From 5b55c0bea7410e60566e329b28070ce77269a322 Mon Sep 17 00:00:00 2001 From: Francesco Fusco <159429316+fuscof-ibm@users.noreply.github.com> Date: Sat, 31 Jan 2026 05:50:30 +0100 Subject: [PATCH] [Attention] Clarify comment explaining attn_logits +1 dimension (#33427) Signed-off-by: Francesco Fusco --- vllm/v1/attention/backends/mla/triton_mla.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/v1/attention/backends/mla/triton_mla.py b/vllm/v1/attention/backends/mla/triton_mla.py index 2403dcc61..f6c1790f6 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,