[Kimi-K3][AMD] Return KDA and MLA projection outputs directly (#50592)

Signed-off-by: Liuyinfeng01 <yinfeliu@amd.com>
Co-authored-by: Liuyinfeng01 <199041580+LiuYinfeng01@users.noreply.github.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
yinfengLiu
2026-09-21 15:44:41 -05:00
committed by GitHub
co-authored by Liuyinfeng01 Andreas Karatzas
parent 9fe1cbd463
commit 382970ee6c
7 changed files with 170 additions and 33 deletions
-11
View File
@@ -922,14 +922,12 @@ class KimiDecoderLayer(nn.Module):
aux_stream=aux_stream,
run_gemm_rs_ar=run_gemm_rs_ar,
)
self._self_attn_writes_output = False
else:
self.self_attn = KimiLinearGatedDeltaNetAttention(
config,
vllm_config,
prefix=f"{prefix}.self_attn",
)
self._self_attn_writes_output = True
else:
qk_nope_head_dim = config.qk_nope_head_dim
qk_rope_head_dim = config.qk_rope_head_dim
@@ -959,7 +957,6 @@ class KimiDecoderLayer(nn.Module):
aux_stream=aux_stream,
run_gemm_rs_ar=run_gemm_rs_ar,
)
self._self_attn_writes_output = False
if self.use_sequence_parallel:
self.self_attn.o_proj.reduce_results = False
@@ -1025,14 +1022,6 @@ class KimiDecoderLayer(nn.Module):
positions: torch.Tensor,
hidden_states: torch.Tensor,
) -> torch.Tensor:
if self._self_attn_writes_output:
output = torch.empty_like(hidden_states)
self.self_attn(
hidden_states=hidden_states,
positions=positions,
output=output,
)
return output
return self.self_attn(
hidden_states=hidden_states,
positions=positions,