[Bugfix] Fix structured output crash on CPU due to pin_memory=True

vllm/v1/structured_output/utils.py

@@ -106,28 +106,33 @@ def apply_grammar_bitmask(
     # since the bitmask is already aligned with the logits.
     skip_out_indices = len(out_indices) == logits.shape[0]
 
-    index_tensor = None
+    indices: torch.Tensor | list[int] | None = None
     if not skip_out_indices:
-        # xgrammar expects a python list of indices but it will actually work with
-        # a tensor. If we copy the tensor ourselves here we can do it in a non_blocking
-        # manner and there should be no cpu sync within xgrammar.
-        index_tensor = torch.tensor(
-            out_indices, dtype=torch.int32, device="cpu", pin_memory=True
-        )
-        index_tensor = index_tensor.to(logits.device, non_blocking=True)
+        if logits.device.type == "cpu":
+            # On CPU, pass indices as a plain list — pin_memory requires CUDA,
+            # and the xgrammar CPU kernel expects Sequence[int], not a tensor.
+            indices = out_indices
+        else:
+            index_tensor = torch.tensor(
+                out_indices,
+                dtype=torch.int32,
+                device="cpu",
+                pin_memory=True,
+            )
+            indices = index_tensor.to(logits.device, non_blocking=True)
 
     # Handle dtype conversion for CPU (older xgrammar CPU kernels require float32)
     # See: https://github.com/vllm-project/vllm/issues/31901
     if logits.device.type == "cpu" and logits.dtype != torch.float32:
         # Convert to float32, apply bitmask, then convert back
         logits_float32 = logits.to(torch.float32)
         xgr.apply_token_bitmask_inplace(
-            logits_float32, grammar_bitmask, indices=index_tensor
+            logits_float32, grammar_bitmask, indices=indices
         )
         # Copy the modified values back to the original tensor
         logits.copy_(logits_float32.to(logits.dtype))
     else:
-        xgr.apply_token_bitmask_inplace(logits, grammar_bitmask, indices=index_tensor)
+        xgr.apply_token_bitmask_inplace(logits, grammar_bitmask, indices=indices)
 
 
 class OutlinesVocabulary: