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 8, 05:00 UTC
[Bugfix] Fix ARC touch KeyError for non-ready T1 blocks in kv offload (#34576)
Signed-off-by: Yiqi Xue <xuey666@gmail.com>
Yiqi Xue committed
Feb 16, 2026 at 07:33 UTC
bc56a1d56e98a747de7a6c0610673cecc681d808
1 file changed
+2
-1
vllm/v1/kv_offload/arc_manager.py
+2
-1
@@ -90,7 +90,8 @@ class ARCOffloadingManager(OffloadingManager):
90
block = self.t1.pop(block_hash)
91
if not block.is_ready:
92
# block was just prepared to be stored, not really touched twice
93
- self.t1.move_to_end(block_hash)
93
+ # keep it in T1 and mark as most recently used
94
+ self.t1[block_hash] = block
95
else:
96
self.t2[block_hash] = block
97