Enable HPU Fused SDPA for Qwen3-VL vision attention using attention masks #787
Conversation
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
Signed-off-by: slokesha <slokeshappa@habana.ai>
slokesha
force-pushed
the
slokesha/qwen3_enable_mask
branch
from
c245eb8 to
e07204d
Compare
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
slokesha
requested review from
adobrzyn,
afierka-intel,
iboiko-habana,
kamil-kaczor,
ksmusz,
kzawora-intel,
mgawarkiewicz-intel,
michalkuligowski and
xuechendi
as code owners
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
Signed-off-by: Spurthi Lokeshappa <slokeshappa@habana.ai>
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
slokesha
changed the title
Signed-off-by: slokesha <spurthi.lokeshappa@intel.com>
🚧 CI BlockedThe main CI workflow was not started for the following reason:
|
✅ CI PassedAll checks passed successfully against the following vllm commit: |
2 participants
Qwen3-VL vision attention is updated to use FusedSDPA.apply directly when the query sequence length is within the supported fused range (q_len ≤ 65536).
This removes the per-block Q/K/V attention loop and enables the optimized HPU fused SDPA kernel for vision attention.
The change aligns Qwen3-VL with the optimized path already used by Qwen2.5-VL on Gaudi, improving efficiency while preserving identical model outputs.