Skip to content

Remove grpo requirement bs=num_generations #3609

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

Merged
danielhanchen merged 3 commits into from
Nov 20, 2025
+25 −1
Merged

Remove grpo requirement bs=num_generations #3609

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c6485c8
Remove grpo requirement bs=num_generations
mmathew23 Nov 17, 2025
cc38926
Update rl.py
danielhanchen Nov 20, 2025
e27afa5
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Nov 20, 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
17 changes: 13 additions & 4 deletions unsloth/models/rl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -822,12 +822,21 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
extra_args += check_dr_grpo

# Check GRPO num_generations mismatch
if "per_device_train_batch_size" in call_args and "num_generations" in call_args:
if (
"per_device_train_batch_size" in call_args
and "num_generations" in call_args
and "steps_per_generation" in call_args
and "generation_batch_size" in call_args
):
# if world size is not set by accelerate or torchrun at this point it will be 1
check_num_generations = (
"if (per_device_train_batch_size // num_generations) * num_generations != per_device_train_batch_size:\n"
" print('Unsloth: We now expect `per_device_train_batch_size` to be a multiple of `num_generations`.\\n"
"if steps_per_generation is None and generation_batch_size is None:\n"
" ga = gradient_accumulation_steps\n"
" world_size = int(os.environ.get('WORLD_SIZE', '1'))\n"
" if (ga * world_size * per_device_train_batch_size) % num_generations != 0:\n"
" print('Unsloth: We now expect `per_device_train_batch_size` * `gradient_accumulation_steps` * `world_size` to be a multiple of `num_generations`.\\n"
"We will change the batch size of ' + str(per_device_train_batch_size) + ' to the `num_generations` of ' + str(num_generations))\n"
" per_device_train_batch_size = num_generations\n"
" per_device_train_batch_size = num_generations\n"
"\n"
)
extra_args += check_num_generations
Comment on lines 825 to 842
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot Nov 18, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

While this new relaxed check is a good improvement, it appears the old, stricter check from grpo_trainer_fix_batch_size in unsloth/models/rl_replacements.py is still being applied after this new logic.

The old check unconditionally forces per_device_train_batch_size to be a multiple of num_generations, which counteracts the goal of this pull request.

To ensure this change is effective, the old validation logic should be removed from the grpo_trainer_fix_batch_size function.

Copy link
Collaborator Author

@mmathew23 mmathew23 Nov 18, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No this is incorrect. It only forces the constraint in a specific case.

Expand Down
1 change: 1 addition & 0 deletions unsloth/models/rl_replacements.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -853,6 +853,7 @@ def compute_loss(
# TRL warns if batch size is not a multiple of num_generations -> fix this.
def grpo_trainer_fix_batch_size(RLTrainer_source, RLConfig_source):
if "divisible by the number of generations" not in RLTrainer_source:
# in later trl versions this doesn't exist anymore
return ""
if "num_generations" not in RLConfig_source:
return ""
Expand Down