Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Adds ignore_index to sequence parallel cross entropy #6882

Closed
wants to merge 5 commits into from
Closed

Adds ignore_index to sequence parallel cross entropy #6882

Changes from 2 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8d865bd
adds ignore_index to sp cross entropy
Dec 17, 2024
69443da
Merge branch 'master' into add-ignore-index-sp-loss
tjruwase Dec 17, 2024
e58080e
Formatting fixes
loadams Dec 18, 2024
e82b795
Merge branch 'master' into add-ignore-index-sp-loss
loadams Dec 18, 2024
87643c3
Merge branch 'master' into add-ignore-index-sp-loss
hwchen2017 Dec 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 13 additions & 5 deletions deepspeed/sequence/cross_entropy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,17 +9,21 @@


class _VocabSequenceParallelCrossEntropy(torch.autograd.Function):

@staticmethod
def forward(ctx, vocab_seq_parallel_logits, target, sp_group):
def forward(ctx, vocab_seq_parallel_logits, target, sp_group, ignore_index=-100):
# vocab_seq_parallel_logits: [S/P, B, V]
# target: [S/P, B]
# return: [S, B]

# Need softmax for backward
softmax = torch.nn.functional.softmax(vocab_seq_parallel_logits, dim=-1)
ctx.vocab_size = vocab_seq_parallel_logits.size(2)
loss = torch.nn.functional.nll_loss(softmax.log().view(-1, ctx.vocab_size), target.view(-1), reduction='none')
loss = torch.nn.functional.nll_loss(
softmax.log().view(-1, ctx.vocab_size),
target.view(-1),
ignore_index=ignore_index,
reduction="none",
)

sp_world_size = dist.get_world_size(sp_group)
sp_rank = dist.get_rank(sp_group)
Expand Down Expand Up @@ -56,5 +60,9 @@ def backward(ctx, grad_output):
return grad_input, None, None, None


def vocab_sequence_parallel_cross_entropy(vocab_parallel_logits, target, sp_group):
return _VocabSequenceParallelCrossEntropy.apply(vocab_parallel_logits, target, sp_group)
def vocab_sequence_parallel_cross_entropy(
loadams marked this conversation as resolved.
Show resolved Hide resolved
vocab_parallel_logits, target, sp_group, ignore_index=-100
):
return _VocabSequenceParallelCrossEntropy.apply(
vocab_parallel_logits, target, sp_group, ignore_index
)
Loading