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
[Bugfix][ROCm][DeepSeek] Fix for forward_hip in rope for DeepSeek (#27373)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Gregory Shtrasberg committed
Oct 23, 2025 at 13:43 UTC
0825197bee8dea547f2ab25f48afd8aea0cd2578
1 file changed
+2
-5
vllm/model_executor/layers/rotary_embedding/base.py
+2
-5
@@ -165,11 +165,8 @@ class RotaryEmbedding(CustomOp):
165
self.rotary_dim,
166
self.is_neox_style,
167
)
168
- else:
169
- # ops.rotary_embedding() is an in-place operation
170
- # that updates the query and key tensors.
171
- self.forward_cuda(positions, query, key)
172
- return query, key
168
+ return query, key
169
+ return self.forward_cuda(positions, query, key)
170
171
def forward_xpu(
172
self,