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 5, 08:30 UTC
[feat]: make DCP error msg clearer (#28443)
Signed-off-by: vnadathur <glvikramn@gmail.com> Signed-off-by: WorldExplored <srreyansh.sethi@gmail.com> Signed-off-by: Srreyansh Sethi <107075589+WorldExplored@users.noreply.github.com> Co-authored-by: vnadathur <glvikramn@gmail.com> Co-authored-by: vnadathur <236933696+vnadathur@users.noreply.github.com>
Srreyansh Sethi committed
Apr 9, 2026 at 22:27 UTC
3dd60971de5ac6752cf7c546db43534b39c996e1
1 file changed
+5
-4
vllm/v1/worker/cp_utils.py
+5
-4
@@ -29,10 +29,11 @@ def check_attention_cp_compatibility(vllm_config: VllmConfig) -> None:
29
)
30
if dcp_size > 1:
31
assert layer_impl.need_to_return_lse_for_decode, (
32
- "DCP requires attention impls to return"
33
- " the softmax lse for decode, but the impl "
34
- f"{layer_impl.__class__.__name__} "
35
- "does not return the softmax lse for decode."
32
+ "Decode Context Parallelism (DCP) requires attention "
33
+ "implementations to return the softmax LSE during decode, "
34
+ f"but {layer_impl.__class__.__name__} does not. "
35
+ "Try a different backend by setting "
36
+ "VLLM_ATTENTION_BACKEND or disable DCP."
37
)
38
39
if pcp_size > 1: