marian + m2m_100 / translation: all models pass wmk perf

Summary

All Marian (Helsinki-NLP opus-mt) and M2M-100 (facebook/nllb) translation models fail during ONNX export due to EncoderDecoderCache being an unsupported TorchScript type — the same root cause as the T5 and BART issues.

Eval Results (2026-03-11)

6/6 models fail — 0 pass.

Error: Only tuples, lists and Variables are supported as JIT inputs/outputs. Here, received an input of unsupported type: EncoderDecoderCache

Root Cause

Same as T5/BART: recent transformers EncoderDecoderCache type is not serializable by TorchScript. Marian (MarianMTModel) is an encoder-decoder seq2seq model.

Current State

• No marian.py model config in modelkit/models/hf/
• Relies on Optimum default export, which does not bypass EncoderDecoderCache
• facebook/nllb-200-distilled-600M uses m2m_100 model_type — same issue

Desired State

All 6 translation models above pass wmk perf.

Acceptance Criteria

• All 6 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 6/6 pass

Technical Notes

• If a shared encoder-decoder ONNX config fix is implemented for T5/BART, it should automatically cover marian and m2m_100 as well
• Marian model_type: marian; NLLB model_type: m2m_100
• Both use MarianMTModel / M2M100ForConditionalGeneration — both are seq2seq with decoder cache
• Fix: register ONNX config overrides or patch export to set use_cache=False for all encoder-decoder models universally

Related Files

• modelkit/models/hf/bert.py — pattern for ONNX config registration
• modelkit/export/io.py — register_onnx_overwrite()
• eval_results/2026-03-11/models/Helsinki-NLP__opus-mt-nl-en__translation/result.json
• eval_results/2026-03-11/models/facebook__nllb-200-distilled-600M__translation/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
• Related: T5 issue, BART issue (same root cause — coordinate fix)

Start with modelkit/export/io.py and the ONNX configuration pattern in modelkit/models/hf/bert.py, then compare the related T5 and BART fixes. Reproduce the failures with the listed Marian and NLLB models using wmk perf, and inspect the referenced result.json files. Done means all six models pass, the fix is universal rather than model-name-specific, and uv run pytest tests/ passes.