mobiuslabsgmbh/Llama-2-7b-chat-hf_2bitgs8_hqq

This is an experimental HQQ 2-bit quantized Llama2-7B-chat model using a low-rank adapter to improve the performance (referred to as HQQ+).

Quantizing small models at extreme low-bits is a challenging task. The purpose of this model is to show the community what to expect when fine-tuning such models. We notice that, when given more specialized data, the low-bit model can even outperform the full-precision model at some tasks.

This version offloads the meta-data to the CPU, so only the 2-bit weights and the low-rank adapters are stored in the GPU memory.

Datasets

The adapter was trained via SFT on random subsets of the following:

Base Model

• wikitext-2-raw-v1 (full)

Chat Model

• timdettmers/openassistant-guanaco (full)
• microsoft/orca-math-word-problems-200k (10K)
• meta-math/MetaMathQA (10K)
• HuggingFaceH4/ultrafeedback_binarized (10K - chosen answers only)

Performance

Models	Llama2-7B (fp16)	Llama2-7B (HQQ 2-bit)	Llama2-7B (HQQ+ 2-bit)	Quip# (2-bit)
Wiki Perpexlity	5.18	6.06	5.14	8.54
VRAM (GB)	13.5	2.6	2.69	2.72
forward time (sec)	0.1	0.221	0.27	0.353

Models	Llama2-7B-chat (fp16)	Llama2-7B-chat (HQQ 2-bit)	Llama2-7B-chat (HQQ+ 2-bit)
ARC (25-shot)	53.67	45.56	47.01
HellaSwag (10-shot)	78.56	73.59	73.74
MMLU (5-shot)	48.16	43.18	43.33
TruthfulQA-MC2	45.32	43.1	42.66
Winogrande (5-shot)	72.53	67.32	71.51
GSM8K (5-shot)	23.12	9.7	28.43
Average	53.56	47.08	51.11

Usage

First, install the latest version of HQQ:

pip install git+https://github.com/mobiusml/hqq.git

Then you can use the sample code below:

from hqq.engine.hf import HQQModelForCausalLM, AutoTokenizer

#Load the model
model_id = 'mobiuslabsgmbh/Llama-2-7b-chat-hf_2bitgs8_hqq' 
model     = HQQModelForCausalLM.from_quantized(model_id, adapter='adapter_v0.1.lora')
tokenizer = AutoTokenizer.from_pretrained(model_id)

#Setup Inference Mode
tokenizer.add_bos_token = False
tokenizer.add_eos_token = False
if not tokenizer.pad_token: tokenizer.add_special_tokens({'pad_token': '[PAD]'})
model.config.use_cache  = True
model.eval();

# Optional: torch compile for faster inference
# model = torch.compile(model)

#Streaming Inference
import torch, transformers
from threading import Thread

def chat_processor(chat, max_new_tokens=100, do_sample=True, device='cuda'):
    tokenizer.use_default_system_prompt = False
    streamer = transformers.TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)

    generate_params = dict(
        tokenizer("<s> [INST] " + chat + " [/INST] ", return_tensors="pt").to(device),
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=do_sample,
        pad_token_id=tokenizer.pad_token_id,
        top_p=0.90 if do_sample else None,
        top_k=50 if do_sample else None,
        temperature= 0.6 if do_sample else None,
        num_beams=1,
        repetition_penalty=1.2,
    )

    t = Thread(target=model.generate, kwargs=generate_params)
    t.start()
    
    print("User: ", chat); 
    print("Assistant: ");
    outputs = ""
    for text in streamer:
        outputs += text
        print(text, end="", flush=True)

    torch.cuda.empty_cache()
  
    return outputs

Example

outputs = chat_processor("If you had 5 apples yesterday and you ate 2 today morning, how many apples do you have this evening?", max_new_tokens=1000, do_sample=False)

User:  If you had 5 apples yesterday and you ate 2 today morning, how many apples do you have this evening?
Assistant: 
You started with 5 apples.You ate 2 of them so now you have 5-2=3 apples left.So by the evening you will still have 3 apples.