[Bugfix][KV Offload] Bound unaligned SWA loads by physical GPU blocks (#49052)
Signed-off-by: Colton Ottley <colton@ottleyengineering.com> Co-authored-by: Colton Ottley <colton@ottleyengineering.com> Co-authored-by: jasl <jasl9187@hotmail.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> Co-authored-by: Or Ozeri <oro@il.ibm.com>
coltonottley committed
Jul 26, 2026 at 05:53 UTC
5559679229bc961848b121ccdeaa8fa5d79bec98
1 file changed
+1
vllm/distributed/kv_transfer/kv_connector/v1/offloading/scheduler.py
+1
@@ -860,6 +860,7 @@ class OffloadingConnectorScheduler:
860
num_pending_gpu_blocks
861
<= group_config.sliding_window_size_in_chunks
862
* self.config.blocks_per_chunk
863
+ + 1
864
)
865
866
num_chunks = cdiv(num_cached_tokens, tokens_per_chunk)