Unverified Commit 3dd60971 authored by Srreyansh Sethi's avatar Srreyansh Sethi Committed by GitHub
Browse files

[feat]: make DCP error msg clearer (#28443)


Signed-off-by: default avatarvnadathur <glvikramn@gmail.com>
Signed-off-by: default avatarWorldExplored <srreyansh.sethi@gmail.com>
Signed-off-by: default avatarSrreyansh Sethi <107075589+WorldExplored@users.noreply.github.com>
Co-authored-by: default avatarvnadathur <glvikramn@gmail.com>
Co-authored-by: default avatarvnadathur <236933696+vnadathur@users.noreply.github.com>
parent a5b17fba
...@@ -29,10 +29,11 @@ def check_attention_cp_compatibility(vllm_config: VllmConfig) -> None: ...@@ -29,10 +29,11 @@ def check_attention_cp_compatibility(vllm_config: VllmConfig) -> None:
) )
if dcp_size > 1: if dcp_size > 1:
assert layer_impl.need_to_return_lse_for_decode, ( assert layer_impl.need_to_return_lse_for_decode, (
"DCP requires attention impls to return" "Decode Context Parallelism (DCP) requires attention "
" the softmax lse for decode, but the impl " "implementations to return the softmax LSE during decode, "
f"{layer_impl.__class__.__name__} " f"but {layer_impl.__class__.__name__} does not. "
"does not return the softmax lse for decode." "Try a different backend by setting "
"VLLM_ATTENTION_BACKEND or disable DCP."
) )
if pcp_size > 1: if pcp_size > 1:
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment