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 3, 10:30 UTC
[Bugfix][KV Offload] count appended GPU blocks in store group_sizes (#42945)
Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>
Kfir Toledo committed
May 18, 2026 at 14:36 UTC
e414e1f1c020108593526b706efaf89e427c05a2
1 file changed
+1
-1
vllm/distributed/kv_transfer/kv_connector/v1/offloading/scheduler.py
+1
-1
@@ -710,7 +710,6 @@ class OffloadingConnectorScheduler:
710
711
offloaded_block_idx = start_block_idx + idx
712
gpu_block_idx = offloaded_block_idx * block_size_factor
713
- num_group_blocks += block_size_factor
713
for i in range(block_size_factor):
714
block_id = block_ids[gpu_block_idx + i]
715
if block_id == 0:
@@ -720,6 +719,7 @@ class OffloadingConnectorScheduler:
719
elif start_gpu_block_idx is None:
720
start_gpu_block_idx = gpu_block_idx + i
721
src_block_ids.append(block_id)
722
+ num_group_blocks += 1
723
if is_sliding_window:
724
sliding_window_block_ids.append(block_id)
725
else: