diff --git a/vllm/model_executor/guided_decoding/xgrammar_decoding.py b/vllm/model_executor/guided_decoding/xgrammar_decoding.py
index f10a8fb8e03cf..2d8594cb8aafa 100644
--- a/vllm/model_executor/guided_decoding/xgrammar_decoding.py
+++ b/vllm/model_executor/guided_decoding/xgrammar_decoding.py
@@ -298,8 +298,11 @@ class XGrammarLogitsProcessor:
         # token_bitmask is a CPU tensor for use with accept_token and
         # fill_next_token_bitmask so we move it to the device of scores
         device_type = scores.device.type
+        dtype = scores.dtype
         if device_type != "cuda":
-            scores = scores.to("cpu").unsqueeze(0)
+            # xgrammar on cpu only supports float32 scores
+            # see: https://github.com/mlc-ai/xgrammar/blob/c1b64920cad24f44f235778c1c00bb52d57da01a/python/xgrammar/kernels/apply_token_bitmask_inplace_cpu.py#L22
+            scores = scores.to("cpu").float().unsqueeze(0)
 
         # Note: In this method, if the tensors have different dimensions
         # on CPU device fails, but on GPU it runs without error. Hence the
@@ -307,7 +310,7 @@ class XGrammarLogitsProcessor:
         xgr.apply_token_bitmask_inplace(scores,
                                         self.token_bitmask.to(scores.device))
         if device_type != "cuda":
-            scores = scores.to(device_type).squeeze()
+            scores = scores.to(dtype).to(device_type).squeeze()
 
         return scores