Speedup problem with GPTQModel

[ChenMnZ]

Hi

I test bitblas models with the https://github.com/ModelCloud/GPTQModel repo.

I found that the output is correct. However, BitBLAS obtains similar token generation speed in low-bits (2-bit and 4-bit) model with FP16 model. Detailed results are as follow:

the corresponding test code is:

import torch
from transformers import AutoTokenizer
from gptqmodel import GPTQModel, QuantizeConfig, get_backend

import time

def main():
    import argparse

    parser = argparse.ArgumentParser()
    parser.add_argument("--model", default=None, type=str, help="direction for saving quantization model")
    parser.add_argument("--wbits", type=int, default=4, help="quantization bits")
    parser.add_argument("--group_size", type=int, default=128, help="quantization group size")
    parser.add_argument("--test_speed", action="store_true")

    


    args = parser.parse_args()
    tokenizer = AutoTokenizer.from_pretrained(args.model, use_fast=False,legacy=False)
    model = GPTQModel.from_quantized(args.model, device_map='auto',torch_dtype=torch.float16,backend=get_backend('BITBLAS'))
    model.cuda()
    print(f"memory footprint after loading quantized model: {torch.cuda.max_memory_allocated('cuda') / 1024**3:.2f}GiB")


    if args.test_speed:
        prompt = "Write a poem about large language model:"
        input_ids = tokenizer(prompt, return_tensors='pt').input_ids.cuda()
        start_time = time.time()
        output = model.generate(inputs=input_ids, do_sample=True, top_k=10, max_new_tokens=256)
        end_time = time.time()
        speed = len(output[0])/(end_time-start_time)
        print(tokenizer.decode(output[0]))
        print(f"generation speed:{speed}token/s")
        

if __name__ =='__main__':
    main()

Do you know what is the potential problem to hinder speedup. Thank you.

[LeiWang1999]

Hi @ChenMnZ , would you mind provide the reproduce scripts of the triton v2 backend? :)

[ChenMnZ]

@LeiWang1999
Thanks for the quik reply. For the testing of triton v2, just replace model loading manner from

GPTQModel.from_quantized(args.model, device_map='auto',torch_dtype=torch.float16,backend=get_backend('BITBLAS'))

to

GPTQModel.from_quantized(args.model, device_map='auto',torch_dtype=torch.float16)

args.model should be the path of a standard GPTQ packed model. And the code will automatically choose the triton v2 kernel for 2-bit quantization.

[LeiWang1999]

@ChenMnZ Thanks, that's interesting, I‘ll take a look.

[ChenMnZ]

Hi, @LeiWang1999
Have you found a solution to this inference speed problem.

[LeiWang1999]

hi @ChenMnZ , can you provide huggingface model repos for us to reproduce?

[ChenMnZ]

@LeiWang1999
You can find the related model at https://huggingface.co/collections/ChenMnZ/efficientqatgptq-format-669e050e0060107f091edc95.

[LeiWang1999]

hi @ChenMnZ , have you met this error when loading python gptq.py --model ChenMnZ/Llama-3-8b-instruct-EfficientQAT-w4g128-BitBLAS --test_speed

Traceback (most recent call last):
  File "/root/BitBLAS/debug/gptq.py", line 35, in <module>
    main()
  File "/root/BitBLAS/debug/gptq.py", line 27, in main
    output = model.generate(inputs=input_ids, do_sample=True, top_k=10, max_new_tokens=256)
  File "/opt/conda/lib/python3.10/site-packages/gptqmodel-0.9.3.dev0+cu1201010-py3.10-linux-x86_64.egg/gptqmodel/models/base.py", line 466, in generate
    return self.model.generate(**kwargs)
  File "/opt/conda/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/opt/conda/lib/python3.10/site-packages/transformers/generation/utils.py", line 2024, in generate
    result = self._sample(
  File "/opt/conda/lib/python3.10/site-packages/transformers/generation/utils.py", line 3020, in _sample
    next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
RuntimeError: probability tensor contains either `inf`, `nan` or element < 0

[ChenMnZ]

@LeiWang1999
Sorry for the misleading.

Replace the model.cuda() to model.model.cuda() in code can solve this problem.

[w32zhong]

@ChenMnZ what GPU(s) are you running for the experiments?

[ChenMnZ]

@w32zhong Nvidia-A100 80GB