Skip to content

Fix mis-aligned prompts and completions in colocate mode #3491

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 5 commits into from
May 24, 2025
Merged

Fix mis-aligned prompts and completions in colocate mode #3491

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
1573855
Slice prompts_text back
toslali-ibm May 24, 2025
889402e
Use all_prompts_text in generation for consistency
toslali-ibm May 24, 2025
3c9f5c0
Update trl/trainer/grpo_trainer.py
toslali-ibm May 24, 2025
4f50b59
Update trl/trainer/grpo_trainer.py
toslali-ibm May 24, 2025
577f4b4
Merge branch 'main' into tppromptalign
toslali-ibm May 24, 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: 4 additions & 2 deletions trl/trainer/grpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1057,10 +1057,12 @@ def _generate_and_score_completions(
orig_size = len(prompts_text)
gathered_prompts = [None for _ in range(self.vllm_tensor_parallel_size)]
torch.distributed.all_gather_object(gathered_prompts, prompts_text, group=self.tp_group)
prompts_text = [p for sublist in gathered_prompts for p in sublist]
all_prompts_text = [p for sublist in gathered_prompts for p in sublist]
else:
all_prompts_text = prompts_text

with profiling_context(self, "vLLM.generate"):
all_outputs = self.llm.generate(prompts_text, sampling_params=sampling_params, use_tqdm=False)
all_outputs = self.llm.generate(all_prompts_text, sampling_params=sampling_params, use_tqdm=False)

completion_ids = [output.token_ids for outputs in all_outputs for output in outputs.outputs]

Expand Down
Loading