Loss compilation issues with Autogen#
Custom loss functions with AutoGen#
When creating custom losses, you might encounter compilation failures. To address this, wrap your custom loss class with the @autogen_loss decorator, which enables AutoGen to handle the compilation of these custom losses efficiently.
from cerebras_pytorch/src/cerebras/pytorch/nn/modules import autogen_loss
@autogen_loss
class CustomLoss(nn.Module):
def __init__(self, ...):
Improving loss function performance#
Enable autogen to use fused autogenerated graphs for losses in PyTorch, enhancing performance. Set use_autogen = True
when defining your loss:
loss = MSELoss(..., use_autogen=True)
Supported losses include L1Loss, MSELoss, and others. Note that CosineEmbeddingLoss is not supported and will default to primitive kernels.