Script to merge Flux diffuser into 1 safetensor file

[pivtienduc]

Hi everyone

I found Flux model which is split into 3 parts. So far I can't find any script to merge splitted parts of Flux model into one

Can anyone help me to do it

Thank in advance

[a-r-r-o-w]

You can load the model using from_pretrained and save it again, but by specifying max_shard_size="30GB" to get a single large file. The docs for loading/saving pipelines contains more info about from_pretrained and save_pretrained.

[punitvara]

@hlky Should we close this issue as PR is already raised for this ?

[pivtienduc]

@hlky Should we close this issue as PR is already raised for this ?

Yes I will close it

[hlky]

Typically issues are not closed until they are resolved i.e. by a PR merging.

[pivtienduc]

Typically issues are not closed until they are resolved i.e. by a PR merging.

Actually I found another way to solve it so don't need this one anymore

[clort81]

Typically issues are not closed until they are resolved i.e. by a PR merging.

Actually I found another way to solve it so don't need this one anymore

It would have been helpful to others if you had shared the solution you found here.

Please take note that yours was uncooperative behavior, then reflect upon it before making future requests for help from the community.

[hlky]

Hi @clort81, here's an example of max_shard_size.

from diffusers import DiffusionPipeline, FluxTransformer2DModel

pipeline = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell")
pipeline.save_pretrained("FLUX.1-schnell/", max_shard_size="30GB")

transformer = FluxTransformer2DModel.from_pretrained("black-forest-labs/FLUX.1-schnell", subfolder="transformer")
transformer.save_pretrained("FLUX.1-schnell/transformer/", max_shard_size="30GB")