diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py
index 03814ea3..a2e1193d 100644
--- a/vllm/nvfp4_cutedsl.py
+++ b/vllm/nvfp4_cutedsl.py
@@ -86,12 +86,6 @@ class CuTeDSLMoERunner:
             self.max_num_tokens, device=self.device, dtype=torch.int32
         ).unsqueeze(1).expand(-1, self.top_k).contiguous().view(-1)
         
-        # Force GPU sync and verify
-        torch.cuda.synchronize()
-        _check = self._token_indices[:8].cpu().tolist()
-        assert _check == [0, 0, 1, 1, 2, 2, 3, 3], \
-            f"Token indices corrupted: {_check}"
-        
         self._expert_id_range = torch.arange(self.num_experts, device=self.device)
         
         self._expert_offsets_buf = torch.zeros(