Method forward

mpu/cross_entropy.py:28–81

(ctx, vocab_parallel_logits, target)

    @staticmethod
    def forward(ctx, vocab_parallel_logits, target):

        # Copy so the input remains unchanged.
        logits = vocab_parallel_logits.clone()
        # Maximum value along vocab dimension across all GPUs.
        logits_max = torch.max(logits, dim=-1)[0]
        torch.distributed.all_reduce(logits_max,
                                     op=torch.distributed.ReduceOp.MAX,
                                     group=get_model_parallel_group())
        # Subtract the maximum value.
        logits.sub_(logits_max.unsqueeze(dim=-1))
        # Sum of exponential of logits along vocab dimension across all GPUs.
        exp_logits = logits.exp()
        sum_exp_logits = exp_logits.sum(dim=-1)
        torch.distributed.all_reduce(sum_exp_logits,
                                     op=torch.distributed.ReduceOp.SUM,
                                     group=get_model_parallel_group())

        # Get the partition's vocab indecies
        get_vocab_range = VocabUtility.vocab_range_from_per_partition_vocab_size
        partition_vocab_size = vocab_parallel_logits.size()[-1]
        rank = get_model_parallel_rank()
        world_size = get_model_parallel_world_size()
        vocab_start_index, vocab_end_index = get_vocab_range(
            partition_vocab_size, rank, world_size)

        # Create a mask of valid vocab ids (1 means it needs to be masked).
        target_mask = (target < vocab_start_index) | (target >= vocab_end_index)
        masked_target = target.clone() - vocab_start_index
        masked_target[target_mask] = 0

        # Get predicted-logits = logits[target].
        # For Simplicity, we convert logits to a 2-D tensor with size
        # [*, partition-vocab-size] and target to a 1-D tensor of size [*].
        logits_2d = logits.view(-1, partition_vocab_size)
        masked_target_1d = masked_target.view(-1)
        arange_1d = torch.arange(start=0, end=logits_2d.size()[0],
                                 device=logits_2d.device)
        predicted_logits_1d = logits_2d[arange_1d, masked_target_1d]
        predicted_logits = predicted_logits_1d.view_as(target)
        predicted_logits[target_mask] = 0.0
        # All reduce is needed to get the chunks from other GPUs.
        torch.distributed.all_reduce(predicted_logits,
                                     op=torch.distributed.ReduceOp.SUM,
                                     group=get_model_parallel_group())

        # Loss = log(sum(exp(logits))) - predicted-logit.
        loss = torch.log(sum_exp_logits) - predicted_logits

        # Store softmax, target-mask and masked-target for backward pass.
        exp_logits.div_(sum_exp_logits.unsqueeze(dim=-1))
        ctx.save_for_backward(exp_logits, target_mask, masked_target_1d)

        return loss

    @staticmethod
    def backward(ctx, grad_output):