xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-12 17:26:43 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Models][Qwen] Replace pad with cat for better performance (#26486)

Browse Source 
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
This commit is contained in:
Lukas Geiger 2025-10-09 15:51:26 +01:00 committed by GitHub
parent e246ad6f0c
commit 2c1c7dfb35
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 6 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/models/dots_ocr.py
View File

@ -680,7 +680,7 @@ class DotsVisionTransformer(nn.Module):
dim=0, dim=0,
dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32, dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
) )
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0) cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
max_seqlen, seqlens = self.compute_attn_mask_seqlen(cu_seqlens) max_seqlen, seqlens = self.compute_attn_mask_seqlen(cu_seqlens)
for blk in self.blocks: for blk in self.blocks:

5
vllm/model_executor/models/ernie45_vl.py
View File

@ -574,11 +574,12 @@ class Ernie4_5_VisionTransformer(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)
zeros = cu_seqlens.new_zeros(1)
if num_pad > 0: if num_pad > 0:
cu_seqlens = F.pad(cu_seqlens, (1, 1), value=0) cu_seqlens = torch.cat([zeros, cu_seqlens, zeros])
cu_seqlens[-1] = cu_seqlens[-2] + num_pad cu_seqlens[-1] = cu_seqlens[-2] + num_pad
else: else:
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0) cu_seqlens = torch.cat([zeros, cu_seqlens])
# add batch size # add batch size
if hidden_states.ndim == 2: if hidden_states.ndim == 2:

2
vllm/model_executor/models/qwen3_vl.py
View File

@ -539,7 +539,7 @@ class Qwen3_VisionTransformer(nn.Module):
dim=0, dim=0,
dtype=grid_thw_tensor.dtype if torch.jit.is_tracing() else torch.int32, dtype=grid_thw_tensor.dtype if torch.jit.is_tracing() else torch.int32,
) )
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0) cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
hidden_states = hidden_states.unsqueeze(1) hidden_states = hidden_states.unsqueeze(1)
rotary_pos_emb = rotary_pos_emb.to(hidden_states.device) rotary_pos_emb = rotary_pos_emb.to(hidden_states.device)

2
vllm/model_executor/models/siglip2navit.py
View File

@ -592,7 +592,7 @@ class Siglip2Encoder(nn.Module):
# for more information # for more information
dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32, dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
) )
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0) cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
reverse_indices = torch.argsort(window_index) reverse_indices = torch.argsort(window_index)