def full_eval()

in trainer.py [0:0]

• 30 lines of code
• 5 McCabe index (conditional complexity)

def full_eval(model, optimizer, scheduler, data, block_size, hidden_size):
    model.eval()
    train_pos = 0
    nb_batches_per_iter_max = math.ceil(data.size(1) / block_size)
    h_cache = [
        torch.zeros(
            data.size(0),
            layer.attn.attn.get_cache_size(),
            hidden_size).to(data.device)
        for layer in model.module.layers]

    loss_all = 0
    actual_nb_batches_per_iter = 0
    for _ in range(nb_batches_per_iter_max):
        actual_nb_batches_per_iter += 1
        X = data[:, train_pos: train_pos + block_size].contiguous()
        Y = data[:, train_pos + 1: train_pos + block_size + 1].contiguous()

        loss, h_cache = _train_batch(
            model=model,
            optimizer=optimizer,
            scheduler=scheduler,
            X=X, Y=Y,
            h_cache=h_cache,
            eval_only=True,
            batch_split=1)
        loss_all += loss
        train_pos += block_size
        if train_pos >= data.size(1) - block_size:
            # Skip the remaining tokens as it can't make a whole block.
            # An effect on performance should be negligable for a large data.
            break

    loss_all = loss_all / actual_nb_batches_per_iter
    return loss_all