Fix CPP Radix Cache and add test to CI (#11645)

python/sglang/srt/mem_cache/radix_cache_cpp.py

@@ -167,7 +167,7 @@ class RadixCacheCpp(BasePrefixCache):
         token_ids = (req.origin_input_ids + req.output_ids)[:kv_committed_len]
         kv_indices = self.req_to_token_pool.req_to_token[
             req.req_pool_idx, :kv_committed_len
-        ]
+        ].to(dtype=torch.int64, copy=True)
 
         # NOTE: our C++ implementation don't need `token_ids` and `kv_indices` to be page-aligned
         # it will automatically align them, but length of them should be equal
@@ -204,7 +204,9 @@ class RadixCacheCpp(BasePrefixCache):
         assert req.req_pool_idx is not None
         token_ids = req.fill_ids
         prefill_len = len(token_ids)  # prefill only (maybe chunked)
-        kv_indices = self.req_to_token_pool.req_to_token[req.req_pool_idx, :prefill_len]
+        kv_indices = self.req_to_token_pool.req_to_token[
+            req.req_pool_idx, :prefill_len
+        ].to(dtype=torch.int64, copy=True)
 
         # NOTE: our C++ implementation don't need `token_ids` and `kv_indices` to be page-aligned
         # it will automatically align them, but length of them should be equal

test/srt/run_suite.py

@@ -217,6 +217,7 @@ suites = {
         TestFile("openai_server/features/test_lora_openai_compatible.py", 150),
         TestFile("batch_invariant/test_batch_invariant_ops.py", 10),
         TestFile("test_deepseek_v3_deterministic.py", 240),
+        TestFile("test_cpp_radix_cache.py", 60),
     ],
     "nightly-4-gpu-b200": [
         TestFile("test_fp4_moe.py", 300),