You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hello.
When using tensor parallel on bloom (tp_size = 8), we find that the cross_entropy loss computed by mpu.cross_entropy is different from torch.nn.functional.cross_entropy. The difference is about 1% for our data.
For the implementation of mpu.cross_entropy, we find that the loss is computed on the partition_vocab_size which is 8 times smaller than vocab_size (tp_size = 8). We think maybe this implementation causes the difference above.
In this case, is this implementation correct? Or can this implementation ensure the performance when using tensor parallel?
The text was updated successfully, but these errors were encountered:
Hello.
When using tensor parallel on bloom (tp_size = 8), we find that the cross_entropy loss computed by mpu.cross_entropy is different from torch.nn.functional.cross_entropy. The difference is about 1% for our data.
For the implementation of mpu.cross_entropy, we find that the loss is computed on the partition_vocab_size which is 8 times smaller than vocab_size (tp_size = 8). We think maybe this implementation causes the difference above.
In this case, is this implementation correct? Or can this implementation ensure the performance when using tensor parallel?
The text was updated successfully, but these errors were encountered: