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 14, 05:30 UTC
[Bugfix] Add missing VIT backend dispatch on CPU (#24623)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Sep 11, 2025 at 13:28 UTC
29799ddacc9234591ce33e0e8e818fae20565a10
1 file changed
+2
-1
vllm/attention/layer.py
+2
-1
@@ -399,7 +399,8 @@ class MultiHeadAttention(nn.Module):
399
key,
400
value,
401
scale=self.scale)
402
- elif self.attn_backend == _Backend.TORCH_SDPA:
402
+ elif (self.attn_backend == _Backend.TORCH_SDPA
403
+ or self.attn_backend == _Backend.TORCH_SDPA_VLLM_V1):
404
query, key, value = (x.transpose(1, 2)
405
for x in (query, key, value))
406
out = F.scaled_dot_product_attention(query,