bench/generation/evaluate_model.py [35:48]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
@torch.no_grad()
def calibrate(model, tokenizer, batch_size, batches):
    samples = batch_size * batches
    cal_dataset = load_dataset("lambada", split=["validation"])[0]
    model.eval()
    total = 0
    for batch in cal_dataset.iter(batch_size=batch_size):
        inputs = tokenizer(batch["text"], return_tensors="pt", padding=True)
        input_ids = inputs.input_ids.to(model.device)
        attention_mask = inputs.attention_mask.to(model.device)
        model(input_ids, attention_mask=attention_mask)
        total += input_ids.size(0)
        if total >= samples:
            break
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



bench/generation/setup/quanto.py [24:37]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
@torch.no_grad()
def calibrate(model, tokenizer, batch_size, batches):
    samples = batch_size * batches
    cal_dataset = load_dataset("lambada", split=["validation"])[0]
    model.eval()
    total = 0
    for batch in cal_dataset.iter(batch_size=batch_size):
        inputs = tokenizer(batch["text"], return_tensors="pt", padding=True)
        input_ids = inputs.input_ids.to(model.device)
        attention_mask = inputs.attention_mask.to(model.device)
        model(input_ids, attention_mask=attention_mask)
        total += input_ids.size(0)
        if total >= samples:
            break
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -