Language Models

Megatron Core supports the following language model architectures for large-scale training.

Converting HuggingFace Models

Use Megatron Bridge to convert HuggingFace models to Megatron format. Megatron Bridge is the official standalone converter with support for an extensive list of models including LLaMA, Mistral, Mixtral, Qwen, DeepSeek, Gemma, Phi, Nemotron, and many more.

See the Megatron Bridge supported models list for the complete and up-to-date list.

Decoder-Only Models

Model	Description	Key Features
GPT	Generative Pre-trained Transformer	Standard autoregressive LM, foundational architecture
LLaMA	Meta’s LLaMA family	Efficient architecture with RoPE, SwiGLU, RMSNorm
Mistral	Mistral AI models	Sliding window attention, efficient inference
Mixtral	Sparse Mixture-of-Experts	8x7B MoE architecture for efficient scaling
Qwen	Alibaba’s Qwen series	HuggingFace integration, multilingual support
Mamba	State Space Model	Subquadratic sequence length scaling, efficient long context

Encoder-Only Models

Model	Description	Key Features
BERT	Bidirectional Encoder Representations	Masked language modeling, classification tasks

Encoder-Decoder Models

Model	Description	Key Features
T5	Text-to-Text Transfer Transformer	Unified text-to-text framework, sequence-to-sequence

Retrieval-Augmented Models

Model	Description	Key Features
RETRO	Retrieval-Enhanced Transformer	Retrieval-augmented generation, knowledge grounding

Example Scripts

Training examples for these models can be found in the examples/ directory:

• examples/gpt3/ - GPT-3 training scripts

• examples/llama/ - LLaMA training scripts

• examples/mixtral/ - Mixtral MoE training

• examples/mamba/ - Mamba training scripts

• examples/bert/ - BERT training scripts

• examples/t5/ - T5 training scripts

• examples/retro/ - RETRO training scripts

Model Implementation

All language models are built using Megatron Core’s composable transformer blocks, enabling:

• Flexible parallelism strategies (TP, PP, DP, EP, CP)

• Mixed precision training (FP16, BF16, FP8)

• Distributed checkpointing

• Efficient memory management