A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 05:00 UTC
[Bugfix] Remove `tile_size=64` for mm_prefix triton attention (#30973)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Dec 19, 2025 at 03:42 UTC
d2dc5dfc6ecafbd3d725c1c42dd019db2b1efd30
1 file changed
-7
vllm/attention/ops/triton_unified_attention.py
-7
@@ -800,7 +800,6 @@ def _get_tile_size(
800
head_size: int,
801
sliding_window: int,
802
element_size: int,
803
- is_mm_prefix: bool,
803
is_prefill: bool,
804
) -> int:
805
"""Select tile size with Gemma3-specific optimization.
@@ -809,10 +808,6 @@ def _get_tile_size(
808
the larger head dimension (128/256). For other models, use
809
the default vLLM behavior.
810
"""
812
- if is_mm_prefix:
813
- # Multimodal bidirectional attention needs a larger tile size
814
- return 64
815
-
811
if _is_gemma3_attention(head_size, sliding_window):
812
# Gemma3: use 32 for decode (default is 16)
813
return 32
@@ -903,14 +898,12 @@ def unified_attention(
898
head_size,
899
sliding_window_val,
900
q.element_size(),
906
- is_mm_prefix=use_mm_prefix,
901
is_prefill=True,
902
)
903
TILE_SIZE_DECODE = _get_tile_size(
904
head_size,
905
sliding_window_val,
906
q.element_size(),
913
- is_mm_prefix=use_mm_prefix,
907
is_prefill=False,
908
)
909