Note, when running generate.py
and asking your first question, it will download the model(s), which for the 6.9B model takes about 15 minutes per 3 pytorch bin files if have 10MB/s download.
If all data has been put into ~/.cache
by HF transformers, then these following steps (those related to downloading HF models) are not required.
- Download model and tokenizer of choice
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = 'h2oai/h2ogpt-oasst1-512-12b'
model = AutoModelForCausalLM.from_pretrained(model_name)
model.save_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.save_pretrained(model_name)
- Download reward model, unless pass
--score_model='None'
togenerate.py
# and reward model
reward_model = 'OpenAssistant/reward-model-deberta-v3-large-v2'
from transformers import AutoModelForSequenceClassification, AutoTokenizer
model = AutoModelForSequenceClassification.from_pretrained(reward_model)
model.save_pretrained(reward_model)
tokenizer = AutoTokenizer.from_pretrained(reward_model)
tokenizer.save_pretrained(reward_model)
- For LangChain support, download embedding model:
hf_embedding_model = "sentence-transformers/all-MiniLM-L6-v2"
model_kwargs = 'cpu'
from langchain.embeddings import HuggingFaceEmbeddings
embedding = HuggingFaceEmbeddings(model_name=hf_embedding_model, model_kwargs=model_kwargs)
- For HF inference server and OpenAI, this downloads the tokenizers used for Hugging Face text generation inference server and gpt-3.5-turbo:
import tiktoken
encoding = tiktoken.get_encoding("cl100k_base")
encoding = tiktoken.encoding_for_model("gpt-3.5-turbo")
- Run generate with transformers in Offline Mode
HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 python generate.py --base_model='h2oai/h2ogpt-oasst1-512-12b' --gradio_offline_level=2 --share=False
Some code is always disabled that involves uploads out of user control: Huggingface telemetry, gradio telemetry, chromadb posthog.
The additional option --gradio_offline_level=2
changes fonts to avoid download of google fonts. This option disables google fonts for downloading, which is less intrusive than uploading, but still required in air-gapped case. The fonts don't look as nice as google fonts, but ensure full offline behavior.
If the front-end can still access internet, but just backend should not, then one can use --gradio_offline_level=1
for slightly better-looking fonts.
Note that gradio attempts to download iframeResizer.contentWindow.min.js, but nothing prevents gradio from working without this. So a simple firewall block is sufficient. For more details, see: AUTOMATIC1111/stable-diffusion-webui#10324.
- Disable access or port
To ensure nobody can access your gradio server, disable the port via firewall. If that is a hassle, then one can enable authentication by adding to CLI when running python generate.py
:
--auth=[('jon','password')]
with no spaces. Run python generate.py --help
for more details.