Source codes to quantize the LLaMA 3.1 405B model

#10
by shuyuej - opened

Could you please share the source codes for quantizing the LLaMA 3.1 405B model?

Thank you very much in advance!

Hugging Quants org

Hi here @shuyuej , those are already within the model card 🤗

See https://huggingface.co./hugging-quants/Meta-Llama-3.1-405B-Instruct-GPTQ-INT4#quantization-reproduction

import random

import numpy as np
import torch

from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
from datasets import load_dataset
from transformers import AutoTokenizer

pretrained_model_dir = "meta-llama/Meta-Llama-3.1-405B-Instruct"
quantized_model_dir = "meta-llama/Meta-Llama-3.1-405B-Instruct-GPTQ-INT4"

print("Loading tokenizer, dataset, and tokenizing the dataset...")
tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True)
dataset = load_dataset("Salesforce/wikitext", "wikitext-2-raw-v1", split="train")
encodings = tokenizer("\n\n".join(dataset["text"]), return_tensors="pt")

print("Setting random seeds...")
random.seed(0)
np.random.seed(0)
torch.random.manual_seed(0)

print("Setting calibration samples...")
nsamples = 128
seqlen = 2048
calibration_samples = []
for _ in range(nsamples):
    i = random.randint(0, encodings.input_ids.shape[1] - seqlen - 1)
    j = i + seqlen
    input_ids = encodings.input_ids[:, i:j]
    attention_mask = torch.ones_like(input_ids)
    calibration_samples.append({"input_ids": input_ids, "attention_mask": attention_mask})

quantize_config = BaseQuantizeConfig(
    bits=4,  # quantize model to 4-bit
    group_size=128,  # it is recommended to set the value to 128
    desc_act=True,  # set to False can significantly speed up inference but the perplexity may slightly bad
    sym=True,  # using symmetric quantization so that the range is symmetric allowing the value 0 to be precisely represented (can provide speedups)
    damp_percent=0.1,  # see https://github.com/AutoGPTQ/AutoGPTQ/issues/196
)

# load un-quantized model, by default, the model will always be loaded into CPU memory
print("Load unquantized model...")
model = AutoGPTQForCausalLM.from_pretrained(pretrained_model_dir, quantize_config)

# quantize model, the examples should be list of dict whose keys can only be "input_ids" and "attention_mask"
print("Quantize model with calibration samples...")
model.quantize(calibration_samples)

# save quantized model using safetensors
model.save_quantized(quantized_model_dir, use_safetensors=True)

@alvarobartt Thank you very much for sharing the codes.
How can we enable multi-GPU quantization while using the model.quantize()?
It seems that only one GPU (cuda:0) is running, and the GPU memory is not enough to conduct quantization for such a model.

Thank you very much again, and have a nice day!

Best regards,

Shuyue
August 3rd, 2024

@alvarobartt I always get this error on my side while I am using these codes:

torch._C._LinAlgError: linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 1 is not positive-definite).

Sign up or log in to comment