How nvidia-nemo/megatron-bridge works
NeMo Megatron Bridge is a model-development library for converting model checkpoints and configuring large-scale LLM, multimodal, and diffusion workloads. The principal user workflow is to bring a Hugging Face model into Megatron-compatible form, configure and train or fine-tune it, then save or convert checkpoints for inference. The graph also preserves inference, data preparation, PEFT, and diffusion as distinct capabilities. Most runtime relationships are not exposed in the excerpts, so edges are limited to documented capabilities or indexed source references; unsampled implementation links are omitted.
Main entry points: auto_bridge.py AutoBridge · model_load_save.py Model load and save · model_provider.py Model providers · config.py Training configuration
Components 24
Model conversion
- AutoBridge
src/megatron/bridge/models/conversion/auto_bridge.py - Model bridge
src/megatron/bridge/models/conversion/model_bridge.py - Parameter mapping
src/megatron/bridge/models/conversion/param_mapping.py - Mapping registry
src/megatron/bridge/models/conversion/mapping_registry.py - HF checkpoint state
src/megatron/bridge/models/hf_pretrained/state.py - PEFT conversion
src/megatron/bridge/models/conversion/peft_bridge.py - MiMo conversion
src/megatron/bridge/models/megatron_mimo/conversion/orchestrator.py
Training and data
- Model providers
src/megatron/bridge/models/model_provider.py - GPT provider
src/megatron/bridge/models/gpt_provider.py - Training data
src/megatron/bridge/data - Training configuration
src/megatron/bridge/training/config.py - Model load and save
src/megatron/bridge/training/model_load_save.py - Checkpointing
src/megatron/bridge/training/checkpointing.py - Training launcher
scripts/training/run_recipe.py
Inference
- Text generation
src/megatron/bridge/inference/text_generation.py - VLM controller
src/megatron/bridge/inference/vlm/vlm_inference_controller.py - VLM engine
src/megatron/bridge/inference/vlm/vlm_engine.py - Inference scripts
scripts/inference
Diffusion
- Flow training
src/megatron/bridge/diffusion/common/flow_matching/flow_matching_pipeline.py - FLUX inference
src/megatron/bridge/diffusion/models/flux/flow_matching/flux_inference_pipeline.py - WAN inference
src/megatron/bridge/diffusion/models/wan/flow_matching/flow_inference_pipeline.py - Diffusion recipes
src/megatron/bridge/diffusion/recipes
Other components
- Model engineer
- Hugging Face
Connections 23
- Model engineer → AutoBridge: convert models
- Hugging Face → AutoBridge: load checkpoint
- AutoBridge → Model bridge: uses bridge
src/megatron/bridge/models/conversion/auto_bridge.py - AutoBridge → Parameter mapping: uses mappings
src/megatron/bridge/models/conversion/auto_bridge.py - AutoBridge → HF checkpoint state: loads state
src/megatron/bridge/models/conversion/auto_bridge.py - Mapping registry → Parameter mapping: registers mappings
src/megatron/bridge/models/conversion/mapping_registry.py - AutoBridge → PEFT conversion: supports PEFT
src/megatron/bridge/models/conversion/auto_bridge.py - Model engineer → Training launcher: launch training
- Training launcher → Training configuration: configures run
- Training configuration → Model load and save: loads model
src/megatron/bridge/training/model_load_save.py - Model load and save → Checkpointing: uses checkpointing
src/megatron/bridge/training/model_load_save.py - Model load and save → Model providers: builds model
src/megatron/bridge/training/model_load_save.py - Model providers → GPT provider: selects provider
- Model engineer → Training data: prepare data
- Model engineer → Inference scripts: run inference
- Inference scripts → Text generation: serves text
- VLM engine → VLM controller: schedules requests
src/megatron/bridge/inference/vlm/vlm_engine.py - VLM controller → Text generation: extends generation
- Model engineer → Diffusion recipes: run diffusion
- Diffusion recipes → Flow training: configures training
- Model engineer → FLUX inference: generate images
- Model engineer → WAN inference: generate video
- MiMo conversion → AutoBridge: extends conversion
src/megatron/bridge/models/megatron_mimo/conversion/orchestrator.py