diff --git a/vllm/v1/worker/cp_utils.py b/vllm/v1/worker/cp_utils.py index 2c2e0b5cdb..c5febda691 100644 --- a/vllm/v1/worker/cp_utils.py +++ b/vllm/v1/worker/cp_utils.py @@ -29,10 +29,11 @@ def check_attention_cp_compatibility(vllm_config: VllmConfig) -> None: ) if dcp_size > 1: assert layer_impl.need_to_return_lse_for_decode, ( - "DCP requires attention impls to return" - " the softmax lse for decode, but the impl " - f"{layer_impl.__class__.__name__} " - "does not return the softmax lse for decode." + "Decode Context Parallelism (DCP) requires attention " + "implementations to return the softmax LSE during decode, " + f"but {layer_impl.__class__.__name__} does not. " + "Try a different backend by setting " + "VLLM_ATTENTION_BACKEND or disable DCP." ) if pcp_size > 1: