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, 00:30 UTC
[Bugfix][CPU] Fix MRoPE dispatch on the CPU backend (#27800)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Oct 30, 2025 at 23:12 UTC
eebf00cb0c925404672d407674b319ebc5ae3a84
1 file changed
+9
vllm/model_executor/layers/rotary_embedding/mrope.py
+9
@@ -357,6 +357,15 @@ class MRotaryEmbedding(RotaryEmbeddingBase):
357
key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
358
return query, key
359
360
+ def forward_cpu(
361
+ self,
362
+ positions: torch.Tensor,
363
+ query: torch.Tensor,
364
+ key: torch.Tensor | None = None,
365
+ offsets: torch.Tensor | None = None,
366
+ ) -> tuple[torch.Tensor, torch.Tensor | None]:
367
+ return self.forward_native(positions, query, key, offsets)
368
+
369
@staticmethod
370
def get_next_input_positions(
371
mrope_position_delta: int,