Remove hf_auth_token use

[Abhishek-Varma]

-- This commit removes --hf_auth_token uses from vicuna.py.
-- It adds llama2 models based on daryl49's HF.

Signed-off-by: Abhishek Varma abhishek@nod-labs.com

[Abhishek-Varma]

Currently marking it as draft since 13B and 70B paths need testing.
CC: @powderluv

[powderluv]

If we only download the mlir we wouldn't hit the token right?

[Abhishek-Varma]

If we only download the mlir we wouldn't hit the token right?

I did try doing that but during the run saw that we will hit that issue - because we're using tokenizers to decode each generated token. And this tokenizer is being instantiated as per the HF repo we use.

[Abhishek-Varma]

If we only download the mlir we wouldn't hit the token right?

I did try doing that but during the run saw that we will hit that issue - because we're using tokenizers to decode each generated token. And this tokenizer is being instantiated as per the HF repo we use.

Even this would work since we're anyway blocking the IR generation.
It'd then essentially download the tokenizer's config files from daryl149/llama-2-7b-hf and we already have the MLIR generated from meta-llama/Llama-2-7b-chat-hf.

I verified it on CPU for llama2 7B.

With this PR we don't need to maintain config files for tokenizer but we're changing the base HF repo and this would impact the workflow when the IR generation is given a green signal.

But with the other PR we only need to incur an overhead for maintaining the config files - keeping rest of the infra same.