Upgrade transformers to 4.42.3 for ipex models (#812)

* upgrade ipex model to transformers 4.42.3

* fix tests

Author: jiqing-feng

.github/workflows/test_ipex.yml

@@ -21,7 +21,7 @@ jobs:
       fail-fast: false
       matrix:
         python-version: [3.8, 3.9]
-        transformers-version: [4.39.0, 4.41.2]
+        transformers-version: [4.39.0, 4.42.3]
         os: [ubuntu-latest]
 
     runs-on: ${{ matrix.os }}

optimum/exporters/ipex/model_patcher.py

@@ -34,7 +34,7 @@
 
 # Please also update in the setup.py and .github/workflows/test_ipex.yml if you change the transformers version
 _TRANSFORMERS_MIN_VERSION = "4.39.0"
-_TRANSFORMERS_MAX_VERSION = "4.41.2"
+_TRANSFORMERS_MAX_VERSION = "4.42.3"
 
 _IPEX_EXPORTED_GENERATION_TASKS = ("text-generation",)
 

optimum/intel/ipex/modeling_base.py

@@ -134,7 +134,6 @@ class IPEXModel(OptimizedModel):
     base_model_prefix = "ipex_model"
     main_input_name = "input_ids"
     output_name = "last_hidden_state"
-    _supports_cache_class = False
 
     def __init__(
         self,
@@ -428,6 +427,8 @@ def forward(
 class IPEXModelForCausalLM(IPEXModel, GenerationMixin):
     auto_model_class = AutoModelForCausalLM
     export_feature = "text-generation"
+    _supports_cache_class = False
+    _is_stateful = False
 
     def __init__(
         self,
@@ -476,8 +477,8 @@ def __init__(
             else:
                 self._reorder_cache = self.model_cls._reorder_cache.__get__(self)
 
-        if is_transformers_version(">=", "4.38.0") and model_type in {"llama", "phi", "persimmon"}:
-            self.prepare_inputs_for_generation = _prepare_inputs_for_generation_for_llama
+        if is_transformers_version(">=", "4.38.0") and model_type in {"llama", "phi", "persimmon", "mistral"}:
+            self.prepare_inputs_for_generation = _ipex_prepare_inputs_for_generation
         else:
             self.prepare_inputs_for_generation = self.model_cls.prepare_inputs_for_generation.__get__(self)
 
@@ -613,7 +614,7 @@ def generate(self, *args, **kwargs):
         return super().generate(*args, **kwargs)
 
 
-def _prepare_inputs_for_generation_for_llama(
+def _ipex_prepare_inputs_for_generation(
     input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
 ):
     from transformers.cache_utils import Cache

setup.py

@@ -62,7 +62,7 @@
     "neural-compressor": ["neural-compressor>=2.2.0", "onnxruntime<1.15.0", "accelerate"],
     "openvino": ["openvino>=2023.3", "nncf>=2.11.0", "openvino-tokenizers[transformers]"],
     "nncf": ["nncf>=2.11.0"],
-    "ipex": ["intel-extension-for-pytorch", "transformers>=4.39.0,<=4.41.2"],
+    "ipex": ["intel-extension-for-pytorch", "transformers>=4.39.0,<=4.42.3"],
     "diffusers": ["diffusers"],
     "quality": QUALITY_REQUIRE,
     "tests": TESTS_REQUIRE,