minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Delete dead rematch path in SessionAwareCache.release_session (#22735)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-04-13 17:02:40 -07:00
committed by GitHub
parent 9fb00ede15
commit 33a3ba256f
3 changed files with 10 additions and 155 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/unit/mem_cache/test_streaming_session_unit.py
-82
View File
@@ -111,88 +111,6 @@ def test_streaming_release_kv_cache_trims_overallocated_tail(monkeypatch):
assert allocator.freed[0].tolist() == list(range(32, 40))
def test_release_session_recomputes_current_tree_owned_prefix():
page_size = 16
req_to_token = torch.arange(128, dtype=torch.int32).reshape(1, 128)
req_to_token_pool = SimpleNamespace(req_to_token=req_to_token, free_slots=[])
allocator = _FakeAllocator()
full_match = MatchResult(
device_indices=torch.tensor(list(range(16)) + list(range(64, 96))),
last_device_node="stale-expanded",
last_host_node="stale-expanded",
)
protected_match = MatchResult(
device_indices=torch.tensor(list(range(16))),
last_device_node="current-protected",
last_host_node="current-protected",
)
inner = _FakeInnerCache(
req_to_token_pool,
allocator,
page_size,
match_results=[full_match, protected_match],
)
tree_cache = SessionAwareCache(inner)
tree_cache.slots["session-a"] = SessionSlot(
req_pool_idx=0,
kv_committed_len=48,
kv_allocated_len=48,
last_node="outdated-node",
cache_protected_len=32,
)
req = _FakeReq("session-a", req_pool_idx=0, committed=48, allocated=48)
tree_cache.release_session("session-a", req)
assert inner.dec_lock_ref_calls == ["current-protected"]
assert req_to_token_pool.free_slots == [0]
assert len(allocator.freed) == 1
assert allocator.freed[0].tolist() == list(range(16, 48))
def test_release_session_never_grows_tree_owned_prefix():
page_size = 16
req_to_token = torch.arange(128, dtype=torch.int32).reshape(1, 128)
req_to_token_pool = SimpleNamespace(req_to_token=req_to_token, free_slots=[])
allocator = _FakeAllocator()
overmatched = MatchResult(
device_indices=torch.tensor(list(range(48))),
last_device_node="overmatched-node",
last_host_node="overmatched-node",
)
capped_match = MatchResult(
device_indices=torch.tensor(list(range(16))),
last_device_node="original-lock-node",
last_host_node="original-lock-node",
)
inner = _FakeInnerCache(
req_to_token_pool,
allocator,
page_size,
match_results=[overmatched, capped_match],
)
tree_cache = SessionAwareCache(inner)
tree_cache.slots["session-a"] = SessionSlot(
req_pool_idx=0,
kv_committed_len=48,
kv_allocated_len=48,
last_node="outdated-node",
cache_protected_len=16,
)
req = _FakeReq("session-a", req_pool_idx=0, committed=48, allocated=48)
tree_cache.release_session("session-a", req)
assert inner.dec_lock_ref_calls == ["original-lock-node"]
assert req_to_token_pool.free_slots == [0]
assert len(allocator.freed) == 1
assert allocator.freed[0].tolist() == list(range(16, 48))
def test_match_prefix_abort_does_not_restore_live_session_slot():
req_to_token = torch.arange(256, dtype=torch.int32).reshape(2, 128)
req_to_token_pool = SimpleNamespace(req_to_token=req_to_token, free_slots=[])