fix: fallback token indices allocation with verify+rebuild

vllm/nvfp4_cutedsl.py

@@ -130,6 +130,16 @@ class CuTeDSLMoERunner:
         self._token_indices = torch.arange(
             self.max_num_tokens, dtype=torch.int32
         ).unsqueeze(1).expand(-1, self.top_k).contiguous().view(-1).to(self.device)
+        # Verify the tensor was transferred correctly (CuTeDSL JIT can corrupt GPU state)
+        _verify = self._token_indices[:8].cpu().tolist()
+        if _verify != [0, 0, 1, 1, 2, 2, 3, 3]:
+            # Fallback: allocate directly on GPU
+            self._token_indices = torch.zeros(
+                self.max_num_tokens * self.top_k, dtype=torch.int32, device=self.device
+            )
+            for i in range(self.max_num_tokens):
+                self._token_indices[i * self.top_k] = i
+                self._token_indices[i * self.top_k + 1] = i
         self._expert_id_range = torch.arange(
             self.num_experts, dtype=torch.int32
         ).to(self.device)