Skip to content

[rl] Using JobConfig as the centralized config system for inference and simple GRPO #2188

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wwwjn wants to merge 3 commits into from
Closed

[rl] Using JobConfig as the centralized config system for inference and simple GRPO #2188

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
7f3d05f
config sys v1
wwwjn Dec 30, 2025
1092b4d
config sys for simple_grpo v1
wwwjn Dec 30, 2025
ac14471
config sys for simple_grpo v2
wwwjn Dec 31, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions torchtitan/experiments/rl/unified/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,10 +66,10 @@ python torchtitan/experiments/rl/unified/infer.py --model-ckpt-path <path_to_mod

5. Run simple rl loop
```
VLLM_BATCH_INVARIANT=1 VLLM_ATTENTION_BACKEND=FLASH_ATTN python3 torchtitan/experiments/rl/unified/simple_rl_multiprocess.py
VLLM_BATCH_INVARIANT=1 VLLM_ATTENTION_BACKEND=FLASH_ATTN python3 torchtitan/experiments/rl/unified/simple_grpo.py \
--job.config_file torchtitan/experiments/rl/unified/run_configs/qwen3_0.6b.toml
```
Right now we only support VLLM_COMPAT mode, which could achieve trainer and generator bitwise identical. We are working on support UNIFIED mode,
which uses a unified model definition for trainer and generator.
We uses a unified model definition for trainer and generator, which could achieve trainer and generator bitwise identical.

## TODO
Work on batch invariance:
Expand Down
Loading
Loading