diff --git a/vllm/models/kimi_k3/amd/mla.py b/vllm/models/kimi_k3/amd/mla.py index 1e4a51e8bf..dd93ce439a 100644 --- a/vllm/models/kimi_k3/amd/mla.py +++ b/vllm/models/kimi_k3/amd/mla.py @@ -50,8 +50,9 @@ class KimiK3MultiHeadLatentAttentionWrapper(MultiHeadLatentAttentionWrapper): if attn.hisparse_cache is not None or attn.q_pad_num_heads is not None: return False # Only the W_UK variants reproduced in _fused_decode are supported. - return attn.is_aiter_triton_fp4_bmm_enabled or ( - not attn.is_aiter_triton_fp8_bmm_enabled and attn.W_UK_T is not None + return ( + attn.is_aiter_triton_fp4_bmm_enabled + or not attn.is_aiter_triton_fp8_bmm_enabled ) def get_input_quant_key(self) -> QuantKey | None: