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 9, 06:15 UTC
[Attention] Clarify comment explaining attn_logits +1 dimension (#33427)
Signed-off-by: Francesco Fusco <ffu@zurich.ibm.com>
Francesco Fusco committed
Jan 31, 2026 at 05:50 UTC
5b55c0bea7410e60566e329b28070ce77269a322
1 file changed
+2
-2
vllm/v1/attention/backends/mla/triton_mla.py
+2
-2
@@ -143,8 +143,8 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
143
B,
144
q_num_heads,
145
num_kv_splits,
146
- # NOTE(lucas) idk why the +1 is here but sglang has it so we
147
- # just mirror that
146
+ # NOTE: the +1 stores the LogSumExp (LSE) that the stage2
147
+ # kernel uses to merge partial attention outputs across splits.
148
self.kv_lora_rank + 1,
149
),
150
dtype=torch.float32,