How yongliang-wu/dft works
This repository implements Dynamic Fine-Tuning (DFT) experiments on top of VERL: the documented path is to prepare math data, train a model with the FSDP DFT trainer, then generate and grade benchmark answers. VERL also contains distributed RL training, rollout, checkpointing, and model support. The graph combines README-documented workflow with source-index references and shown excerpts; links involving unsampled trainer or evaluation internals are omitted or marked unverified.
Main entry points: fsdp_workers.py FSDP Workers · fsdp_dft_trainer.py DFT Trainer · predict_results_16.py Benchmark Generation · evaluate.py Math Evaluation
Components 17
DFT Training
- DFT Trainer
verl/verl/trainer/fsdp_dft_trainer.py - Training Data
verl/verl/utils/dataset/rl_dataset.py
VERL Runtime
- FSDP Workers
verl/verl/workers/fsdp_workers.py - Model Loading
verl/verl/utils/model.py - Generation Rollout
verl/verl/workers/rollout/vllm_rollout/vllm_rollout_spmd.py - Worker Dispatch
verl/verl/single_controller/base/decorator.py - Batch Protocol
verl/verl/protocol.py - FSDP Checkpoints
verl/verl/utils/checkpoint/fsdp_checkpoint_manager.py - Model Support
verl/verl/models/transformers/monkey_patch.py
Math Evaluation
- Benchmark Generation
verl/eval/math500/predict_results_16.py - Math Evaluation
math_evaluation/evaluate.py - Answer Parser
math_evaluation/parser.py - LaTeX Parser
math_evaluation/latex2sympy/latex2sympy2.py - Math Grader
math_evaluation/grader.py - Benchmark Results
verl/eval/math500/eval_verify.py
Other components
- Researcher
- Evaluator
Connections 15
- Researcher → DFT Trainer: launches
- DFT Trainer → Training Data: uses data
- DFT Trainer → FSDP Workers: dispatches training
- FSDP Workers → Model Loading: loads model
verl/verl/workers/fsdp_workers.py - FSDP Workers → Generation Rollout: creates rollout
verl/verl/workers/fsdp_workers.py - FSDP Workers → FSDP Checkpoints: uses checkpoints
verl/verl/workers/fsdp_workers.py - FSDP Workers → Worker Dispatch: registers methods
verl/verl/workers/fsdp_workers.py - Worker Dispatch → Batch Protocol: dispatches batches
verl/verl/single_controller/base/decorator.py - Model Loading → Model Support: uses model support
- Evaluator → Benchmark Generation: runs evaluation
- Benchmark Generation → Model Support: uses model support
- Math Evaluation → Answer Parser: parses answers
- Answer Parser → LaTeX Parser: converts LaTeX
math_evaluation/parser.py - Math Evaluation → Math Grader: grades answers
- Benchmark Generation → Benchmark Results: produces predictions