Skip to content

[Bugfix][CPU] Voxtral realtime: scale attention metadata out of place #55218

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

Closed
alex-chaiko wants to merge 1 commit into from
+21 −7
Closed

[Bugfix][CPU] Voxtral realtime: scale attention metadata out of place #55218

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
28 changes: 21 additions & 7 deletions vllm/model_executor/models/whisper_causal.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,13 +158,27 @@ def build(
fast_build: bool = False,
) -> AttentionMetadata:
new_common_attn_metadata = copy.deepcopy(common_attn_metadata)
new_common_attn_metadata.query_start_loc *= block_pool_size
new_common_attn_metadata.query_start_loc_cpu *= block_pool_size
new_common_attn_metadata.seq_lens *= block_pool_size
if new_common_attn_metadata._seq_lens_cpu is not None:
new_common_attn_metadata._seq_lens_cpu *= block_pool_size
if new_common_attn_metadata._num_computed_tokens_cpu is not None:
new_common_attn_metadata._num_computed_tokens_cpu *= block_pool_size
# Scale out of place: the V1 CPU model runner points every
# `CpuGpuBuffer.gpu` at its `.cpu` tensor, so a device field and its
# `_cpu` counterpart can share storage (`deepcopy` preserves the
# sharing), and in-place scaling would apply `block_pool_size` twice.
new_common_attn_metadata.query_start_loc = (
common_attn_metadata.query_start_loc * block_pool_size
)
new_common_attn_metadata.query_start_loc_cpu = (
common_attn_metadata.query_start_loc_cpu * block_pool_size
)
new_common_attn_metadata.seq_lens = (
common_attn_metadata.seq_lens * block_pool_size
)
if common_attn_metadata._seq_lens_cpu is not None:
new_common_attn_metadata._seq_lens_cpu = (
common_attn_metadata._seq_lens_cpu * block_pool_size
)
if common_attn_metadata._num_computed_tokens_cpu is not None:
new_common_attn_metadata._num_computed_tokens_cpu = (
common_attn_metadata._num_computed_tokens_cpu * block_pool_size
)
new_common_attn_metadata.num_actual_tokens *= block_pool_size
new_common_attn_metadata.max_query_len *= block_pool_size
new_common_attn_metadata.max_seq_len *= block_pool_size
Expand Down
Loading