diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py index d713a374a0..81230b84c3 100644 --- a/python/sglang/srt/managers/scheduler.py +++ b/python/sglang/srt/managers/scheduler.py @@ -2750,7 +2750,6 @@ class Scheduler( def _maybe_namespace_elastic_radix_cache(self, req: Req) -> None: if ( get_exec().moe.elastic_ep_backend is None - or self.disable_radix_cache or not self.tree_cache.supports_prefix_sharing() ): return diff --git a/python/sglang/srt/managers/scheduler_components/invariant_checker.py b/python/sglang/srt/managers/scheduler_components/invariant_checker.py index d2ae577144..e1fd4233c6 100644 --- a/python/sglang/srt/managers/scheduler_components/invariant_checker.py +++ b/python/sglang/srt/managers/scheduler_components/invariant_checker.py @@ -484,9 +484,8 @@ class SchedulerInvariantChecker: def _check_tree_cache(self): if not envs.SGLANG_ENABLE_TREE_CACHE_SANITY_CHECK.get(): return - if ( - self.tree_cache.supports_prefix_sharing() - and (self.is_hybrid_swa and self.tree_cache.supports_swa()) + if self.tree_cache.supports_prefix_sharing() and ( + (self.is_hybrid_swa and self.tree_cache.supports_swa()) or (self.is_hybrid_ssm and self.tree_cache.supports_mamba()) ): self.tree_cache.sanity_check() diff --git a/python/sglang/srt/managers/scheduler_components/pool_stats_observer.py b/python/sglang/srt/managers/scheduler_components/pool_stats_observer.py index 422c50aa5d..1c29d6e06b 100644 --- a/python/sglang/srt/managers/scheduler_components/pool_stats_observer.py +++ b/python/sglang/srt/managers/scheduler_components/pool_stats_observer.py @@ -33,10 +33,13 @@ def kv_private_tokens(kv: ReqKvInfo, page_size: int) -> int: def kv_private_swa_tokens(kv: ReqKvInfo, page_size: int) -> int: + # Window eviction frees only [swa_dead_lo, evicted); [protected, swa_dead_lo) + # stays held until the request finishes, even after the cursor jumps past it. allocated = ceil_align(kv.kv_allocated_len, page_size) - return allocated - max( - kv.cache_protected_len, kv.get_evicted_seqlen(ComponentType.SWA) + window_freed = max( + 0, kv.get_evicted_seqlen(ComponentType.SWA) - kv.swa_dead_lo(page_size) ) + return allocated - kv.cache_protected_len - window_freed def kv_mamba_slots(kv: ReqKvInfo) -> int: diff --git a/test/registered/scheduler/test_retract_decode.py b/test/registered/scheduler/test_retract_decode.py index 77e9edb0d1..ab9affddea 100644 --- a/test/registered/scheduler/test_retract_decode.py +++ b/test/registered/scheduler/test_retract_decode.py @@ -101,12 +101,5 @@ class TestRetractDecodeLongOutput(CustomTestCase): assert self.process.poll() is None, "Server crashed during test" -@unittest.skipIf(is_in_ci(), "Skipped in CI due to long runtime") -class TestRetractDecodeLongOutputRadixDisabled(TestRetractDecodeLongOutput): - """python -m unittest test_retract_decode.TestRetractDecodeLongOutputRadixDisabled""" - - other_args = ["--disable-radix-cache"] - - if __name__ == "__main__": unittest.main() diff --git a/test/registered/unit/managers/scheduler_components/test_invariant_checker.py b/test/registered/unit/managers/scheduler_components/test_invariant_checker.py index 4a60d4ee44..2930dfac91 100644 --- a/test/registered/unit/managers/scheduler_components/test_invariant_checker.py +++ b/test/registered/unit/managers/scheduler_components/test_invariant_checker.py @@ -14,12 +14,14 @@ import torch from sglang.srt.disaggregation.utils import DisaggregationMode from sglang.srt.environ import envs +from sglang.srt.managers.schedule_batch import ReqKvInfo from sglang.srt.managers.scheduler_components import invariant_checker from sglang.srt.managers.scheduler_components.invariant_checker import ( SchedulerInvariantChecker, ) from sglang.srt.managers.scheduler_components.pool_stats_observer import ( SchedulerPoolStatsObserver, + kv_private_swa_tokens, ) from sglang.srt.mem_cache.allocator.page_interleave import PageInterleavePoolAllocator from sglang.srt.mem_cache.base_prefix_cache import InsertParams, MatchPrefixParams @@ -40,13 +42,14 @@ class TestCheckTreeCacheGate(CustomTestCase): envs.SGLANG_ENABLE_TREE_CACHE_SANITY_CHECK.clear() yield - def _make_checker(self): + def _make_checker(self, prefix_sharing=True, hybrid_swa=True, hybrid_ssm=False): tree_cache = MagicMock() - tree_cache.supports_prefix_sharing.return_value = True - tree_cache.supports_swa.return_value = True + tree_cache.supports_prefix_sharing.return_value = prefix_sharing + tree_cache.supports_swa.return_value = hybrid_swa + tree_cache.supports_mamba.return_value = hybrid_ssm return SchedulerInvariantChecker( - is_hybrid_swa=True, - is_hybrid_ssm=False, + is_hybrid_swa=hybrid_swa, + is_hybrid_ssm=hybrid_ssm, disaggregation_mode=DisaggregationMode.NULL, page_size=1, full_tokens_per_layer=None, @@ -101,6 +104,32 @@ class TestCheckTreeCacheGate(CustomTestCase): checker.tree_cache.sanity_check.assert_called_once() + def test_skipped_without_prefix_sharing(self): + with envs.SGLANG_ENABLE_TREE_CACHE_SANITY_CHECK.override(True): + checker = self._make_checker( + prefix_sharing=False, hybrid_swa=False, hybrid_ssm=True + ) + + checker._check_tree_cache() + + checker.tree_cache.sanity_check.assert_not_called() + + +class TestPrivateSwaTokens(CustomTestCase): + def test_floor_shielded_prefix_stays_private(self): + # A prefill-aware SWA request with radix off: the window cursor jumps to + # the floor (278) without freeing it, then frees [278, 300). + kv = ReqKvInfo( + req_pool_idx=0, + kv_allocated_len=280, + swa_evict_floor=278, + component_evicted_seqlens={ComponentType.SWA: 278}, + ) + self.assertEqual(kv_private_swa_tokens(kv, page_size=1), 280) + kv.kv_allocated_len = 450 + kv.set_evicted_seqlen(ComponentType.SWA, 300) + self.assertEqual(kv_private_swa_tokens(kv, page_size=1), 428) + class TestShardedFullPoolInvariant(CustomTestCase): PAGE_SIZE = 16