obscura/vllm
2
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2026-08-13 01:08:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

[KV Offload] Return None from lookup() for in-flight blocks (#41795)

Browse Source 
Signed-off-by: Ronen Schaffer <[email protected]>
This commit is contained in:
Ronen Schaffer
2026-05-06 17:31:21 +03:00
committed by GitHub
parent 6467213a9f
commit f39bcf1e30
2 changed files with 11 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/v1/kv_offload/cpu/test_manager.py
+6 -6
View File
@@ -163,9 +163,9 @@ def test_cpu_manager():
),
)
# lookup [1, 2] -> not ready
assert cpu_manager.lookup(to_key(1), _EMPTY_REQ_CTX) is False
assert cpu_manager.lookup(to_key(2), _EMPTY_REQ_CTX) is False
# lookup [1, 2] -> write in-flight, not yet ready
assert cpu_manager.lookup(to_key(1), _EMPTY_REQ_CTX) is None
assert cpu_manager.lookup(to_key(2), _EMPTY_REQ_CTX) is None
# no events so far
assert list(cpu_manager.take_events()) == []
@@ -296,9 +296,9 @@ class TestARCPolicy:
),
)
# lookup [1, 2] -> not ready
assert cpu_manager.lookup(to_key(1), _EMPTY_REQ_CTX) is False
assert cpu_manager.lookup(to_key(2), _EMPTY_REQ_CTX) is False
# lookup [1, 2] -> write in-flight, not yet ready
assert cpu_manager.lookup(to_key(1), _EMPTY_REQ_CTX) is None
assert cpu_manager.lookup(to_key(2), _EMPTY_REQ_CTX) is None
# no events so far
assert list(cpu_manager.take_events()) == []
vllm/v1/kv_offload/cpu/manager.py
+5 -1
View File
@@ -86,7 +86,11 @@ class CPUOffloadingManager(OffloadingManager):
def lookup(self, key: OffloadKey, req_context: ReqContext) -> bool | None:
block = self._policy.get(key)
return block is not None and block.is_ready
if block is None:
return False
if not block.is_ready:
return None # write in-flight; caller should retry
return True
def prepare_load(
self,