Upgrade transformers to v4.51.3 (#3685)

Author: blzheng

docs/tutorials/features/fast_bert.md

@@ -9,7 +9,7 @@ Currently `ipex.fast_bert` API is only well optimized for training. For inferenc
 
 ### Prerequisite
 
-- Transformers 4.6.0 ~ 4.48.0
+- Transformers 4.6.0 ~ 4.51.3
 
 ### Usage Example
 

examples/cpu/features/fast_bert/README.md

@@ -5,7 +5,7 @@
 Currently `ipex.fast_bert` API is only well optimized for training. For inference, it ensures functionality, while to get peak perf, please use `ipex.optimize` API + torchscript.
 
 # Prerequisite:
-Transformers 4.6.0 ~ 4.48.0
+Transformers 4.6.0 ~ 4.51.3
 
 # Usage Example:
 Training:

examples/cpu/llm/fine-tuning/requirements.txt

@@ -6,6 +6,6 @@ black[jupyter]
 datasets
 fire
 peft
-transformers==4.48.0
+transformers==4.51.3
 gradio
 sentencepiece

examples/cpu/llm/inference/requirements.txt

@@ -1 +1 @@
-transformers==4.48.0
+transformers==4.51.3

intel_extension_for_pytorch/cpu/tpp/fused_bert.py

@@ -1250,7 +1250,7 @@ def fast_bert(model, dtype=torch.float, optimizer=None, unpad=False):
     # tpp bert optimization depends on the transformers repo to implementate the related module
     installed_pkg = {dist.metadata["Name"].lower() for dist in distributions()}
     min_version = "4.6.0"
-    max_version = "4.48.0"
+    max_version = "4.51.3"
     if "transformers" not in installed_pkg:
         raise RuntimeError(
             "Please installed the transformers with version: between {} and {}".format(
@@ -1275,9 +1275,10 @@ def fast_bert(model, dtype=torch.float, optimizer=None, unpad=False):
         position_ids_persistent = True
     PT_OPTIMIZER_TO_TPP_OPTIMIZER = {
         torch.optim.AdamW: AdamW,
-        transformers.optimization.AdamW: AdamW,
         torch.optim.SGD: SGD,
     }
+    if hasattr(transformers.optimization, "AdamW"):
+        PT_OPTIMIZER_TO_TPP_OPTIMIZER[transformers.optimization.AdamW] = AdamW
     if dtype not in (
         torch.float,
         torch.bfloat16,

intel_extension_for_pytorch/transformers/generation/__init__.py

@@ -1,5 +1,5 @@
-from .beam_search import _beam_search
+from .beam_search import _beam_search, _beam_search_legacy
 from .greedy_search import _greedy_search
 from .sample import _sample
-from .beam_sample import _beam_sample
+from .beam_sample import _beam_sample, _beam_sample_legacy
 from .utils import whisper_generate