slapo.op.cross_entropy

Functions:

vocab_parallel_cross_entropy(...[, ...])

Performs cross entropy loss when logits are split across tensor parallel ranks

Classes:

ParallelCrossEntropy([group])

slapo.op.cross_entropy.vocab_parallel_cross_entropy(vocab_parallel_logits, target, label_smoothing=0.0, group=None)

Performs cross entropy loss when logits are split across tensor parallel ranks

Parameters

• vocab_parallel_logits – logits split across tensor parallel ranks dimension is [sequence_length, batch_size, hidden_size]

• target – correct vocab ids of dimension [sequence_length, micro_batch_size]

• label_smoothing – smoothing factor, must be in range [0.0, 1.0) default is no smoothing (=0.0)

• group – torch.distributed group

class slapo.op.cross_entropy.ParallelCrossEntropy(group=None)

Methods:

forward(outputs, labels)

Defines the computation performed at every call.

forward(outputs, labels)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.