megatron-hf-converter — Claude Code skill guide

To convert HuggingFace checkpoints to Megatron-LM format, run hf_to_megatron.sh with your HF model path and target Megatron path. This automates the HF→Megatron workflow via Megatron-Bridge, enabling scalable training pipelines.

Yes, you can export Megatron checkpoints back to HuggingFace format for inference. The conversion process includes tokenizer handling options and optional patching to align the exported model with the target architecture configuration.

You validate weight equivalence after model conversion by running round-trip validation. This checks weight mappings and inference consistency to catch mapping errors early and ensures numerical equivalence between the converted checkpoints.

Yes, you need Megatron-Bridge to convert checkpoints between HuggingFace and Megatron-LM. The Skill requires Megatron-Bridge along with PyTorch, transformers, and safetensors to automate the conversion and validate numerical equivalence.

The best way to handle tokenizer mismatches during Megatron to HuggingFace export is to use tokenizer patching. This adjusts tokenizers post-export to align with the target architecture configuration when needed.