From f359b967fe7b78a90d336cb093b46e7f790b1e3f Mon Sep 17 00:00:00 2001
From: Yishuo Wang <yishuo.wang@intel.com>
Date: Tue, 17 Dec 2024 10:31:32 +0800
Subject: [PATCH] update

---
 .../test_transformers_api_layernorm.py        | 38 +++++++++----------
 1 file changed, 19 insertions(+), 19 deletions(-)

diff --git a/python/llm/test/inference_gpu/test_transformers_api_layernorm.py b/python/llm/test/inference_gpu/test_transformers_api_layernorm.py
index 68a15d8a579..b0cd8a178f7 100644
--- a/python/llm/test/inference_gpu/test_transformers_api_layernorm.py
+++ b/python/llm/test/inference_gpu/test_transformers_api_layernorm.py
@@ -13,39 +13,39 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
- 
+
 import os
 import pytest
 import gc
- 
+
 import torch
 from ipex_llm.transformers import AutoModelForCausalLM, AutoModel
 from transformers import LlamaTokenizer, AutoTokenizer
- 
+
 device = os.environ['DEVICE']
 print(f'Running on {device}')
- 
+
 PROMPT = "Once upon a time, there existed a little girl who liked to have adventures. She wanted to go to places and meet new people, and have fun"
 TEST_MODEL_LIST = [
     ("Falcon-7B", AutoModelForCausalLM, AutoTokenizer, os.environ.get('FALCON_7B_ORIGIN_PATH'))
 ]
- 
+
 class Test_Optimize_Gpu_Model:
     def setup_method(self):
         self.layer_outputs = []
         self.pre_layer_outputs = []
- 
+
     def run_optimize_gpu_model(self, Name, Model, Tokenizer, model_path, LayerNorm_layer, layer_before_LayerNorm, lower_bound):
         with torch.inference_mode():
             def pre_forward_hook(module, input, output, layer_name):
                 self.pre_layer_outputs.append(output)
-                
+
             def forward_hook(module, input, output, layer_name):
                 self.layer_outputs.append(output)
- 
+
             tokenizer = Tokenizer.from_pretrained(model_path, trust_remote_code=True)
             input_ids = tokenizer.encode(PROMPT, return_tensors="pt").to(device)
- 
+
             model = Model.from_pretrained(model_path,
                                         load_in_4bit=True,
                                         optimize_model=False,
@@ -64,18 +64,18 @@ def forward_hook(module, input, output, layer_name):
             # the list `layer_output` has only one element.
             layer_tensor = self.layer_outputs.pop()
             model.to('cpu')
- 
+
             opt_model = Model.from_pretrained(model_path,
                                             load_in_4bit=True,
                                             optimize_model=True,
                                             trust_remote_code=True)
             opt_model = opt_model.to(device)
- 
- 
+
+
             def replace_forward_hook(module, input, output, layer_name):
                 output = self.pre_layer_outputs[0]
                 return output
- 
+
             for layer_name, layer_module in opt_model.named_modules():
                 if layer_name == layer_before_LayerNorm:
                     layer_module.register_forward_hook(
@@ -89,12 +89,12 @@ def replace_forward_hook(module, input, output, layer_name):
             # the list `layer_output` has only one element.
             opt_layer_tensor = self.layer_outputs[0]
             opt_model.to('cpu')
- 
- 
+
+
             LayerNorm_output_diff = []
             for i, (t1, t2) in enumerate(zip(layer_tensor, opt_layer_tensor)):
                 LayerNorm_output_diff.append(t1 - t2)
- 
+
             max_diff_tensor = [torch.max(item).item() for item in LayerNorm_output_diff]
             print(max_diff_tensor)
             torch.xpu.empty_cache()
@@ -102,16 +102,16 @@ def replace_forward_hook(module, input, output, layer_name):
             del opt_model
             gc.collect()
             assert all(max_diff <= lower_bound for max_diff in max_diff_tensor)
-   
+
     @pytest.mark.parametrize('Name, Model, Tokenizer, model_path',TEST_MODEL_LIST)
     def test_dynamic_functions(self, Name, Model, Tokenizer, model_path):
         if Name == "Falcon-7B":
             self.Falcon_7B_gpu_model(Name, Model, Tokenizer, model_path)
 
-    
+
     def Falcon_7B_gpu_model(self, Name, Model, Tokenizer, model_path):
         # currently only compare the output of the last LayerNorm layer.
         layer_before_LayerNorm = "transformer.h.30"
         LayerNorm_layer = "transformer.h.31.input_layernorm"
-        lower_bound = 0
+        lower_bound = 1e-5
         self.run_optimize_gpu_model(Name, Model, Tokenizer, model_path, LayerNorm_layer, layer_before_LayerNorm, lower_bound)
\ No newline at end of file