Closed dacorvo closed 7 months ago
This modifies the decoder models export code to reduce the disk usage when creating checkpoints:
torch_dtype = auto
float32
snapshot_download
pytorch
safetensors
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
What does this PR do?
This modifies the decoder models export code to reduce the disk usage when creating checkpoints:
torch_dtype = auto
when loading the model to avoid casting weigths tofloat32
(the default),snapshot_download
before exporting to avoid downloading bothpytorch
andsafetensors
weights.