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 13, 23:15 UTC
[Bugfix] Fix KDA output (#27905)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Nov 1, 2025 at 11:54 UTC
3a5de7d2d6e65b6580c3ceb905334843a7b6dd6f
1 file changed
+2
-3
vllm/model_executor/layers/kda.py
+2
-3
@@ -259,7 +259,7 @@ class KimiDeltaAttention(nn.Module, MambaBase):
259
hidden_states: torch.Tensor,
260
positions: torch.Tensor,
261
output: torch.Tensor,
262
- ) -> torch.Tensor:
262
+ ) -> None:
263
num_tokens = hidden_states.size(0)
264
q = self.q_proj(hidden_states)[0]
265
k = self.k_proj(hidden_states)[0]
@@ -291,8 +291,7 @@ class KimiDeltaAttention(nn.Module, MambaBase):
291
)
292
core_attn_out = self.o_norm(core_attn_out, g2)
293
core_attn_out = rearrange(core_attn_out, "1 n h d -> n (h d)")
294
-
295
- return self.o_proj(core_attn_out)[0]
294
+ output[:] = self.o_proj(core_attn_out)[0]
295
296
def _forward(
297
self,