diff --git a/vllm/v1/attention/backends/mla/triton_mla.py b/vllm/v1/attention/backends/mla/triton_mla.py index 2403dcc613..f6c1790f60 100644 --- a/vllm/v1/attention/backends/mla/triton_mla.py +++ b/vllm/v1/attention/backends/mla/triton_mla.py @@ -143,8 +143,8 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]): B, q_num_heads, num_kv_splits, - # NOTE(lucas) idk why the +1 is here but sglang has it so we - # just mirror that + # NOTE: the +1 stores the LogSumExp (LSE) that the stage2 + # kernel uses to merge partial attention outputs across splits. self.kv_lora_rank + 1, ), dtype=torch.float32,