mpnet / fill-mask + feature-extraction + sentence-similarity: all models pass wmk perf

Summary

MPNet models fail on fill-mask and sentence-similarity with index out of range in self during ONNX export or inference. Feature-extraction also fails on one model due to a Windows symlink issue in the HF cache.

Eval Results (2026-03-11)

9/9 models fail — 0 pass.

Root Cause

Two distinct issues:

1. index out of range in self: MPNet's ONNX export produces a graph with a tensor indexing operation that goes out of bounds. Likely a sequence length or vocab size mismatch in the dummy input generator — MPNet uses a different positional embedding scheme than BERT.
2. onnx.ModelProto size: Multilingual MPNet models are large (>2GB ONNX) — same issue as xlm-roberta (#429). Fix: use external data format for large exports.
3. Windows symlink (exit 3): HF hub cache symlinks broken on Windows without Developer Mode — environment issue, not a ModelKit code bug.

Current State

• No mpnet.py in modelkit/models/hf/ — MPNet relies on Optimum defaults
• modelkit/models/hf/roberta.py exists (RoBERTa config) but MPNet has a different architecture
• No ONNX config registered for mpnet model type

Desired State

All 9 MPNet models pass wmk perf (excluding the Windows-symlink-only failure which is environment).

Acceptance Criteria

• sentence-transformers/all-mpnet-base-v2 passes fill-mask and sentence-similarity
• sentence-transformers/multi-qa-mpnet-base-dot-v1 passes fill-mask, feature-extraction, and sentence-similarity
• sentence-transformers/paraphrase-multilingual-mpnet-base-v2 passes feature-extraction and sentence-similarity
• 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 index out of range: debug MPNet's dummy input shape — MPNet uses max_position_embeddings=514 (514 not 512) and a custom positional embedding offset. The dummy input generator may generate sequence lengths that cause an off-by-one in the embedding lookup.
• For ONNX size: same fix as xlm-roberta — enable use_external_data_format=True for models above size threshold
• Consider adding modelkit/models/hf/mpnet.py with a MpnetIOConfig similar to BertIOConfig

Related Files

• modelkit/models/hf/bert.py — pattern for custom ONNX config with max_position_embeddings
• modelkit/export/io.py — register_onnx_overwrite()
• modelkit/models/hf/__init__.py — where new mpnet.py import would be added
• eval_results/2026-03-11/models/sentence-transformers__all-mpnet-base-v2__fill-mask/result.json