Skip to content

[Core] Remove FlashAttention block size restriction for hybrid models #36701

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
tdoublep wants to merge 1 commit into
base: main
Choose a base branch
from
+0 −17
Open

[Core] Remove FlashAttention block size restriction for hybrid models #36701

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: 0 additions & 17 deletions vllm/v1/attention/backends/flash_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,7 +35,6 @@
)
from vllm.config import (
VllmConfig,
get_current_vllm_config,
get_current_vllm_config_or_none,
get_layers_from_vllm_config,
)
Expand Down Expand Up @@ -67,22 +66,6 @@ class FlashAttentionBackend(AttentionBackend):

@staticmethod
def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
vllm_config = get_current_vllm_config()
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
if (
model_config
and model_config.is_hybrid
and (
cache_config.mamba_ssm_cache_dtype == "float32"
or cache_config.mamba_cache_dtype == "float32"
)
):
# NOTE(tdoublep): while in principle, FA supports
# MultipleOf(16), these are the block sizes that do not
# suffer from the NaN propagation problem described here:
# https://github.com/Dao-AILab/flash-attention/issues/1974
return [16, 32, 64]
return [MultipleOf(16)]

forward_includes_kv_cache_update: bool = False
Expand Down