[BUG] Trainer have no attribute 'use_cuda_amp'

[nghiant03]

Bug description

Trainer.train() when finishing training return:
AttributeError: 'Trainer' object has no attribute 'use_cuda_amp'
During handling of the above exception, another exception occurred:
AttributeError: 'Trainer' object has no attribute 'use_amp'

Steps/Code to reproduce bug

Train a model with these configurations:
training_args = T4RecTrainingArguments(
output_dir=lb_out.as_posix(),
max_sequence_length=20,
data_loader_engine='nvtabular',
num_train_epochs=1,
dataloader_drop_last=False,
per_device_train_batch_size=256,
per_device_eval_batch_size=16,
gradient_accumulation_steps=1,
learning_rate=0.0005,
report_to=[],
save_strategy='no',
logging_steps=500
)

trainer = Trainer(
model=model,
args=training_args,
schema=schema,
compute_metrics=True,
)

Expected behavior

Model finish training without returning error

Environment details

• Transformers4Rec version: 23.08.00
• Platform:Ubuntu 22.04.5 LTS
• Python version:3.10.16
• Huggingface Transformers version:4.45.0
• PyTorch version (GPU?):2.4.1
• Tensorflow version (GPU?):

Additional context

It seems that transformers have remove CUDA amp entirely for CPU amp. huggingface/transformers#27760

[greekera1000]

Hi @nghiant03 ! I’d like to take a look at this issue and investigate a fix for the use_cuda_amp / use_amp compatibility problem with newer versions of Hugging Face Transformers. I’ll first try to reproduce the issue and then work on a small compatibility fix and regression test if possible.