[Bugfix][KV Offload] count appended GPU blocks in store group_sizes (#42945)
Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>
This commit is contained in:
@@ -710,7 +710,6 @@ class OffloadingConnectorScheduler:
|
||||
|
||||
offloaded_block_idx = start_block_idx + idx
|
||||
gpu_block_idx = offloaded_block_idx * block_size_factor
|
||||
num_group_blocks += block_size_factor
|
||||
for i in range(block_size_factor):
|
||||
block_id = block_ids[gpu_block_idx + i]
|
||||
if block_id == 0:
|
||||
@@ -720,6 +719,7 @@ class OffloadingConnectorScheduler:
|
||||
elif start_gpu_block_idx is None:
|
||||
start_gpu_block_idx = gpu_block_idx + i
|
||||
src_block_ids.append(block_id)
|
||||
num_group_blocks += 1
|
||||
if is_sliding_window:
|
||||
sliding_window_block_ids.append(block_id)
|
||||
else:
|
||||
|
||||
Reference in New Issue
Block a user