compressa-ai/Llama-3-8B-Instruct-OmniQuant

Llama 3 8B Instruct – OmniQuant

Based on Llama 3 8B Instruct.

Quantized with OmniQuant.

Evaluation

PPL (↓)

	wiki
FP	8,29
Quantized	8,97

Accuracy on English Benchmarks, % (↑)

	piqa	arc_easy	arc_challenge	boolq	hellaswag	winogrande	mmlu_humanities	mmlu_social_sciences	mmlu_stem	mmlu_other
FP	78,7	81,6	53,0	83,1	57,7	72,1	67,0	70,9	54,5	68,2
Quantized	77,2	80,7	51,8	82,8	56,8	72,5	63,4	67,6	50,1	65,0

Accuracy on Russian Benchmarks, % (↑)

	danetqa	terra	rwsd	muserc	rucos	lidirus	parus	rcb	russe	rucola
FP	78,6	60,9	65,7	56,1	64,9	63,2	71,0	34,1	60,8	64,1
Quantized	71,6	60,6	52,5	63,7	57,3	57,2	74,0	33,6	36,9	67,5

Summary

	Avg acc diff on Eng, % (↑)	Avg acc diff on Rus, % (↑)	Occupied disk space, % (↓)
FP	0	0	100
Quantized	-1,9	-4,5	35,7

Examples

Imports and Model Loading

Expand

import gc

import auto_gptq.nn_modules.qlinear.qlinear_cuda as qlinear_cuda
import auto_gptq.nn_modules.qlinear.qlinear_triton as qlinear_triton
import torch

from accelerate import (
    init_empty_weights,
    infer_auto_device_map,
    load_checkpoint_in_model,
)
from tqdm import tqdm
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    pipeline,
)


def get_named_linears(model):
    return {
        name: module for name, module in model.named_modules()
        if isinstance(module, torch.nn.Linear)
    }


def set_module(model, name, module):
    parent = model
    levels = name.split('.')

    for i in range(len(levels) - 1):
        cur_name = levels[i]

        if cur_name.isdigit():
            parent = parent[int(cur_name)]
        else:
            parent = getattr(parent, cur_name)

    setattr(parent, levels[-1], module)


def load_model(model_path):
    # Based on: https://github.com/OpenGVLab/OmniQuant/blob/main/runing_quantized_mixtral_7bx8.ipynb

    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)

    if not hasattr(config, 'quantization_config'):
        raise AttributeError(
            f'No quantization info found in model config "{model_path}"'
            f' (`quantization_config` section is missing).'
        )

    wbits = config.quantization_config['bits']
    group_size = config.quantization_config['group_size']

    # We are going to init an ordinary model and then manually replace all Linears with QuantLinears
    del config.quantization_config

    with init_empty_weights():
        model = AutoModelForCausalLM.from_config(config=config, torch_dtype=torch.float16, trust_remote_code=True)

    layers = model.model.layers

    for i in tqdm(range(len(layers))):
        layer = layers[i]
        named_linears = get_named_linears(layer)

        for name, module in named_linears.items():
            params = (
                wbits, group_size,
                module.in_features, module.out_features,
                module.bias is not None
            )

            if wbits in [2, 4]:
                q_linear = qlinear_triton.QuantLinear(*params)
            elif wbits == 3:
                q_linear = qlinear_cuda.QuantLinear(*params)
            else:
                raise NotImplementedError("Only 2, 3 and 4 bits are supported.")

            q_linear.to(next(layer.parameters()).device)
            set_module(layer, name, q_linear)

    torch.cuda.empty_cache()
    gc.collect()

    model.tie_weights()
    device_map = infer_auto_device_map(model)

    print("Loading pre-computed quantized weights...")

    load_checkpoint_in_model(
        model, checkpoint=model_path,
        device_map=device_map, offload_state_dict=True,
    )

    print("Model loaded successfully!")

    return model

Inference

model_path = "compressa-ai/Llama-3-8B-Instruct-OmniQuant"

model = load_model(model_path).cuda()
tokenizer = AutoTokenizer.from_pretrained(
    model_path, use_fast=False, trust_remote_code=True
)

# Llama 3 "specifics"
# https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/discussions/4
terminators = [
    tokenizer.convert_tokens_to_ids("<|end_of_text|>"),
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

system_message = "You are a friendly chatbot who always responds in the style of a pirate."
user_message = "Where are we going, Captain?"
messages = [
    {"role": "system", "content": system_message},
    {"role": "user", "content": user_message},
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

inputs = tokenizer(prompt, return_tensors="pt")
inputs = {k: v.cuda() for k, v in inputs.items()}

outputs = model.generate(
    **inputs, max_new_tokens=512,
    do_sample=True, temperature=0.7, top_p=0.95,
    eos_token_id=terminators,
)

response = tokenizer.decode(outputs[0])
continuation = response.removeprefix(prompt).removesuffix(tokenizer.eos_token)

print(f'Prompt:\n{prompt}')
print(f'Continuation:\n{continuation}\n')

Inference Using Pipeline

pipe = pipeline(
    "text-generation",
    model=model, tokenizer=tokenizer,
    eos_token_id=terminators,
    max_new_tokens=512, do_sample=True,
    temperature=0.7, top_p=0.95,
    device=0,
)

prompt = pipe.tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

outputs = pipe(prompt)

response = outputs[0]["generated_text"]
continuation = response.removeprefix(prompt)

print(f'Prompt:\n{prompt}')
print(f'Continuation:\n{continuation}\n')