Skip to content

fix: DRY sampler never applies during decoding (stale token history views) #1720

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

Open
Nakanokensetsu wants to merge 1 commit into
base: main
Choose a base branch
from
+7 −10
Open

fix: DRY sampler never applies during decoding (stale token history views) #1720

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
17 changes: 7 additions & 10 deletions aphrodite/v1/worker/gpu_input_batch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1261,9 +1261,13 @@ def _make_sampling_metadata(self) -> SamplingMetadata:
output_token_ids = cast(list[list[int]], self.req_output_token_ids) if needs_output_token_ids else []
output_token_ids_tensor = None
token_history_ids, token_history_lens = None, None
token_history_ids_cpu, token_history_lens_cpu = (
self._get_token_history_cpu_views(num_reqs) if not self.no_dry else (None, None)
)
# The DRY CPU kernel cannot be fed from token_ids_cpu here: cached views
# freeze at metadata-build width and hide every generated token, while
# live full-width views expose async-scheduling placeholder ids (-1) at
# sampling time. Passing None makes Sampler.apply_dry fall back to the
# persistent state, which update_dry_state keeps in sync with the real
# sampled ids each step.
token_history_ids_cpu, token_history_lens_cpu = (None, None)

allowed_token_ids_mask: torch.Tensor | None = None
if not self.no_allowed_token_ids:
Expand Down Expand Up @@ -1414,13 +1418,6 @@ def _make_output_token_ids_tensor(self, num_reqs: int) -> torch.Tensor:
)
return output_token_ids_cpu_tensor.to(device=self.device, non_blocking=True)

def _get_token_history_cpu_views(self, num_reqs: int) -> tuple[torch.Tensor, torch.Tensor]:
max_history_len = int(self.num_tokens_no_spec[:num_reqs].max()) if num_reqs else 0
return (
self.token_ids_cpu_tensor[:num_reqs, :max_history_len],
self.num_tokens_no_spec_cpu_tensor[:num_reqs],
)

def _make_token_history_tensors(self, num_reqs: int) -> tuple[torch.Tensor, torch.Tensor]:
history_lens = torch.as_tensor(
self.num_tokens_no_spec[:num_reqs],
Expand Down
Loading