diff --git a/python/llm/src/bigdl/llm/transformers/low_bit_linear.py b/python/llm/src/bigdl/llm/transformers/low_bit_linear.py
index ba527a15..16c8d334 100644
--- a/python/llm/src/bigdl/llm/transformers/low_bit_linear.py
+++ b/python/llm/src/bigdl/llm/transformers/low_bit_linear.py
@@ -426,6 +426,7 @@ class LowBitLinear(nn.Linear):
             try:
                 import intel_extension_for_pytorch
                 import linear_q4_0
+                from bigdl.llm.utils.xmx_checker import use_xmx
             except ModuleNotFoundError:
                 invalidInputError(False,
                                   "Please `pip install bigdl_core_xe` first.")
@@ -440,7 +441,8 @@ class LowBitLinear(nn.Linear):
                 # current workaround to reduce first token latency of fp32 input
                 # sometimes fp16 cause nan and training instability
                 # disable the conversion when training
-                if self.conver_to_half and x_2d.shape[0] > 1 and x_2d.dtype == torch.float32:
+                if self.conver_to_half and x_2d.shape[0] > 1 and x_2d.dtype == torch.float32 and \
+                        not use_xmx(x_2d, self.weight.qtype):
                     x_2d = x_2d.half()
                     result = linear_q4_0.forward_new(x_2d, self.weight.data, self.weight.qtype,
                                                      input_seq_size)
diff --git a/python/llm/src/bigdl/llm/utils/xmx_checker.py b/python/llm/src/bigdl/llm/utils/xmx_checker.py
new file mode 100644
index 00000000..e26916eb
--- /dev/null
+++ b/python/llm/src/bigdl/llm/utils/xmx_checker.py
@@ -0,0 +1,51 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import torch
+import intel_extension_for_pytorch as ipex
+from bigdl.llm.ggml.quantize import ggml_tensor_qtype
+
+
+SYM_INT4 = ggml_tensor_qtype["sym_int4"]
+SYM_INT8 = ggml_tensor_qtype["sym_int8"]
+NF4 = ggml_tensor_qtype["nf4"]
+NF3 = ggml_tensor_qtype["nf3"]
+FP8 = ggml_tensor_qtype["fp8"]
+FP4 = ggml_tensor_qtype["fp4"]
+MOFQ4 = ggml_tensor_qtype["mixed_4bit"]
+
+
+class XMXChecker:
+    def __init__(self):
+        self.support_xmx = self.check_xmx()
+        self.supported_qtype = [SYM_INT4, SYM_INT8, FP8]
+
+    @staticmethod
+    def check_xmx():
+        name = torch.xpu.get_device_name(0)
+        # todo: not sure how to check xmx or how to get device name for now
+        return "Arc(TM)" in name or "GPU Max" in name or "GPU Flex" in name
+
+    def check(self, input_tensor: torch.Tensor, qtype: int):
+        return self.support_xmx and 1 < input_tensor.shape[0] <= 8 and \
+            qtype in self.supported_qtype
+
+
+xmx_checker = XMXChecker()
+
+
+def use_xmx(input_tensor: torch.Tensor, qtype: int):
+    return xmx_checker.check(input_tensor, qtype)