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
llama4_vision_rope: add HIP override to accept (q, k) and avoid (positions, q, k) mismatch (#26790)
Signed-off-by: Huamin Li <3ericli@gmail.com>
Huamin Li committed
Oct 14, 2025 at 11:54 UTC
87efc681dbd57ab8e522c1d5cf50dc6ee5f7db2d
1 file changed
+7
vllm/model_executor/layers/rotary_embedding/llama4_vision_rope.py
+7
@@ -78,3 +78,10 @@ class Llama4VisionRotaryEmbedding(RotaryEmbedding):
78
key: torch.Tensor | None = None,
79
) -> tuple[torch.Tensor, torch.Tensor | None]:
80
return self.forward_native(query, key)
81
+
82
+ def forward_hip( # type: ignore[override]
83
+ self,
84
+ query: torch.Tensor,
85
+ key: torch.Tensor | None = None,
86
+ ) -> tuple[torch.Tensor, torch.Tensor | None]:
87
+ return self.forward_native(query, key)