diff --git a/python/sglang/srt/mem_cache/swa_radix_cache.py b/python/sglang/srt/mem_cache/swa_radix_cache.py
index 4268130e9..01c8814db 100644
--- a/python/sglang/srt/mem_cache/swa_radix_cache.py
+++ b/python/sglang/srt/mem_cache/swa_radix_cache.py
@@ -519,7 +519,7 @@ class SWARadixCache(KVCacheEventMixin, BasePrefixCache):
             ]
 
             # `req.prefix_indices` will be used in `PrefillAdder::add_chunked_req` later
-            req.prefix_indices = kv_indices
+            req.prefix_indices = kv_indices.to(dtype=torch.int64, copy=True)
             return
 
         token_ids = req.get_fill_ids()
diff --git a/python/sglang/srt/mem_cache/unified_radix_cache.py b/python/sglang/srt/mem_cache/unified_radix_cache.py
index 375c17832..597de340d 100644
--- a/python/sglang/srt/mem_cache/unified_radix_cache.py
+++ b/python/sglang/srt/mem_cache/unified_radix_cache.py
@@ -749,7 +749,7 @@ class UnifiedRadixCache(BasePrefixCache):
             kv_indices = self.req_to_token_pool.req_to_token[
                 req.req_pool_idx, : len(token_ids)
             ]
-            req.prefix_indices = kv_indices
+            req.prefix_indices = kv_indices.to(dtype=torch.int64, copy=True)
             return
 
         kv_indices_orig = self.req_to_token_pool.req_to_token[