debug: clone + sync token indices before JIT

vllm/nvfp4_cutedsl.py

@@ -111,10 +111,13 @@ class CuTeDSLMoERunner:
         
         # Allocate token indices FIRST, before any CUDA JIT compilation
         # (CuTeDSL's cute.compile can corrupt GPU memory state during JIT)
+        # Use .clone() to ensure a fresh allocation and cuda.synchronize() to
+        # ensure the kernel completes before any JIT compilation starts
         if self._token_indices is None:
             self._token_indices = torch.arange(
                 self.max_num_tokens, device=self.device, dtype=torch.int32
-            ).unsqueeze(1).expand(-1, self.top_k).contiguous().view(-1)
+            ).unsqueeze(1).expand(-1, self.top_k).contiguous().view(-1).clone()
+            torch.cuda.synchronize()
         if self._expert_id_range is None:
             self._expert_id_range = torch.arange(self.num_experts, device=self.device)
         if self._expert_offsets_buf is None: