Skip to content

fix grpo single card #3246

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
merged 1 commit into from
Feb 24, 2025
Merged

fix grpo single card #3246

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 7 additions & 4 deletions swift/trainers/rlhf_trainer/grpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -243,14 +243,13 @@ def __init__(self,
# Buffer the batch to reuse generated outputs across multiple updates. For more details, see
# `_get_train_sampler` and `_prepare_inputs`.
self._buffered_inputs = [None] * args.gradient_accumulation_steps

self.add_callback(GRPOCallback(self))
if self.args.async_generate:
self.add_callback(GRPOCallback(self))

@property
def infer_rank(self):
rank, local_rank, world_size, local_world_size = get_dist_setting()
assert local_world_size % self.args.num_infer_workers == 0
assert local_world_size + self.args.num_infer_workers == get_device_count()
step = local_world_size // self.args.num_infer_workers
for _vllm_rank in range(self.args.num_infer_workers):
_assigned = _vllm_rank * step
Expand All @@ -263,7 +262,6 @@ def infer_rank(self):
def local_infer_rank(self):
rank, local_rank, world_size, local_world_size = get_dist_setting()
assert local_world_size % self.args.num_infer_workers == 0
assert local_world_size + self.args.num_infer_workers == get_device_count()
step = local_world_size // self.args.num_infer_workers
for _vllm_rank in range(self.args.num_infer_workers):
_assigned = _vllm_rank * step
Expand Down Expand Up @@ -502,6 +500,11 @@ def _generate_and_score_completions(
mode = 'eval' if self.control.should_evaluate else 'train'
completion_length = self.accelerator.gather_for_metrics(outputs['completion_mask'].sum(1)).float().mean().item()
self._metrics[mode]['completion_length'].append(completion_length)
# clip ratio
response_clip_ratio = torch.gt(
self.accelerator.gather_for_metrics(outputs['completion_mask'].sum(1)),
self.args.max_completion_length).float().mean().item()
self._metrics[mode]['response_clip_ratio'].append(response_clip_ratio)
reward_per_func = rewards_per_func.mean(0)
for i, reward_func in enumerate(self.reward_funcs):
if isinstance(reward_func, nn.Module): # Module instead of PretrainedModel for compat with compiled models
Expand Down