2022-11-11 09:08:17 +00:00
|
|
|
import torch
|
|
|
|
from torch.cuda.amp import custom_bwd, custom_fwd
|
|
|
|
|
|
|
|
|
|
|
|
class _VocabCrossEntropy(torch.autograd.Function):
|
|
|
|
@staticmethod
|
|
|
|
@custom_fwd
|
|
|
|
def forward(ctx, vocab_parallel_logits, target):
|
|
|
|
# Maximum value along vocab dimension across all GPUs.
|
|
|
|
logits_max = torch.max(vocab_parallel_logits, dim=-1)[0]
|
|
|
|
|
|
|
|
# Subtract the maximum value.
|
|
|
|
vocab_parallel_logits.sub_(logits_max.unsqueeze(dim=-1))
|
|
|
|
|
|
|
|
# Create a mask of valid vocab ids (1 means it needs to be masked).
|
|
|
|
target_mask = target < 0
|
|
|
|
masked_target = target.clone()
|
|
|
|
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 = vocab_parallel_logits.view(-1, vocab_parallel_logits.size(-1))
|
|
|
|
masked_target_1d = masked_target.view(-1)
|
2023-09-18 08:31:06 +00:00
|
|
|
arange_1d = torch.arange(start=0, end=logits_2d.size()[0], device=logits_2d.device)
|
2022-11-11 09:08:17 +00:00
|
|
|
predicted_logits_1d = logits_2d[arange_1d, masked_target_1d]
|
|
|
|
predicted_logits_1d = predicted_logits_1d.clone().contiguous()
|
|
|
|
predicted_logits = predicted_logits_1d.view_as(target)
|
|
|
|
predicted_logits[target_mask] = 0.0
|
|
|
|
|
|
|
|
# Sum of exponential of logits along vocab dimension across all GPUs.
|
|
|
|
exp_logits = vocab_parallel_logits
|
|
|
|
torch.exp(vocab_parallel_logits, out=exp_logits)
|
|
|
|
sum_exp_logits = exp_logits.sum(dim=-1)
|
|
|
|
|
|
|
|
# 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
|
|
|
|
@custom_bwd
|
|
|
|
def backward(ctx, grad_output):
|
|
|
|
# Retreive tensors from the forward path.
|
|
|
|
softmax, target_mask, masked_target_1d = ctx.saved_tensors
|
|
|
|
|
|
|
|
# All the inputs have softmax as their gradient.
|
|
|
|
grad_input = softmax
|
|
|
|
# For simplicity, work with the 2D gradient.
|
|
|
|
partition_vocab_size = softmax.size()[-1]
|
|
|
|
grad_2d = grad_input.view(-1, partition_vocab_size)
|
|
|
|
|
|
|
|
# Add the gradient from matching classes.
|
2023-09-18 08:31:06 +00:00
|
|
|
arange_1d = torch.arange(start=0, end=grad_2d.size()[0], device=grad_2d.device)
|
2023-09-19 06:20:26 +00:00
|
|
|
grad_2d[arange_1d, masked_target_1d] -= 1.0 - target_mask.view(-1).float()
|
2022-11-11 09:08:17 +00:00
|
|
|
|
|
|
|
# Finally elementwise multiplication with the output gradients.
|
|
|
|
grad_input.mul_(grad_output.unsqueeze(dim=-1))
|
|
|
|
|
|
|
|
return grad_input, None
|
|
|
|
|
|
|
|
|
|
|
|
def vocab_cross_entropy(vocab_logits, target):
|
|
|
|
"""helper function for the cross entropy."""
|
|
|
|
|
|
|
|
return _VocabCrossEntropy.apply(vocab_logits, target)
|