xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-04-08 01:07:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Remove tile_size=64 for mm_prefix triton attention (#30973)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
(cherry picked from commit d2dc5dfc6ecafbd3d725c1c42dd019db2b1efd30)
This commit is contained in:
Isotr0py 2025-12-19 03:42:32 +08:00 committed by Kevin H. Luu
parent ac43367ced
commit b2eb84de77
1 changed files with 0 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/attention/ops/triton_unified_attention.py
View File

@ -800,7 +800,6 @@ def _get_tile_size(
head_size: int,
sliding_window: int,
element_size: int,
is_mm_prefix: bool,
is_prefill: bool,
) -> int:
"""Select tile size with Gemma3-specific optimization.
@ -809,10 +808,6 @@ def _get_tile_size(
the larger head dimension (128/256). For other models, use
the default vLLM behavior.
"""
if is_mm_prefix:
# Multimodal bidirectional attention needs a larger tile size
return 64
if _is_gemma3_attention(head_size, sliding_window):
# Gemma3: use 32 for decode (default is 16)
return 32
@ -903,14 +898,12 @@ def unified_attention(
head_size,
sliding_window_val,
q.element_size(),
is_mm_prefix=use_mm_prefix,
is_prefill=True,
)
TILE_SIZE_DECODE = _get_tile_size(
head_size,
sliding_window_val,
q.element_size(),
is_mm_prefix=use_mm_prefix,
is_prefill=False,
)