Skip to content

indicconformer_export

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

indicconformer_export

README.md

IndicConformer ONNX Export

This folder will hold the ONNX export pipeline for AI4Bharat's IndicConformer (Hybrid CTC-RNNT Conformer, 22 official Indian languages). It parallels scripts/nemo_export/ but uses AI4Bharat's NeMo fork rather than upstream NeMo, so the two envs must stay isolated.

See ../../README.md for the full project context and IndicConformer_plan.md once promoted, or the 6-phase plan kept in ~/Programming/parakeet_csharp/data/IndicConformer_plan.md.

Status

Phase 1 — Feasibility spike (in progress). Nothing C#-side is worth touching until CTC export + parity are proven.

Environment

Use Python 3.10, 3.11, or 3.12. The AI4Bharat fork requires Python >= 3.10.

python3 scripts/indicconformer_export/setup_indicconformer_env.py
source .venv-indicconformer-export/bin/activate

This env is separate from .venv-nemo-export/ on purpose: the fork pins different NeMo/Transformer internals, and mixing them would break Parakeet export.

Install with a specific CUDA build or CPU-only:

# CUDA 12.1
python3 scripts/indicconformer_export/setup_indicconformer_env.py --cuda-version cu121
# CPU only
python3 scripts/indicconformer_export/setup_indicconformer_env.py --cuda-version ""

Phase 1 — Discovery

Before writing any exporter, load a checkpoint and dump its structure:

python scripts/indicconformer_export/inspect_indicconformer.py \
  --hf-repo ai4bharat/indicconformer_stt_hi_hybrid_ctc_rnnt_large \
  --out /tmp/indicconformer_hi.json

What we're looking for in the report:

  • has_ctc_head — must be true (CTC-head export is the whole plan)
  • ctc_decoder_param_names — are there per-language weight tensors? That would confirm AI4Bharat's "multi-softmax" design
  • tokenizer_langs / tokenizer_class — AggregateTokenizer with 22 langs is the shape we expect for the unified model
  • cfg_preprocessor — confirms we can reuse the DFT preprocessor from nemo_export/ (same 80-mel Conformer frontend, presumably)

Once that report lands, Phase 1's open question — "how is language selected at inference?" — is answered, and we can commit Phase 2's export shape (single ONNX file with a language-id input vs. one ONNX per language).

References

Publishing

After running repackage_600m_indicconformer.py to consolidate the external data, build the manifest and upload to christopherthompson81/indicconformer-600m-onnx using the shared tools at the scripts/ root:

python scripts/make_manifest.py \
    --model-dir ~/models/indicconformer_600m_onnx \
    --files encoder-model.onnx encoder-model.onnx.data \
            ctc_decoder-model.onnx nemo128.onnx \
            vocab.txt language_spans.json config.json

python scripts/upload_to_hf.py \
    --model-dir ~/models/indicconformer_600m_onnx \
    --repo-id christopherthompson81/indicconformer-600m-onnx \
    --sync-readme --create-repo

The model card is sourced from scripts/hf_readmes/indicconformer-600m-onnx/README.md; edit it there, commit, and re-run with --sync-readme to update HF.