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
[Models][QwenVL] Remove unnecessary `.contiguous()` calls (#27106)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
Oct 18, 2025 at 16:05 UTC
5c2acb270aad36e35750f617f6219bf95a4924c4
2 files changed
+2
-2
vllm/model_executor/models/qwen2_5_vl.py
+1
-1
@@ -396,7 +396,7 @@ class Qwen2_5_VisionAttention(nn.Module):
396
q, k, v = self.split_qkv(x)
397
batch_size = q.shape[1]
398
399
- q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous() for x in (q, k, v))
399
+ q, k, v = (rearrange(x, "s b ... -> b s ...") for x in (q, k, v))
400
if rotary_pos_emb is not None:
401
# [2 * b, s, heads, head_dim]
402
qk_concat = torch.cat([q, k], dim=0)
vllm/model_executor/models/qwen2_vl.py
+1
-1
@@ -423,7 +423,7 @@ class Qwen2VisionAttention(nn.Module):
423
q, k, v = self.split_qkv(x)
424
batch_size = q.shape[1]
425
426
- q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous() for x in (q, k, v))
426
+ q, k, v = (rearrange(x, "s b ... -> b s ...") for x in (q, k, v))
427
if rotary_pos_emb is not None:
428
# [2 * b, s, heads, head_dim]
429
qk_concat = torch.cat([q, k], dim=0)