import torch import torch.distributed as dist from torch.autograd import Function from torch.distributed import ProcessGroup __all__ = ['DistCrossEntropy', 'cross_entropy_1d'] class DistCrossEntropy(Function): r""" Overwrite the forward and backward function to calculate the cross entropy loss before gather Args: Function (:class:`torch.autograd.Function`): default """ @staticmethod def forward(ctx, vocab_logits: torch.Tensor, target: torch.Tensor, ignore_index: int, process_group: ProcessGroup): r""" Calculate the cross entropy loss before gather, the origin loss function is as follows: loss = -log(exp(x[class])/sum(exp(x[i])) and can be rewrite as: loss = log(sum(exp(x[i])) - x[class] To avoid the `nan` of log(sum(exp(x[i]))), we minus the max of x[i] Args: vocab_logits (:class:`torch.Tensor`): The logits of the vocabulary, shape is [batch_size, seq_len, vocab_size] labels (:class:`torch.Tensor`): The labels of the vocabulary, shape is [batch_size, seq_len] Returns: :class:`torch.Tensor`: The cross entropy loss """ # get the max logits_max = torch.max(vocab_logits, dim=-1)[0] dist.all_reduce(logits_max, op=dist.ReduceOp.MAX, group=process_group) # minus the max to avoid the result of sum of exp is too large and the log is nan vocab_logits = vocab_logits - logits_max.unsqueeze(dim=-1) # mask the target in the local device partition_vocab_size = vocab_logits.size()[-1] rank = dist.get_rank(group=process_group) world_size = dist.get_world_size(group=process_group) global_vocab_size = partition_vocab_size * world_size # [down, up) => false, other device and -100 => true delta = (global_vocab_size + world_size - 1) // world_size down_threshold = rank * delta up_threshold = down_threshold + delta mask = (target < down_threshold) | (target >= up_threshold) masked_target = target.clone() - down_threshold masked_target[mask] = 0 # reshape the logits and target # reshape the vocab_logits to [bath_size * seq_len, vocab_size] # reshape the labels to [bath_size * seq_len] logits_2d = vocab_logits.view(-1, partition_vocab_size) masked_target_1d = masked_target.view(-1) # extract the x[class] and set the x[other device] to zero pred_logits_1d = logits_2d[torch.arange(start=0, end=logits_2d.shape[0], device=logits_2d.device), masked_target_1d] pred_logits_1d = pred_logits_1d.clone().contiguous() pred_logits = pred_logits_1d.view_as(target) pred_logits[mask] = 0.0 # allreduce the get all x(i,y) dist.all_reduce(pred_logits, op=dist.ReduceOp.SUM, group=process_group) exp_logits = vocab_logits torch.exp(vocab_logits, out=exp_logits) sum_exp_logits = torch.sum(exp_logits, dim=-1) dist.all_reduce(sum_exp_logits, op=dist.ReduceOp.SUM, group=process_group) # calculate the loss # loss = log(sum(exp(x[i]))) - x[class] loss = torch.where(target == ignore_index, 0.0, torch.log(sum_exp_logits) - pred_logits) loss = torch.sum(loss).div_(torch.sum(loss != 0.0)) # calculate the softmax exp_logits.div_(sum_exp_logits.unsqueeze(dim=-1)) ctx.save_for_backward(exp_logits, mask, masked_target_1d) return loss @staticmethod def backward(ctx, grad_output): # retrieve the saved tensors exp_logits, mask, masked_target_1d = ctx.saved_tensors # use exp logits as the input grad grad_logits = exp_logits partion_vocab_size = grad_logits.shape[-1] grad_logits_2d = grad_logits.view(-1, partion_vocab_size) update = 1.0 - mask.view(-1).float() grad_logits_2d[torch.arange(0, grad_logits_2d.shape[0]), masked_target_1d] -= update grad_logits.mul_(grad_output.unsqueeze(dim=-1)) return grad_logits, None, None def cross_entropy_1d(vocab_logits: torch.Tensor, labels: torch.Tensor, ignore_index: int = -100, process_group: ProcessGroup = None) -> torch.Tensor: return DistCrossEntropy.apply(vocab_logits, labels, ignore_index, process_group)