[Fix] use torch.cat instead of torch.concat to prevent entering the Autograd backends. (#4466)

python/sglang/srt/mem_cache/memory_pool.py

@@ -172,7 +172,7 @@ class TokenToKVPoolAllocator:
             return
 
         if self.is_not_in_free_group:
-            self.free_slots = torch.concat((self.free_slots, free_index))
+            self.free_slots = torch.cat((self.free_slots, free_index))
         else:
             self.free_group.append(free_index)
 
@@ -183,7 +183,7 @@ class TokenToKVPoolAllocator:
     def free_group_end(self):
         self.is_not_in_free_group = True
         if self.free_group:
-            self.free(torch.concat(self.free_group))
+            self.free(torch.cat(self.free_group))
 
     def clear(self):
         # The padded slot 0 is used for writing dummy outputs from padded tokens.
@@ -739,7 +739,7 @@ class HostKVCache(abc.ABC):
     @synchronized
     def free(self, indices: torch.Tensor) -> int:
         self.mem_state[indices] = MemoryStateInt.IDLE
-        self.free_slots = torch.concat([self.free_slots, indices])
+        self.free_slots = torch.cat([self.free_slots, indices])
         self.can_use_mem_size += len(indices)
         return len(indices)