[mem_cache] Make release, row-reuse asserts, and presence checks read the KV record (#37167)

test/registered/unit/layers/test_minicpm_sparse_cache.py

@@ -68,7 +68,7 @@ def make_pool_and_req(capacity: int = 64):
     )
     req = SimpleNamespace(
         inflight_middle_chunks=0,
-        kv=SimpleNamespace(req_pool_idx=None),
+        kv=ReqKvInfo(),
     )
     req_pool_idx = pool.alloc([req])[0]
     return pool, req, req_pool_idx, allocator

test/registered/unit/mem_cache/test_decode_retraction_backup.py

@@ -3,6 +3,7 @@ from types import SimpleNamespace
 
 import torch
 
+from sglang.srt.managers.schedule_batch import ReqKvInfo
 from sglang.srt.mem_cache.allocator import TokenToKVPoolAllocator
 from sglang.srt.mem_cache.cache_init_params import CacheInitParams
 from sglang.srt.mem_cache.common import retraction_backup
@@ -129,9 +130,7 @@ class TestDecodeRetractionBackup(unittest.TestCase):
         )
 
     def _admit_req(self, env, num_tokens: int):
-        req = SimpleNamespace(
-            rid="request", kv=SimpleNamespace(req_pool_idx=None), seqlen=num_tokens + 1
-        )
+        req = SimpleNamespace(rid="request", kv=ReqKvInfo(), seqlen=num_tokens + 1)
         self.assertIsNotNone(env.req_to_token_pool.alloc([req]))
         source_indices = env.allocator.alloc(num_tokens)
         self.assertIsNotNone(source_indices)

test/registered/unit/mem_cache/test_dllm_fdfo_kv_reuse.py

@@ -7,6 +7,7 @@ from types import SimpleNamespace
 import torch
 
 from sglang.srt.dllm.mixin.scheduler import DllmManager
+from sglang.srt.managers.schedule_batch import ReqKvInfo
 from sglang.srt.mem_cache.allocation import alloc_for_extend
 from sglang.srt.mem_cache.memory_pool import ReqToTokenPool
 from sglang.srt.runtime_context import get_context
@@ -63,7 +64,7 @@ def _make_req(rid, prefix, block_size, *, req_pool_idx=None, reuse=False):
         prefix_indices=torch.tensor(prefix, dtype=torch.int32),
         dllm_incomplete_ids=array("q", range(block_size)) if reuse else array("q"),
         inflight_middle_chunks=1 if req_pool_idx is not None else 0,
-        kv=SimpleNamespace(
+        kv=ReqKvInfo(
             req_pool_idx=req_pool_idx,
             kv_committed_len=len(prefix) if req_pool_idx is not None else 0,
             kv_allocated_len=(