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 15, 04:45 UTC
[Misc] [ROCm] Prevent surplus tensor reshape (#19803)
Signed-off-by: Zsolt Borbely <zsolt.borbely@htecgroup.com>
zsolt-borbely-htec committed
Jun 19, 2025 at 07:57 UTC
aa20d10a9182677ffc419e2d823e00237f56eb0d
1 file changed
+1
-1
vllm/v1/attention/backends/triton_attn.py
+1
-1
@@ -376,7 +376,7 @@ class TritonAttentionImpl(AttentionImpl):
376
query.reshape(
377
(num_tokens, num_heads * head_size)).contiguous(),
378
layer._q_scale)
379
- query = query.reshape((num_tokens, num_heads, head_size))
379
+ query = query.reshape((num_tokens, num_heads, head_size))
380
381
use_local_attn = \
382
(self.use_irope and attn_metadata.local_attn_metadata is not None)