About transformers containers

[Heartful-echo]

Hi

I am trying to pull a Docker image from the following page in order to experience Hugging Face Transformers library.

I found the following 3 tags:
dustynv/transformers:git-r35.3.1
dustynv/transformers:nvgpt-r35.3.1
dustynv/transformers:r35.3.1

What are the differences between these?

[AastaLLL]

Hi,

Please check the below config for more info:

dusty-nv/jetson-containers/blob/master/packages/llm/transformers/config.py

transformers:git is installed from the source.
transformers:nvgpt is installed with nvgpt support

Thanks.

[Heartful-echo]

@AastaLLL

Thank you for your advice.
I don’t understand nvgpt.
I want to process various Transformers models as quickly as possible.
In this case, which of the following three is generally better to use?
dustynv/transformers:git-r35.3.1
dustynv/transformers:nvgpt-r35.3.1
dustynv/transformers:r35.3.1

Right now I’m using dustynv/transformers:r35.3.1.
Is this choice correct?

Thanks

[dusty_nv]

@Heartful-echo

Right now I’m using dustynv/transformers:r35.3.1.
Is this choice correct?

@Heartful-echo typically yes, that base transformers container is the one to use, and has installed the latest version of transformers released to pip/PyPi at the time it was built.

The transformers:git variant will use the latest code from https://github.com/huggingface/transformers main branch in case you need any of the latest patches/changes which have not yet been released in an official version.

The transformers:nvgpt variant has support for NVIDIA Megatron models that I use during development/testing.

[Heartful-echo]

@dusty_nv

Thank you for your answer.
I understood the difference between the three images.

Thanks.