Upload modular_isaac.py
Hello and congrats for the release!
This PR makes this model load with no additional dependency, only transformers, which is very convenient for the users:
from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, AutoProcessor
tokenizer = AutoTokenizer.from_pretrained("Perceptron/Isaac-0.1", trust_remote_code=True, use_fast=False)
config = AutoConfig.from_pretrained("Perceptron/Isaac-0.1", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("Perceptron/Isaac-0.1", tokenizer=tokenizer, config=config)
model = AutoModelForCausalLM.from_pretrained("Perceptron/Isaac-0.1", trust_remote_code=True)
you can also add a small inference notebook I made by replacing the username to Perceptron: https://colab.research.google.com/drive/1BHl2ZT8cYZ0HlP_q4HllFuCXWIBX_R_2?usp=sharing
if you add the "notebook.ipynb" repo to it's one-click open in the repository, making it easier for people to try out your model as well!
What is your recommendation around this for the core transformers repo: https://github.com/huggingface/transformers/pull/40962
TensorStream is a core abstraction for us which we will continue to optimize and improve - our intention of keeping it in the perceptron package was to make it easier to centralize improvements across open code bases.