-03-11/models/microsoft__deberta-xlarge-mnli__text-classification/result.json. Run the listed wmk perf cases to determine whether the DeBERTa-v2 timeout occurs during export or QNN compilation, and compare external-data handling for related models. Done means all four named models pass wmk perf, uv run pytest tests/ passes, and no model-specific workaround is added.