[LLM] Add model correctness test on ARC for llama and falcon (#9347)

* add correctness test on arc for llama model

* modify layer name

* add falcon ut

* refactor and add ut for falcon model

* modify lambda positions and update docs

* replace loading pre input with last decodelayer output

* switch lower bound to single model instead of using the common one

* make the code implementation simple

* fix gpu action allocation memory issue

python/llm/test/inference_gpu/test_optimize_model.py

@@ -14,11 +14,13 @@
 # limitations under the License.
 #
 
-import pytest
 import os
+import pytest
 
-from bigdl.llm.transformers import AutoModelForCausalLM, AutoModel
+import torch
 from transformers import LlamaTokenizer, AutoTokenizer
+from bigdl.llm.transformers import AutoModelForCausalLM, AutoModel
+
 
 device = os.environ['DEVICE']
 print(f'Running on {device}')
@@ -29,7 +31,7 @@ prompt = "Once upon a time, there existed a little girl who liked to have advent
 
 @pytest.mark.parametrize('Model, Tokenizer, model_path',[
     (AutoModelForCausalLM, AutoTokenizer, os.environ.get('MPT_7B_ORIGIN_PATH')),
-    (AutoModelForCausalLM, AutoTokenizer, os.environ.get('FALCON_7B_ORIGIN_PATH')),
+    (AutoModelForCausalLM, AutoTokenizer, os.environ.get('FALCON_7B_ORIGIN_PATH'))
     ])
 def test_optimize_model(Model, Tokenizer, model_path):
     tokenizer = Tokenizer.from_pretrained(model_path, trust_remote_code=True)
@@ -55,6 +57,107 @@ def test_optimize_model(Model, Tokenizer, model_path):
 
     assert any(diff) is False
 
+class Test_Optimize_Gpu_Model:
+    def setup(self):
+
+        self.layer_outputs = []
+        self.pre_layer_outputs = []
+
+    def run_optimize_gpu_model(self, Model, Tokenizer, model_path, self_attn, layer_norm, lower_bound):
+        def forward_hook(module, input, output, layer_name):
+            self.layer_outputs.append(output)
+
+        def pre_forward_hook(module, input, output, layer_name):
+            self.pre_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,
+                                      trust_remote_code=True)
+        model = model.to(device)
+
+        for layer_name, layer_module in model.named_modules():
+            if layer_name == layer_norm:
+                layer_module.register_forward_hook(
+                    lambda module, input, output, layer_name=layer_name: pre_forward_hook(module, input,
+                                                                                          output, layer_name))
+            if layer_name == self_attn:
+                layer_module.register_forward_hook(
+                    lambda module, input, output, layer_name=layer_name: forward_hook(module, input,
+                                                                                      output, layer_name))
+        logits_base_model = (model(input_ids)).logits
+        # 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_norm:
+                layer_module.register_forward_hook(
+                    lambda module, input, output, layer_name=layer_name: replace_forward_hook(module, input,
+                                                                                              output, layer_name))
+            if layer_name == self_attn:
+                layer_module.register_forward_hook(
+                    lambda module, input, output, layer_name=layer_name: forward_hook(module, input,
+                                                                                      output, layer_name))
+        logits_optimized_model = (opt_model(input_ids)).logits
+        # the list `layer_output` has only one element.
+        opt_layer_tensor = self.layer_outputs[0]
+        opt_model.to('cpu')
+
+        attn_output_diff = []
+        for i, (t1, t2) in enumerate(zip(layer_tensor, opt_layer_tensor)):
+            if t1 is not None and t2 is not None:
+                if isinstance(t1, torch.Tensor) and isinstance(t2, torch.Tensor):
+                    # 'attn_output' is of type torch.Tensor.
+                    attn_output_diff.append(t1 - t2)
+                else:
+                    # 'past_key_value'is of type tuple as default.
+                    for i, (t3, t4) in enumerate(zip(t1, t2)):
+                        attn_output_diff.append(t3 - t4)
+
+        max_diff_tensor = [torch.max(item).item() for item in attn_output_diff]
+        assert all(max_diff <= lower_bound for max_diff in max_diff_tensor)
+
+
+    def test_falcon_gpu_model(self):
+
+        Model = AutoModelForCausalLM
+        Tokenizer = AutoTokenizer
+        model_path = os.environ.get('FALCON_7B_ORIGIN_PATH')
+        # currently only compare the output of the last self-attention layer.
+        layer_norm = "transformer.h.31.input_layernorm"
+        self_attn = "transformer.h.31.self_attention"
+        lower_bound = 0
+
+        self.run_optimize_gpu_model(Model, Tokenizer, model_path, self_attn, layer_norm, lower_bound)
+
+
+    def test_llama_gpu_model(self):
+
+        Model = AutoModelForCausalLM
+        Tokenizer = AutoTokenizer
+        model_path = os.environ.get('LLAMA2_7B_ORIGIN_PATH')
+        # currently only compare the output of the last self-attention layer.
+        layer_norm = "model.layers.31.input_layernorm"
+        self_attn = "model.layers.31.self_attn"
+        lower_bound = 5e-2
+
+        self.run_optimize_gpu_model(Model, Tokenizer, model_path, self_attn, layer_norm, lower_bound)
+
 
 if __name__ == '__main__':
     pytest.main([__file__])

python/llm/test/run-llm-inference-tests-gpu.sh

@@ -5,7 +5,6 @@ export LLM_HOME=${ANALYTICS_ZOO_ROOT}/python/llm/src
 export LLM_INFERENCE_TEST_DIR=${ANALYTICS_ZOO_ROOT}/python/llm/test/inference_gpu
 
 export USE_XETLA=OFF
-export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
 export DEVICE='xpu'
 
 set -e