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] Fix MRoPE dispatch on CPU (#24712)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Sep 12, 2025 at 12:56 UTC
7920de0a2af7c663efae29856f1bac061a553b58
1 file changed
+9
vllm/model_executor/layers/rotary_embedding/mrope.py
+9
@@ -300,6 +300,15 @@ class MRotaryEmbedding(RotaryEmbedding):
300
key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
301
return query, key
302
303
+ def forward_cpu(
304
+ self,
305
+ positions: torch.Tensor,
306
+ query: torch.Tensor,
307
+ key: Optional[torch.Tensor] = None,
308
+ offsets: Optional[torch.Tensor] = None,
309
+ ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
310
+ return self.forward_native(positions, query, key, offsets)
311
+
312
@classmethod
313
def get_input_positions(
314
cls,