Skip to content

Fix HPU prompt_token_ids device placement for penalty sampling #1465

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
iboiko-habana merged 3 commits into from
May 25, 2026
+4 −2
Merged

Fix HPU prompt_token_ids device placement for penalty sampling #1465

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
96c6dba
Fix HPU prompt_token_ids device placement for penalty sampling
yeonsily May 19, 2026
fab67e7
Merge branch 'main' into dev/presence_penalty_fix
iboiko-habana May 20, 2026
286b300
Merge branch 'main' into dev/presence_penalty_fix
iboiko-habana May 21, 2026
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 vllm_gaudi/v1/worker/hpu_input_batch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -640,12 +640,14 @@ def make_selective_sampling_metadata(
# The prompt tokens are used only for applying penalties during
# the sampling process. Hence copy these tensors only when
# there are requests which need penalties to be applied.
prompt_token_ids = self._make_prompt_token_ids_cpu_tensor()[req_indices]
prompt_token_ids = self._make_prompt_token_ids_cpu_tensor()[req_indices].to(device=self.device,
non_blocking=True)
Comment on lines +643 to +644
else:
# Even with skip_copy=True, we need prompt_token_ids for penalties
if not self.no_penalties:
cached_tensor = self._get_cached_prompt_token_ids()
prompt_token_ids = cached_tensor[req_indices] if cached_tensor is not None else None
prompt_token_ids = cached_tensor[req_indices].to(
device=self.device, non_blocking=True) if cached_tensor is not None else None
Comment on lines +649 to +650
Comment on lines +649 to +650
else:
prompt_token_ids = None

Expand Down
Loading