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 1, 21:15 UTC
fix: pad dummy run query_start_loc (#44603)
Signed-off-by: UranusSeven <109661872+UranusSeven@users.noreply.github.com>
Uranus committed
Jun 5, 2026 at 15:43 UTC
d2f70da116257d02f34d12bf478c974ea9179dac
1 file changed
+3
vllm/v1/worker/gpu_model_runner.py
+3
@@ -5787,6 +5787,9 @@ class GPUModelRunner(
5787
num_scheduled_tokens, self.query_pos.np
5788
)
5789
self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
5790
+ self.query_start_loc.np[num_reqs + 1 : num_reqs_padded + 1].fill(
5791
+ cum_num_tokens[-1]
5792
+ )
5793
self.query_start_loc.copy_to_gpu()
5794
5795
# Sync block table CPU->GPU so cleared rows from