Skip to content

[ROCm] Check that AITER MHA is not selected with sinks #34998

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
gshtras wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[ROCm] Check that AITER MHA is not selected with sinks #34998

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0dd8394
Check for sinks requirement for AITER MHA selector
gshtras Feb 20, 2026
1066c4f
Add check for sinks when explicitly selected
gshtras Feb 20, 2026
4f44ba9
Merge branch 'main' into check_sinks_aiter_mha
gshtras Feb 23, 2026
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
15 changes: 13 additions & 2 deletions vllm/platforms/rocm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -323,6 +323,11 @@ def get_attn_backend_cls(
return AttentionBackendEnum.ROCM_ATTN.get_path()

if selected_backend == AttentionBackendEnum.ROCM_AITER_FA:
if attn_selector_config.has_sink:
raise ValueError(
f"The selected backend, {selected_backend.name}, "
"does not support sinks."
)
if on_gfx9():
logger.info("Using Aiter Flash Attention backend.")
return AttentionBackendEnum.ROCM_AITER_FA.get_path()
Expand All @@ -345,7 +350,12 @@ def get_attn_backend_cls(

# Priority 2: Check for AITER MHA (Flash Attention)
# Only use if explicitly enabled (not just VLLM_ROCM_USE_AITER=1)
if envs.VLLM_ROCM_USE_AITER and envs.VLLM_ROCM_USE_AITER_MHA and on_gfx9():
if (
envs.VLLM_ROCM_USE_AITER
and envs.VLLM_ROCM_USE_AITER_MHA
and on_gfx9()
and not attn_selector_config.has_sink
):
logger.info("Using Aiter Flash Attention backend.")
return AttentionBackendEnum.ROCM_AITER_FA.get_path()

Expand All @@ -365,7 +375,8 @@ def get_attn_backend_cls(
if (
envs.VLLM_ROCM_USE_AITER
and on_gfx9()
and envs.VLLM_ROCM_USE_AITER_MHA is not False
and envs.VLLM_ROCM_USE_AITER_MHA
and not attn_selector_config.has_sink
):
logger.info("Using Aiter Flash Attention backend.")
return AttentionBackendEnum.ROCM_AITER_FA.get_path()
Expand Down