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 6, 05:45 UTC
[Perf] Optimize glm4.xv VIT (#37779)
Signed-off-by: Yang <lymailforjob@gmail.com>
Yang Liu committed
Mar 22, 2026 at 02:12 UTC
b0507004625a7e684ee9df1bc9a2058b9258fd5a
1 file changed
+1
-2
vllm/model_executor/models/glm4_1v.py
+1
-2
@@ -758,11 +758,10 @@ class Glm4vVisionTransformer(nn.Module):
758
grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
759
).cumsum(dim=0, dtype=torch.int32)
760
cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
761
- cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
762
-
761
# pre-compute max_seqlen for attn mask to reduce cuMemcpy operations
762
max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens)
763
seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
764
+ cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
765
x = self.embeddings(
766
x, seqlens, grid_thw, image_type_ids[:, 0], image_type_ids[:, 1]
767
)