Important

NeMo 2.0 is an experimental feature and currently released in the dev container only: nvcr.io/nvidia/nemo:dev. Please refer to the Migration Guide for information on getting started.

ONNX Export of Megatron Models

This guide demonstrates the usage of the ONNX export functionality for Megatron models.

Requirements

Set up the development environment by launching the latest NeMo container

The minimum version requirements for NeMo and TransformerEngine are below

nemo > 1.19
transformer_engine > 0.10

Export to ONNX

The export script supports the ONNX export of models with .nemo and .ckpt file extensions. The script also supports the export of the following types of models: GPT, T5, BERT, BART, NMT, RETRO. Commands for both file formats are discussed in the following sections. The model type used for the examples is GPT.

Export using .nemo file

A model with .nemo file extension can be exported using the command below

python3 examples/nlp/language_modeling/megatron_export.py \
    model_type=gpt \
    onnx_model_file=gpt_126m.onnx \
    gpt_model_file=gpt_126m.nemo

Export using .ckpt file

A model with .ckpt file extension can be exported using the command below

python3 examples/nlp/language_modeling/megatron_export.py \
    model_type=gpt \
    onnx_model_file=gpt_126m.onnx \
    checkpoint_dir=./gpt_126m/ \
    checkpoint_name=model_weights.ckpt \
    hparams_file=./gpt_126m/hparams.yaml