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 4, 16:15 UTC
[Bugfix] add seq_lens_cpu_upper_bound to CommonAttentionMetadata in mla_runner.py (#40844)
Signed-off-by: ignaciosica <mignacio.sica@gmail.com>
Ignacio Sica committed
Apr 24, 2026 at 20:13 UTC
f88763efc35f8da4d3cfe611a0497d3d3251b9e9
1 file changed
+1
benchmarks/attention_benchmarks/mla_runner.py
+1
@@ -404,6 +404,7 @@ def _build_attention_metadata(
404
query_start_loc=q_start_gpu,
405
query_start_loc_cpu=q_start_cpu,
406
seq_lens=seq_lens_gpu,
407
+ seq_lens_cpu_upper_bound=seq_lens_cpu,
408
_seq_lens_cpu=seq_lens_cpu,
409
_num_computed_tokens_cpu=num_computed_tokens_cpu,
410
slot_mapping=slot_mapping,