[mem_cache] Move cache_protected_len and swa_evict_floor into ReqKvInfo (#36982)

test/manual/chunked_prefill/test_scripted_radix.py

@@ -304,7 +304,7 @@ class TestRadixNoTailChunked(ScriptedTestCase):
             if req is not None and req.rid == r.rid:
                 observed_mid_chunk = True
                 prefix_len: int = len(req.prefix_indices)
-                protected_len: int = req.cache_protected_len
+                protected_len: int = req.kv.cache_protected_len
                 assert prefix_len == protected_len, (
                     f"page_size=1 must take the no-tail else branch: "
                     f"len(prefix_indices)={prefix_len} != "
@@ -460,7 +460,7 @@ class TestRadixPartialPage(ScriptedTestCase):
             req = s.chunked_req
             if req is not None and req.rid == r.rid:
                 prefix_len: int = len(req.prefix_indices)
-                protected_len: int = req.cache_protected_len
+                protected_len: int = req.kv.cache_protected_len
                 assert prefix_len >= protected_len, (
                     f"len(prefix_indices)={prefix_len} dropped below "
                     f"cache_protected_len={protected_len}: tail was freed "

test/registered/unit/layers/test_minicpm_sparse_cache.py

@@ -226,7 +226,7 @@ def test_streaming_session_release_frees_compressed_slots():
     )
     session.slots["session-a"] = SessionSlot(
         req_pool_idx=req_pool_idx,
-        kv=SimpleNamespace(kv_allocated_len=16),
+        kv=SimpleNamespace(kv_allocated_len=16, cache_protected_len=0),
     )
 
     session.release_session("session-a")

test/registered/unit/managers/test_priority_scheduling_disaggregation.py

@@ -97,7 +97,7 @@ class TestDecodePreallocQueuePriority(unittest.TestCase):
             finished_reason=FINISH_ABORT("failed") if failed else None,
             return_logprob=False,
             sampling_params=SimpleNamespace(max_new_tokens=8),
-            cache_protected_len=0,
+            kv=SimpleNamespace(cache_protected_len=0),
             time_stats=MagicMock(),
         )
         return SimpleNamespace(

test/registered/unit/managers/test_scheduler_chunked_req_gate.py

@@ -12,7 +12,7 @@ from sglang.test.test_utils import CustomTestCase, maybe_stub_sgl_kernel
 
 maybe_stub_sgl_kernel()
 
-from sglang.srt.managers.schedule_batch import NextBatchPlan, Req
+from sglang.srt.managers.schedule_batch import NextBatchPlan, Req, ReqKvInfo
 from sglang.srt.managers.scheduler import Scheduler
 from sglang.srt.mem_cache.chunk_cache import ChunkCache
 from sglang.srt.utils.common import Range
@@ -38,7 +38,7 @@ def _make_req(
     req.extend_range = Range(fill_len - extend_input_len, fill_len)
     req.inflight_middle_chunks = 0
     req.host_hit_length = 0
-    req.cache_protected_len = 0
+    req.kv = ReqKvInfo()
     req.skip_radix_cache_insert = False
     req.last_node = None
     req.swa_uuid_for_lock = None

test/registered/unit/mem_cache/test_decode_radix_lock_ref.py

@@ -76,14 +76,15 @@ class MockReq:
         )
         self.output_ids = array("q", [fill_ids[-1]] if len(fill_ids) > 1 else [])
         self.req_pool_idx = req_pool_idx
-        self.cache_protected_len = cache_protected_len
         self.last_node = last_node
         self.extra_key = None
         self.cache_salt = None
         self.prefix_indices = torch.empty(0, dtype=torch.int64)
         self.priority = 0
         self.kv_committed_len = len(fill_ids)
-        self.kv = SimpleNamespace(kv_allocated_len=len(fill_ids))
+        self.kv = SimpleNamespace(
+            kv_allocated_len=len(fill_ids), cache_protected_len=cache_protected_len
+        )
 
     def get_fill_ids(self):
         return self.full_untruncated_fill_ids[: self.extend_range.end]
@@ -368,7 +369,7 @@ class TestDecodeLockRefScenarios(unittest.TestCase):
         req.output_ids = [99]
         req.last_node = object()
         req.finished_reason = None
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = 123
         req.swa_prefix_lock_released = False
         req.pd_rebootstrap_in_progress = False

test/registered/unit/mem_cache/test_pure_swa_chunk_cache.py

@@ -22,9 +22,11 @@ class _FakeAllocator:
 
 class _FakeReq:
     req_pool_idx = 0
-    swa_evict_floor = 3
-    cache_protected_len = 0
-    kv = SimpleNamespace(swa_evicted_seqlen=6)
+
+    def __init__(self):
+        self.kv = SimpleNamespace(
+            swa_evicted_seqlen=6, swa_evict_floor=3, cache_protected_len=0
+        )
 
     def pop_committed_kv_cache(self):
         return 8
@@ -51,7 +53,7 @@ class TestPureSWAChunkCache(CustomTestCase):
     def test_finished_req_skips_protected_prefix(self):
         cache = self._make_cache()
         req = _FakeReq()
-        req.cache_protected_len = 2
+        req.kv.cache_protected_len = 2
 
         cache.cache_finished_req(req, kv_len_to_handle=8)
 

test/registered/unit/mem_cache/test_radix_cache_unit.py

@@ -27,6 +27,7 @@ import random
 import unittest
 import unittest.mock
 from array import array
+from types import SimpleNamespace
 
 import torch
 
@@ -518,7 +519,7 @@ class TestRadixCache(unittest.TestCase):
         cache.req_to_token_pool = ReqToTokenPool(request_indices.clone())
         req = unittest.mock.Mock(
             req_pool_idx=0,
-            cache_protected_len=0,
+            kv=SimpleNamespace(cache_protected_len=0),
             extra_key=None,
             cache_salt=None,
             priority=0,

test/registered/unit/mem_cache/test_radix_force_miss.py

@@ -12,6 +12,7 @@ register_cpu_ci(est_time=5, suite="base-a-test-cpu")
 import unittest
 import unittest.mock
 from array import array
+from types import SimpleNamespace
 
 import torch
 
@@ -39,7 +40,7 @@ class _StubReq:
         self.host_hit_length = None
         self.num_matched_prefix_tokens = 0
         self.mamba_branching_seqlen = None
-        self.cache_protected_len = None
+        self.kv = SimpleNamespace(cache_protected_len=None)
 
     def _compute_max_prefix_len(self, input_len):
         return max(input_len - 1, 0)

test/registered/unit/mem_cache/test_streaming_session_unit.py

@@ -72,13 +72,13 @@ class _FakeReq:
         self.kv = SimpleNamespace(
             kv_allocated_len=allocated,
             swa_evicted_seqlen=0,
+            cache_protected_len=0,
         )
         self.origin_input_ids = list(range(committed))
         self.output_ids = []
         self.extra_key = None
         self.cache_salt = None
         self.last_node = None
-        self.cache_protected_len = 0
         self.swa_uuid_for_lock = None
         self.skip_lock_node_ids = {}
         self.mamba_pool_idx = None
@@ -114,8 +114,9 @@ def test_preabort_detaches_session_and_preserves_slot():
     tree_cache.slots["session-a"] = SessionSlot(
         req_pool_idx=0,
         kv_committed_len=48,
-        kv=SimpleNamespace(kv_allocated_len=48, swa_evicted_seqlen=0),
-        cache_protected_len=16,
+        kv=SimpleNamespace(
+            kv_allocated_len=48, swa_evicted_seqlen=0, cache_protected_len=16
+        ),
     )
 
     req = _FakeReq("session-a", req_pool_idx=1, committed=1, allocated=1)
@@ -178,9 +179,10 @@ def test_nth_mid_abort_nukes_session_slot():
     tree_cache.slots["session-a"] = SessionSlot(
         req_pool_idx=0,
         kv_committed_len=50,
-        kv=SimpleNamespace(kv_allocated_len=50, swa_evicted_seqlen=0),
+        kv=SimpleNamespace(
+            kv_allocated_len=50, swa_evicted_seqlen=0, cache_protected_len=0
+        ),
         last_node=None,
-        cache_protected_len=0,
     )
 
     # Mid-processing abort: req has the SESSION slot's pool_idx (restore_to_req ran).
@@ -216,9 +218,10 @@ def test_release_session_threads_mamba_skip_ids():
     tree_cache.slots["session-a"] = SessionSlot(
         req_pool_idx=0,
         kv_committed_len=50,
-        kv=SimpleNamespace(kv_allocated_len=50, swa_evicted_seqlen=0),
+        kv=SimpleNamespace(
+            kv_allocated_len=50, swa_evicted_seqlen=0, cache_protected_len=0
+        ),
         last_node=lock_node,
-        cache_protected_len=0,
         skip_lock_node_ids={ComponentType.MAMBA: {42}},
     )
 

test/registered/unit/mem_cache/test_swa_eviction_boundary.py

@@ -18,7 +18,7 @@ from types import SimpleNamespace
 
 import torch
 
-from sglang.srt.managers.schedule_batch import ScheduleBatch
+from sglang.srt.managers.schedule_batch import ReqKvInfo, ScheduleBatch
 from sglang.srt.mem_cache.allocator.swa import SWATokenToKVPoolAllocator
 from sglang.srt.mem_cache.cache_init_params import CacheInitParams
 from sglang.srt.mem_cache.common import free_swa_out_of_window_slots
@@ -107,10 +107,7 @@ def _make_req(req_pool_idx, token_ids, cache_protected_len, tree):
         is_holding_kv=True,
         origin_input_ids=token_ids,
         output_ids=[],
-        cache_protected_len=cache_protected_len,
-        kv=SimpleNamespace(
-            swa_evicted_seqlen=0,
-        ),
+        kv=ReqKvInfo(cache_protected_len=cache_protected_len),
         extra_key=None,
         cache_salt=None,
         last_node=tree.root_node,

test/registered/unit/mem_cache/test_swa_unittest.py

@@ -36,7 +36,7 @@ class _DummyReq:
     def __init__(self):
         self._kv_committed_len = 0
         self.swa_prefix_lock_released = False
-        self.kv = SimpleNamespace(swa_evicted_seqlen=0)
+        self.kv = SimpleNamespace(swa_evicted_seqlen=0, cache_protected_len=0)
 
 
 def _build_swa_tree(
@@ -682,7 +682,7 @@ class TestSWA(unittest.TestCase):
         req.last_node = tree.root_node
         req.swa_uuid_for_lock = None
         req.kv.swa_evicted_seqlen = 0
-        req.cache_protected_len = 1
+        req.kv.cache_protected_len = 1
         # Intentionally mismatch to ensure code does not use len(prefix_indices).
         req.prefix_indices = torch.tensor([7, 8, 9, 10, 11], device=tree.device)
 
@@ -700,7 +700,7 @@ class TestSWA(unittest.TestCase):
             req, is_insert=True, kv_len_to_handle=req._kv_committed_len
         )
 
-        self.assertEqual(captured["prev_prefix_len"], req.cache_protected_len)
+        self.assertEqual(captured["prev_prefix_len"], req.kv.cache_protected_len)
         self.assertTrue(captured["is_bigram"])
         self.assertEqual(captured["key_len"], len(req.origin_input_ids) - 1)
 
@@ -720,7 +720,7 @@ class TestSWA(unittest.TestCase):
         req2.last_node = tree.root_node
         req2.swa_uuid_for_lock = None
         req2.kv.swa_evicted_seqlen = 0
-        req2.cache_protected_len = 1
+        req2.kv.cache_protected_len = 1
         req2.prefix_indices = torch.tensor([21, 22, 23, 24, 25], device=tree.device)
 
         freed_lens = []
@@ -917,7 +917,7 @@ class TestCacheUnfinishedReqEvictedPrefix(CustomTestCase):
         req.get_fill_ids = lambda: token_ids
         req.extra_key = None
         req.cache_salt = None
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.last_node = tree.root_node
         req.swa_uuid_for_lock = None
         req.prefix_indices = torch.empty(0, dtype=torch.int64, device=tree.device)
@@ -928,7 +928,7 @@ class TestCacheUnfinishedReqEvictedPrefix(CustomTestCase):
         # The insert itself frees nothing.
         self.assertEqual(allocator.swa_available_size(), swa_before)
         # The live leaf holds a full window, so the whole key stays matchable.
-        self.assertEqual(req.cache_protected_len, num_tokens)
+        self.assertEqual(req.kv.cache_protected_len, num_tokens)
         # [0, evicted) is a tombstone; only [evicted, num_tokens) counts as SWA.
         (first,) = tree.root_node.children.values()
         self.assertTrue(first.swa_tombstone)

test/registered/unit/mem_cache/test_unified_radix_cache_bench.py

@@ -642,7 +642,7 @@ def bench_cache_finished(
             len(req.prefix_indices), len(req.full_untruncated_fill_ids)
         )
         req.last_node = node
-        req.cache_protected_len = matched_len
+        req.kv.cache_protected_len = matched_len
         req.kv_committed_len = len(seq)
         if hasattr(lr, "swa_uuid_for_lock"):
             req.swa_uuid_for_lock = lr.swa_uuid_for_lock

test/registered/unit/mem_cache/test_unified_radix_cache_unittest.py

@@ -1228,7 +1228,7 @@ class UnifiedRadixCacheSuite:
         req_to_token_pool.write((req.req_pool_idx, slice(0, kv_len)), kv_indices)
         req.kv_committed_len = kv_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
         req.full_untruncated_fill_ids = array("q", input_ids + output_ids)
@@ -1269,7 +1269,7 @@ class UnifiedRadixCacheSuite:
         req.kv_committed_len = kv_len
         req.kv.kv_allocated_len = kv_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
         if self.cfg.has_mamba:
@@ -1313,7 +1313,7 @@ class UnifiedRadixCacheSuite:
         req_to_token_pool.write((req.req_pool_idx, slice(0, kv_len)), kv_indices)
         req.kv_committed_len = kv_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.swa_prefix_lock_released = True
         req.extra_key = None
@@ -1348,7 +1348,7 @@ class UnifiedRadixCacheSuite:
         req_to_token_pool.write((req.req_pool_idx, slice(0, kv_len)), kv_indices)
         req.kv_committed_len = kv_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
         if self.cfg.has_mamba:
@@ -1357,7 +1357,7 @@ class UnifiedRadixCacheSuite:
         cache.cache_unfinished_req(req)
 
         self.assertGreater(len(req.prefix_indices), 0)
-        self.assertEqual(req.cache_protected_len, len(req.prefix_indices))
+        self.assertEqual(req.kv.cache_protected_len, len(req.prefix_indices))
         self.assertIsNotNone(req.last_node)
         self.assertFalse(req.swa_prefix_lock_released)
 
@@ -1385,7 +1385,7 @@ class UnifiedRadixCacheSuite:
         req_to_token_pool.write((req.req_pool_idx, slice(0, len(tokens))), kv_indices)
         req.kv_committed_len = len(tokens)
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
         req.kv.swa_evicted_seqlen = evicted_len
@@ -1487,7 +1487,7 @@ class UnifiedRadixCacheSuite:
         req_to_token_pool.write((req.req_pool_idx, slice(0, kv_len)), kv_indices)
         req.kv_committed_len = kv_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
         req.full_untruncated_fill_ids = array("q", input_ids)
@@ -1612,7 +1612,7 @@ class UnifiedRadixCacheSuite:
         req_to_token_pool.write((req.req_pool_idx, slice(0, kv_len)), fresh_value)
         req.kv_committed_len = kv_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
         req.kv.swa_evicted_seqlen = 0
@@ -1639,7 +1639,7 @@ class UnifiedRadixCacheSuite:
                 swa_value,
             )
         )
-        self.assertEqual(req.cache_protected_len, len(tokens))
+        self.assertEqual(req.kv.cache_protected_len, len(tokens))
 
         cache.dec_lock_ref(
             req.last_node,
@@ -2199,7 +2199,7 @@ class UnifiedRadixCacheSuite:
         req_to_token_pool.write((req.req_pool_idx, slice(0, pre_len)), kv_indices)
         req.kv_committed_len = pre_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
 
@@ -2288,7 +2288,7 @@ class UnifiedRadixCacheSuite:
         req_to_token_pool.write((req.req_pool_idx, slice(0, pre_len)), kv_indices)
         req.kv_committed_len = pre_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
 
@@ -6770,7 +6770,7 @@ class TestUnifiedRadixCacheInt8MambaCheckpoint(CustomTestCase):
         req.output_ids = array("q")
         req.kv_committed_len = len(tokens)
         req.kv.kv_allocated_len = len(tokens)
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
         req.mamba_last_track_seqlen = len(tokens)
@@ -8097,7 +8097,7 @@ class TestSWAWindowUnderBigramKey(CustomTestCase):
         req_to_token_pool.write((req.req_pool_idx, slice(0, seq_len)), kv_indices)
         req.kv_committed_len = seq_len
         req.last_node = cache.root_node_handle()
-        req.cache_protected_len = 0
+        req.kv.cache_protected_len = 0
         req.swa_uuid_for_lock = None
         req.extra_key = None
 
@@ -8113,7 +8113,7 @@ class TestSWAWindowUnderBigramKey(CustomTestCase):
             f"{self.cfg.sliding_window_size} window",
         )
         self.assertEqual(
-            req.cache_protected_len,
+            req.kv.cache_protected_len,
             boundary,
             "the match after the insert must reach the leaf the insert created",
         )

test/registered/xpu/test_lmcache_radix_cache.py

@@ -72,7 +72,7 @@ def _make_req(rid, req_pool_idx, token_ids, tree):
         extra_key=None,
         cache_salt=None,
         last_node=tree.root_node,
-        cache_protected_len=0,
+        kv=SimpleNamespace(cache_protected_len=0),
         priority=0,
         kv_committed_freed=False,
         kv_committed_len=len(token_ids),