bart + mbart / summarization + text-classification + zero-shot-classification: all models pass wmk perf

Summary

All BART-family models fail across summarization, text-classification, and zero-shot-classification tasks. Two distinct errors: (1) EncoderDecoderCache JIT type error during seq2seq export, and (2) index -1 is out of bounds for dimension 1 with size 0 during classification inference.

Eval Results (2026-03-11)

9/9 models fail — 0 pass.

Root Cause

Two separate issues:

1. Seq2seq (summarization): Same EncoderDecoderCache TorchScript incompatibility as T5 family — see T5 issue.
2. Classification (text-classification / zero-shot-classification): BART uses a decoder with a special EOS token as the classification token. The input generation produces an empty decoder input, causing index -1 is out of bounds for dimension 1 with size 0 when the model tries to extract logits from the last token position.

Current State

• No bart.py model config in modelkit/models/hf/
• No ONNX config override registered for bart model type
• Relies on Optimum defaults which do not handle BART's classification head correctly

Desired State

All 9 BART/mBART models above pass wmk perf.

Acceptance Criteria

• All 9 models listed above pass wmk perf
• Fix is universal — no hardcoded model names (CLAUDE.md Cardinal Rule #1)
• uv run pytest tests/ passes (CLAUDE.md Cardinal Rule #3)
• Eval re-run confirms 9/9 pass

Technical Notes

• For seq2seq: Same fix as T5 — use_cache=False during export or custom BartOnnxConfig
• For classification: BART's BartForSequenceClassification uses decoder_input_ids starting with EOS. The dummy input generator must provide valid decoder_input_ids. Register a BartIOConfig via @register_onnx_overwrite("bart", "text-classification", ...) with correct decoder input shape.
• eval_results/2026-03-11/models/facebook__bart-large-mnli__text-classification/result.json shows the ONNX export succeeds, but inference fails — meaning the ONNX graph is valid but the input shape is wrong at perf time.
• Consider a unified encoder-decoder ONNX config base class for bart/t5/marian

Related Files

• modelkit/models/hf/bert.py — pattern for custom ONNX config
• modelkit/export/io.py — register_onnx_overwrite()
• modelkit/models/hf/__init__.py:47-54 — add bart to MODEL_BUILD_CONFIGS
• eval_results/2026-03-11/models/facebook__bart-large-mnli__text-classification/result.json
• eval_results/2026-03-11/models/google-t5__t5-small__summarization/result.json

References

• CLAUDE.md Cardinal Rule #1: No hardcoded model type strings
• CLAUDE.md Cardinal Rule #3: Run uv run pytest tests/ after implementation

Start with modelkit/models/hf/bert.py and modelkit/export/io.py, then inspect modelkit/models/hf/init.py:47-54 and the referenced result.json files. Run the listed BART perf cases and the test suite to confirm the failures; done means all nine models pass wmk perf and uv run pytest tests/ passes without hardcoded model names.