How baidu-baige/loongforge works
LoongForge is a multi-backend framework for training LLMs, VLMs, diffusion, and embodied models. Its principal workflow prepares training data, constructs a model, and executes pre-training or SFT; the embodied stack adds dataset transforms, distributed training support, and policy evaluation. The README documents both Megatron and torch-native backends and broad model coverage. Sampled source confirms the LLM dataset/model paths and selected embodied transform and evaluation behavior; other subsystem links reflect documented capabilities rather than inspected call sites.
Main entry points: runners Benchmark Runners · pretrain_llm.py LLM Pretraining · sft_llm.py LLM SFT · run.py Evaluation Orchestrator
Components 21
Megatron Training
- LLM Pretraining
loongforge/train/pretrain/pretrain_llm.py - LLM SFT
loongforge/train/sft/sft_llm.py - GPT Datasets
loongforge/train/pretrain/pretrain_llm.py - SFT Datasets
loongforge/train/sft/sft_llm.py - LLM Provider
loongforge/models/foundation/llm_model_provider.py - MoE Routing
loongforge/models/foundation/ernie4_5_vl/ernie_router.py
Embodied Stack
- Robot Datasets
loongforge/embodied/data/datasets/dataset_builder.py - Sample Transforms
loongforge/embodied/data/datasets/transforms/pipeline.py - Embodied Models
loongforge/embodied/model/registry.py
Policy Evaluation
- Evaluation Orchestrator
loongforge/embodied/eval/orchestrator/run.py - Benchmark Runners
loongforge/embodied/eval/orchestrator/runners - Server Manager
loongforge/embodied/eval/orchestrator/server_manager.py - Policy Server
loongforge/embodied/eval/servers/loongforge_server.py - Policy Client
loongforge/embodied/eval/transport/client.py - Evaluation Results
loongforge/embodied/eval/metrics/results.py
Training Runtime
- Distributed Runtime
loongforge/embodied/distributed/parallel.py - Checkpointing
loongforge/embodied/distributed/checkpoint.py - Optimizers
loongforge/embodied/optimizer/optimizer.py
Other components
- Training User
- Evaluation User
- Robot Benchmark
Connections 19
- Training User → LLM Pretraining: starts training
- Training User → LLM SFT: starts fine-tuning
- LLM Pretraining → GPT Datasets: builds datasets
- LLM SFT → SFT Datasets: builds datasets
- LLM Pretraining → LLM Provider: uses model
- LLM SFT → LLM Provider: uses model
- LLM Provider → MoE Routing: routes experts
- Training User → Robot Datasets: configures training
- Robot Datasets → Sample Transforms: builds transforms
- Embodied Models → Distributed Runtime: runs across ranks
- Distributed Runtime → Checkpointing: saves state
- Training User → Evaluation Orchestrator: starts evaluation
- Evaluation Orchestrator → Benchmark Runners: dispatches benchmark
- Evaluation Orchestrator → Server Manager: manages server
- Server Manager → Policy Server: starts process
- Benchmark Runners → Policy Client: requests actions
- Policy Client → Policy Server: sends inference
- Robot Benchmark → Benchmark Runners: provides episodes
- Benchmark Runners → Evaluation Results: records metrics