GPU_settings

training_config = TrainingArguments(
    model_dir_str, False, True, True, False, "steps",
    per_device_train_batch_size=64, #76% @ 24 batch size #76% @ 32 batch size try 64 batch size next time 
    per_device_eval_batch_size=64, #was 24 now 32
    gradient_accumulation_steps=3, #change this to 4
    eval_accumulation_steps=None,
    eval_steps=2000,
    learning_rate=1e-4,
    weight_decay=0.01,
    max_grad_norm=1.0,
    max_steps=50000,
    lr_scheduler_type="cosine",
    warmup_ratio=0.08,
    log_level="debug",
    logging_strategy="steps",
    logging_steps=20,
    save_strategy="steps",
    save_steps=1000,
    save_total_limit=5,
    no_cuda=not USE_CUDA,
    seed=444,
    fp16=FP16,
    fp16_full_eval=FP16_EVAL,
    bf16=BF16,
    bf16_full_eval=BF16_EVAL,
    load_best_model_at_end=True,
    label_smoothing_factor=0.,
    optim="adamw_torch",
    report_to=["tensorboard"],
    gradient_checkpointing=True,
    dataloader_num_workers=8, #added to fix trashing isssue with the gpu not having enough data to process
    dataloader_pin_memory=True, #we want the dataset in memory
    torch_compile=True #added to speed up 



Better sugested by ai
training_config = TrainingArguments(
    model_dir_str, False, True, True, False, "steps",
    per_device_train_batch_size=64, #76% @ 24 batch size #76% @ 32 batch size try 64 batch size next time 
    per_device_eval_batch_size=64, #was 24 now 32
    gradient_accumulation_steps=3, #change this to 4
    eval_accumulation_steps=None,
    eval_steps=3000,
    eval_delay=6000,
    learning_rate=1e-4,
    weight_decay=0.01,
    max_grad_norm=1.0,
    max_steps=30000,
    lr_scheduler_type="cosine",
    warmup_ratio=0.08,
    log_level="debug",
    logging_strategy="steps",
    logging_steps=100,
    save_strategy="steps",
    save_steps=3000,
    save_total_limit=5,
    no_cuda=not USE_CUDA,
    seed=444,
    fp16=FP16,
    fp16_full_eval=FP16_EVAL,
    bf16=BF16,
    bf16_full_eval=BF16_EVAL,
    load_best_model_at_end=True,
    label_smoothing_factor=0.05,
    optim="adamw_torch",
    report_to=["tensorboard"],
    gradient_checkpointing=False,
    dataloader_num_workers=8, #added to fix trashing isssue with the gpu not having enough data to process
    dataloader_pin_memory=True, #we want the dataset in memory
    torch_compile=True #added to speed up