[Fix] Respect cache_protected_len in ChunkCache and disabled-radix release paths (#31662)

test/registered/unit/mem_cache/test_pure_swa_chunk_cache.py

@@ -23,6 +23,7 @@ class _FakeAllocator:
 class _FakeReq:
     req_pool_idx = 0
     swa_evict_floor = 3
+    cache_protected_len = 0
     kv = SimpleNamespace(swa_evicted_seqlen=6)
 
     def pop_committed_kv_cache(self):
@@ -30,12 +31,16 @@ class _FakeReq:
 
 
 class TestPureSWAChunkCache(CustomTestCase):
-    def test_finished_req_skips_already_evicted_swa_range(self):
+    def _make_cache(self):
         cache = PureSWAChunkCache.__new__(PureSWAChunkCache)
         cache.req_to_token_pool = SimpleNamespace(
             req_to_token=torch.arange(10, dtype=torch.int64).unsqueeze(0)
         )
         cache.token_to_kv_pool_allocator = _FakeAllocator()
+        return cache
+
+    def test_finished_req_skips_already_evicted_swa_range(self):
+        cache = self._make_cache()
 
         cache.cache_finished_req(_FakeReq(), kv_len_to_handle=8)
 
@@ -43,6 +48,16 @@ class TestPureSWAChunkCache(CustomTestCase):
         freed = cache.token_to_kv_pool_allocator.freed[0]
         self.assertTrue(torch.equal(freed, torch.tensor([0, 1, 2, 6, 7])))
 
+    def test_finished_req_skips_protected_prefix(self):
+        cache = self._make_cache()
+        req = _FakeReq()
+        req.cache_protected_len = 2
+
+        cache.cache_finished_req(req, kv_len_to_handle=8)
+
+        freed = cache.token_to_kv_pool_allocator.freed[0]
+        self.assertTrue(torch.equal(freed, torch.tensor([2, 6, 7])))
+
 
 if __name__ == "__main__":
     unittest.main()