def train_baseline()

in benchmarks/fp8/ms_amp/distrib_deepspeed.py [0:0]

• 57 lines of code
• 4 McCabe index (conditional complexity)

def train_baseline(zero_stage: int = 1, opt_level: str = "O1"):
    set_seed(42)
    accelerator = Accelerator()
    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(
        MODEL_NAME, accelerator=accelerator
    )

    import numpy as np

    config = {
        "train_batch_size": 32,
        "train_micro_batch_size_per_gpu": 16,
        "gradient_accumulation_steps": 1,
        "zero_optimization": {
            "stage": zero_stage,
            "offload_optimizer": {"device": "none", "nvme_path": None},
            "offload_param": {"device": "none", "nvme_path": None},
        },
        "gradient_clipping": 1.0,
        "steps_per_print": np.inf,
        "bf16": {"enabled": True},
        "fp16": {"enabled": False},
        "zero_allow_untested_optimizer": True,
        "msamp": {
            "enabled": True,
            "opt_level": opt_level,
        },
    }
    (
        model,
        optimizer,
        _,
        _,
    ) = msamp_deepspeed.initialize(
        model=model,
        optimizer=optimizer,
        config_params=config,
    )

    base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
    model.train()

    for _ in range(2):
        for batch in train_dataloader:
            outputs = model(**batch)
            loss = outputs.loss
            model.backward(loss)
            model.step()
            for _ in range(accelerator.num_processes):
                lr_scheduler.step()

    trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator)
    model.destroy()
    torch.cuda.empty_cache()
    AcceleratorState()._reset_state(True)
    assert trained_model_results["accuracy"] > base_model_results["accuracy"], (
        f"Accuracy should be higher for the trained model: {trained_model_results['accuracy']} > {base_model_results['accuracy']}"
    )
    assert trained_model_results["f1"] > base_model_results["f1"], (
        f"F1 score should be higher for the trained model: {trained_model_results['f1']} > {base_model_results['f1']}"
    )

    return base_model_results, trained_model_results