09. Unique Engineering Highlights

1. Manual ONNX export past Optimum — custom IndicTrans unsupported; wrappers + torch.onnx.export. Evidence: EXPORT_ISSUES.md, 01_export_encoder_decoder.py.

2. Mask kept via zero-cost dependency — logits + mask.sum() * 0 so tracer retains encoder_attention_mask. Evidence: it2_onnx_wrappers.py.

3. Dummy encoder states in decoder_with_past — compile cross-attn while feeding real past KV. Evidence: wrappers + size docs.

4. Tied-weight dedup + shared decoder sidecar — large download size cut for browsers. Evidence: onnx_bundle_optimize.py, ONNX_SIZE_OPTIMIZATION.md.

5. SPM → Fairseq dict ID remap — language tags as added tokens; TemplateProcessing for </s>. Evidence: 02_build_fast_tokenizers.py.

6. FP16 bias type-fix fix — force FP16 initializers; clear stale value_info. Evidence: 05_convert_fp16.py, quantization_issues.md.

7. Q4F16 accuracy_level / EP awareness — exclude embed/lm_head; tune for CPU vs WebGPU. Evidence: 06_quantize_q4f16.py.

8. Batched lang-pair-aware parity decode — practical throughput on 3300 fixtures. Evidence: it2_inference.py.

9. v2 1B export — dynamic heads/dims + early externalization for 2GB protobuf limit. Evidence: src/v2/01_export_encoder_decoder.py.

10. Stale HF sidecar cleanup — delete obsolete per-decoder .data after shared layout. Evidence: scripts/upload_all_with_delete.sh.

Ordinary

• Makefile orchestration
• Hugging Face upload CLI
• Vite dashboard scaffolding
• Matplotlib report plots