diff --git a/README.md b/README.md
index 6334282..b8ea758 100644
--- a/README.md
+++ b/README.md
@@ -182,7 +182,7 @@ Using the fine-tuned model to generate new text can be done using the `generate_
 The arguments for the `generate_text.py` script are as follows:
 
 ```
-usage: generate_text.py [-h] --model_config MODEL_CONFIG [--benchmark_mode] [--benchmark_seq_length BENCHMARK_SEQ_LENGTH]
+usage: generate_text.py [-h] --model_config MODEL_CONFIG [--benchmark_mode] [--benchmark_seq_length BENCHMARK_SEQ_LENGTH] [--device cpu]
 
 optional arguments:
   -h, --help            show this help message and exit
@@ -191,6 +191,8 @@ optional arguments:
   --benchmark_mode      use intel pytorch extension to optimize model.
   --benchmark_seq_length BENCHMARK_SEQ_LENGTH
                         length of generation if benchmark mode is used.
+  --device cpu
+           Choose run on cpu or xpu, default is cpu.
 ```
 
 **Configuration Parameters**
@@ -242,14 +244,17 @@ Within the yaml configuration file, the following optional arguments can be spec
 
 **YAML file**                                 | **Environment Name** |  **Configuration** |
 | :---: | :---: | :---: |
-`env/intel/text-intel-torch.yml`             | `text-intel-torch` | Python=3.9.7, PyTorch v1.13, Intel® Extension for PyTorch v1.13, Intel® Neural Compressor v2.0.0 |
+`env/intel/text-intel-torch-cpu.yml`             | `text-intel-torch` | Python=3.9.7, PyTorch v1.13, Intel® Extension for PyTorch v1.13, Intel® Neural Compressor v2.0.0 |
+`env/intel/text-intel-torch-xpu.yml`             | `text-intel-torch` | Python=3.9.7, PyTorch v1.13, Intel® Extension for PyTorch v1.13+xpu, Intel® Neural Compressor v2.0.0 |
 
 ### Optimized Solution Setup
 
-Follow the below conda installation commands to setup the Intel® oneAPI optimized PyTorch environment for model training and text generation. 
+Follow the below conda installation commands to setup the Intel® oneAPI optimized PyTorch environment for model training and text generation. Please choose `env/intel/text-intel-torch-xpu.yml` if you have Intel GPU available.
 
 ```sh
-conda env create -f env/intel/text-intel-torch.yml
+conda env create -f env/intel/text-intel-torch-cpu.yml # For CPU
+or
+conda env create -f env/intel/text-intel-torch-xpu.yml # For XPU
 ```
 
 *Activate stock conda environment*
@@ -257,12 +262,17 @@ Use the following command to activate the environment that was created:
 ```sh
 conda activate text-intel-torch
 ```
+Please perform this additional installation step only if you are using an Intel GPU, CPU users can skip this step:
+```sh
+pip install -r env/intel/update_to_torch_xpu.txt
+```
 
-This script utilizes the dependencies found in the `env/intel/text-intel-torch.yml` file to create an environment as follows:
+This script utilizes the dependencies found in the `env/intel/text-intel-torch-cpu.yml` or `env/intel/text-intel-torch-xpu.yml` file to create an environment as follows:
 
 |     **YAML file**     | **Environment Name** |        **Configuration**        |
 | :-------------------: | :------------------: | :-----------------------------: |
-| `env/intel/text-intel-torch.yml` |     `text-intel-torch`      | Python=3.9.x with PyTorch v1.13, Intel® Extension For PyTorch v 1.13.0, Intel® Neural Compressor v2.0.0 |
+| `env/intel/text-intel-torch-cpu.yml` |     `text-intel-torch`      | Python=3.9.x with PyTorch v1.13, Intel® Extension For PyTorch v 1.13.0, Intel® Neural Compressor v2.0.0 |
+| `env/intel/text-intel-torch-xpu.yml, update_to_torch_xpu.txt` |     `text-intel-torch`      | Python=3.9.x with PyTorch v1.13, Intel® Extension For PyTorch v 1.13.0+xpu, Intel® Neural Compressor v2.0.0 |
 
 ### Intel® oneAPI Optimized Implementation
 
@@ -282,6 +292,33 @@ The command to fine-tune the model with Intel® optimizations enabled is:
 ipexrun --use_logical_core --enable_tcmalloc src/finetune_model.py --model_config configs/config_base.yml --data_path data/abcnews-date-text.csv --save_path saved_models/gpt2-medium-finetuned-intel --intel
 ```
 
+**For Intel GPU training**, device 'xpu' must be added to 'python/site-packages/transformers/training_args.py' as shown in the below diff at lines.
+It can be changed manually follows:
+```python
+diff --git a/training_args.py b/training_args.py
+index 1a90710..9421ccd 100644
+--- a/training_args.py
++++ b/training_args.py
+@@ -1466,6 +1466,8 @@ class TrainingArguments:
+                 torch.distributed.init_process_group(
+                     backend=self.xpu_backend, rank=rank, world_size=size, timeout=self.ddp_timeout_delta
+                 )
++        elif torch.xpu.is_available():
++            device = torch.device("xpu")
+         elif is_torch_tpu_available():
+             device = xm.xla_device()
+             self._n_gpu = 0
+```
+Or  to apply the patch automatically, please run the following Python script only once.
+```bash
+$ cd src
+$ python ./apply_xpu_patch.py
+```
+and then run the fine tuning step as follows:
+```sh
+ipexrun --use_logical_core --enable_tcmalloc src/finetune_model.py --model_config configs/config_base.yml --data_path data/abcnews-date-text.csv --save_path saved_models/gpt2-medium-finetuned-intel --intel
+```
+
 **Expected Output**<br>
 The output-trained model will be saved as a `huggingface` pre-trained model saved to the 'saved_models/gpt2-medium-finetuned-intel'. Training time in seconds would be generated at the end of the training module.
 
@@ -308,6 +345,11 @@ As above, this trained model can be used to generate text using the provided `ge
 python src/generate_text.py --model_config configs/config_finetuned_intel.yml
 ```
 
+To run inference on an Intel GPU, please run the following
+```sh
+python src/generate_text.py --model_config configs/config_finetuned_inc.yml --device xpu
+```
+
 **Expected Output**<br>
 The converted ONNX fine-tuned model will be saved to `saved_models/gpt2-medium-finetuned-intel-onnx`.
 
@@ -348,6 +390,11 @@ Once the quantized model is created, we can use the `generate_text.py` script on
 python src/generate_text.py --model_config configs/config_finetuned_inc.yml
 ```
 
+To run inference on an Intel GPU, please run the following
+```sh
+python src/generate_text.py --model_config configs/config_finetuned_inc.yml --device xpu
+```
+
 ## Performance Observations
 
 In the following, we report some results comparing Intel® technologies vs the stock alternative on the task of generating new texts of various lengths.  
diff --git a/env/intel/text-intel-torch.yml b/env/intel/text-intel-torch-cpu.yml
similarity index 100%
rename from env/intel/text-intel-torch.yml
rename to env/intel/text-intel-torch-cpu.yml
diff --git a/env/intel/text-intel-torch-xpu.yml b/env/intel/text-intel-torch-xpu.yml
new file mode 100644
index 0000000..ce5ad9e
--- /dev/null
+++ b/env/intel/text-intel-torch-xpu.yml
@@ -0,0 +1,15 @@
+# Refer to https://intel.github.io/intel-extension-for-pytorch/xpu/1.13.10+xpu/tutorials/getting_started.html
+# for intel_extension_for_pytorch installation.
+name: text-intel-torch
+channels:
+  - conda-forge
+dependencies:
+  - python=3.9.7
+  - pip=21.2.4
+  - pip:
+    - neural-compressor==2.0
+    - pandas==1.5.3
+    - numpy==1.23.5
+    - transformers==4.26.1
+    - datasets==2.10.1
+    - optimum[onnxruntime]==1.6.4
diff --git a/env/intel/update_to_torch_xpu.txt b/env/intel/update_to_torch_xpu.txt
new file mode 100644
index 0000000..459ee5b
--- /dev/null
+++ b/env/intel/update_to_torch_xpu.txt
@@ -0,0 +1,3 @@
+-i https://developer.intel.com/ipex-whl-stable-xpu      
+torch==1.13.0a0      
+intel_extension_for_pytorch==1.13.10+xpu
diff --git a/src/apply_xpu_patch.py b/src/apply_xpu_patch.py
new file mode 100644
index 0000000..4cd0478
--- /dev/null
+++ b/src/apply_xpu_patch.py
@@ -0,0 +1,12 @@
+import transformers
+import os
+import subprocess
+
+module_path = transformers.__path__[0]
+# get path for training_args.py
+target_file_path = os.path.join(module_path, "training_args.py")
+
+# apply patch to training_args.py file in the transformers package
+subprocess.run(["patch", target_file_path, "transformers_xpu.patch"])
+
+print("patch applied successfully")
diff --git a/src/generate_text.py b/src/generate_text.py
index c0bd13e..64c3ae3 100644
--- a/src/generate_text.py
+++ b/src/generate_text.py
@@ -26,7 +26,7 @@
     AutoModelForCausalLM,
     set_seed
 )
-
+import intel_extension_for_pytorch as ipex
 
 def generate_text(
     tokenizer: PreTrainedTokenizer,
@@ -36,7 +36,8 @@ def generate_text(
     min_length: int = 0,
     max_length: int = 10,
     algorithm: str = 'greedy',
-    stop_token: str = '.') -> Tuple[List[int],float]:
+    stop_token: str = '.',
+    device: str = 'cpu') -> Tuple[List[int],float]:
     """Generate text using the provided model and algorithm.
 
     Args:
@@ -60,18 +61,26 @@ def generate_text(
         Tuple(List[int], float): generated sequence of token_ids, total time for
             of model inference calls
     """
-    
 
     all_token_ids = input_ids.clone()
     all_attention_masks = attention_mask.clone()
     eos_token_id = tokenizer([stop_token], return_tensors='np')[
         'input_ids'][0][0]
     has_eos = torch.zeros(1, dtype=torch.bool)
+    ones_mask = torch.ones([1, 1])
+
+    if device == 'xpu':
+        print("#### xpu")
+        all_token_ids = input_ids.to("xpu")
+        all_attention_masks = attention_mask.to("xpu")
+        eos_token_id = torch.tensor([eos_token_id], device=torch.device("xpu"))
+        has_eos = has_eos.to("xpu")
+        ones_mask = ones_mask.to("xpu")
+        model = model.to("xpu")
 
     total_time = 0
     for step in range(max_length):
-
-        if isinstance(model, torch.nn.Module):
+        if isinstance(model, torch.nn.Module) and device == 'cpu':
             start = time.time()
             next_token_logits = torch.nn.functional.softmax(
                 model(
@@ -79,6 +88,14 @@ def generate_text(
                     attention_mask=all_attention_masks)[:, -1, :], dim=1)
             end = time.time()
             total_time += end - start
+        elif isinstance(model, torch.nn.Module) and device == 'xpu':
+            start = time.time()
+            next_token_logits = torch.nn.functional.softmax(
+                model(
+                    input_ids=all_token_ids,
+                    attention_mask=all_attention_masks)[0][:, -1, :], dim=1)
+            end = time.time()
+            total_time += end - start
         elif isinstance(model, ort.InferenceSession):
             ort_input = {
                 "input_ids": np.array(all_token_ids),
@@ -101,7 +118,7 @@ def generate_text(
         all_token_ids = torch.cat(
             [all_token_ids, tokens_to_add.unsqueeze(-1)], dim=-1)
         all_attention_masks = torch.cat(
-            [all_attention_masks, torch.ones([1, 1])], dim=-1).type_as(all_attention_masks)
+            [all_attention_masks, ones_mask], dim=-1).type_as(all_attention_masks)
         if step > min_length and next_tokens == eos_token_id:
             break
 
@@ -115,8 +132,8 @@ def generate(
         max_length: int = 10,
         prompt_file: str = None,
         benchmark_mode: bool = False,
-        n_runs: int = 100):
-
+        n_runs: int = 100,
+        device: str = 'cpu'):
     # read prompts from file into a list for batch processing
     tokenized_input = []
     if not benchmark_mode and prompt_file is not None:
@@ -143,7 +160,8 @@ def generate(
                     tokenized_prompt.attention_mask,
                     min_length=max_length,
                     max_length=max_length,
-                    algorithm='greedy')
+                    algorithm='greedy',
+                    device=device)
                 if i > 10:
                     times.append(total_time)
         print(f"Average Generation time: {np.mean(times)}s")
@@ -160,7 +178,8 @@ def generate(
                 tokenized_prompt.attention_mask,
                 min_length=min_length,
                 max_length=max_length,
-                algorithm='sample')
+                algorithm='sample',
+                device=device)
             tokenized_output.append(res)
 
         out_json = []
@@ -216,7 +235,8 @@ def main(flags):
         min_length=min_length,
         max_length=max_length,
         prompt_file=prompt_file,
-        benchmark_mode=flags.benchmark_mode
+        benchmark_mode=flags.benchmark_mode,
+        device=flags.device
     )
 
 
@@ -241,6 +261,11 @@ def main(flags):
                         type=int,
                         default=10
                         )
+    parser.add_argument('--device',
+                        type=str,
+                        required=False,
+                        default='cpu',
+                        help='Choose run on cpu or xpu, default is cpu.')
 
     FLAGS = parser.parse_args()
     main(FLAGS)
diff --git a/src/transformers_xpu.patch b/src/transformers_xpu.patch
new file mode 100644
index 0000000..57d5a30
--- /dev/null
+++ b/src/transformers_xpu.patch
@@ -0,0 +1,20 @@
+--- a.py        2023-04-04 11:50:27.000000000 +0000
++++ b.py        2023-04-04 11:52:00.000000000 +0000
+@@ -53,6 +53,7 @@
+     requires_backends,
+ )
+
++import intel_extension_for_pytorch
+
+ if is_torch_available():
+     import torch
+@@ -1466,6 +1467,8 @@
+                 torch.distributed.init_process_group(
+                     backend=self.xpu_backend, rank=rank, world_size=size, timeout=self.ddp_timeout_delta
+                 )
++        elif torch.xpu.is_available():
++            device = torch.device("xpu")
+         elif is_torch_tpu_available():
+             device = xm.xla_device()
+             self._n_gpu = 0
+~