openbmb

/

MiniCPM-V-2_6

like 826

Follow

OpenBMB 373

adding _no_split_module for loading model with device_map="auto"

Thanks for your contribution， but may be it's not necessary to add _no_split_modules in model code
Users can set no_split_module_classes in load_checkpoint_and_dispatch like:

model = load_checkpoint_and_dispatch(model, model_name, dtype=torch.bfloat16, 
    device_map="auto",  no_split_module_classes=['SiglipTextEmbeddings', 'SiglipEncoderLayer'])

@finalf0 Thank you for your comment!

I was trying to deploy it on AWS sagemaker with DJL. And it was configured to load without the option of no_split_module_classes. Let me know if I missed some details.

I checked the Siglip implementation in transformers 4.44.2, they also added these properties. Maybe import from the library is a better way?