biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Perf] Optimize glm4.xv VIT (#37779)

Browse Source 
Signed-off-by: Yang <lymailforjob@gmail.com>
This commit is contained in:
Yang Liu
2026-03-22 02:12:34 -04:00
committed by GitHub
parent 5dac719b2b
commit b050700462
1 changed files with 1 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/model_executor/models/glm4_1v.py
View File

@@ -758,11 +758,10 @@ class Glm4vVisionTransformer(nn.Module):
grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0] grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
).cumsum(dim=0, dtype=torch.int32) ).cumsum(dim=0, dtype=torch.int32)
cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens]) cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
# pre-compute max_seqlen for attn mask to reduce cuMemcpy operations # pre-compute max_seqlen for attn mask to reduce cuMemcpy operations
max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens) max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens)
seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist() seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
x = self.embeddings( x = self.embeddings(
x, seqlens, grid_thw, image_type_ids[:, 0], image_type_ids[:, 1] x, seqlens, grid_thw, image_type_ids[:, 0], image_type_ids[:, 1]
) )