From b0507004625a7e684ee9df1bc9a2058b9258fd5a Mon Sep 17 00:00:00 2001 From: Yang Liu <127183760+KKSK-DON@users.noreply.github.com> Date: Sun, 22 Mar 2026 02:12:34 -0400 Subject: [PATCH] [Perf] Optimize glm4.xv VIT (#37779) Signed-off-by: Yang --- vllm/model_executor/models/glm4_1v.py | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) diff --git a/vllm/model_executor/models/glm4_1v.py b/vllm/model_executor/models/glm4_1v.py index d806562e0f..786b1175cb 100644 --- a/vllm/model_executor/models/glm4_1v.py +++ b/vllm/model_executor/models/glm4_1v.py @@ -758,11 +758,10 @@ class Glm4vVisionTransformer(nn.Module): grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0] ).cumsum(dim=0, dtype=torch.int32) 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 max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens) seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist() + cu_seqlens = cu_seqlens.to(self.device, non_blocking=True) x = self.embeddings( x, seqlens, grid_thw, image_type_ids[:, 0], image_type_ids[:, 1] )