GPTQModel

GPTQ based LLM model compression/quantization toolkit with accelerated inference support for both cpu/gpu via HF, vLLM, and SGLang.

News

• 10/12/2024 ✨ v1.0.9 Move AutoRound to optional and fix pip install regression in v1.0.8.
• 10/11/2024 ✨ v1.0.8 Move QBits to optional and add wheel for python 3.12 and cuda 11.8.
• 10/08/2024 ✨ v1.0.7 Fixed marlin (faster) kernel was not auto-selected for some models.
• 09/26/2024 ✨ v1.0.6 Fixed quantized Llama 3.2 vision quantized loader.
• 09/26/2024 ✨ v1.0.5 Partial Llama 3.2 Vision model support (mllama): only text-layer quantization layers are supported for now.
• 09/26/2024 ✨ v1.0.4 Integrated Liger Kernel support for ~1/2 memory reduction on some models during quantization. Added control toggle disable parallel packing.
• 09/18/2024 ✨ v1.0.3 Added Microsoft GRIN-MoE and MiniCPM3 support.
• 08/16/2024 ✨ v1.0.2 Support Intel/AutoRound v0.3, pre-built whl packages, and PyPI release.
• 08/14/2024 ✨ v1.0.0 40% faster packing, Fixed Python 3.9 compat, added lm_eval api.
• 08/10/2024 🚀 v0.9.11 Added LG EXAONE 3.0 model support. New dynamic per layer/module flexible quantization where each layer/module may have different bits/params. Added proper sharding support to backend.BITBLAS. Auto-heal quantization errors due to small damp values.
• 07/31/2024 🚀 v0.9.10 Ported vllm/nm gptq_marlin inference kernel with expanded bits (8bits), group_size (64,32), and desc_act support for all GPTQ models with FORMAT.GPTQ. Auto calculate auto-round nsamples/seglen parameters based on calibration dataset. Fixed save_quantized() called on pre-quantized models with non-supported backends. HF transformers depend updated to ensure Llama 3.1 fixes are correctly applied to both quant and inference.
• 07/25/2024 🚀 v0.9.9: Added Llama-3.1 support, Gemma2 27B quant inference support via vLLM, auto pad_token normalization, fixed auto-round quant compat for vLLM/SGLang, and more.
• 07/13/2024 🚀 v0.9.8: Run quantized models directly using GPTQModel using fast vLLM or SGLang backend! Both vLLM and SGLang are optimized for dyanamic batching inference for maximum TPS (check usage under examples). Marlin backend also got full end-to-end in/out features padding to enhance current/future model compatibility.

Archived News:

• 07/08/2024 🚀 v0.9.7: InternLM 2.5 model support added.

• 07/08/2024 🚀 v0.9.6: Intel/AutoRound QUANT_METHOD support added for a potentially higher quality quantization with lm_head module quantization support for even more vram reduction: format export to FORMAT.GPTQ for max inference compatibility.

• 07/05/2024 🚀 v0.9.5: Intel/QBits support added for [2,3,4,8] bit quantization/inference on CPU. Cuda kernels have been fully deprecated in favor of Exllama(v1/v2)/Marlin/Triton.

• 07/03/2024 🚀 v0.9.4: HF Transformers integration added and bug fixed Gemma 2 support.

• 07/02/2024 🚀 v0.9.3: Added Gemma 2 support, faster PPL calculations on gpu, and more code/arg refractor.

• 06/30/2024 🚀 v0.9.2: Added auto-padding of model in/out-features for exllama and exllama v2. Fixed quantization of OPT and DeepSeek V2-Lite models. Fixed inference for DeepSeek V2-Lite.

• 06/29/2024 🚀 v0.9.1: With 3 new models (DeepSeek-V2, DeepSeek-V2-Lite, DBRX Converted), BITBLAS new format/kernel, proper batching of calibration dataset resulting > 50% quantization speedup, security hash check of loaded model weights, tons of refractor/usability improvements, bugs fixes and much more.

• 06/20/2924 ✨ GPTQModel v0.9.0: Thanks for all the work from ModelCloud team and the opensource ML community for their contributions!

Mission Statement

We want GPTQModel to be highly focused on GPTQ based quantization and target inference compatibility with HF Transformers, vLLM, and SGLang.

How is GPTQModel different from AutoGPTQ?

GPTQModel started out as a major refractor (fork) of AutoGTQP but has now morphed into a full-stand-in replacement with cleaner api, up-to-date model support, faster inference, faster quantization, higher quality quants and a pledge from the ModelCloud team and that we, along with the open-source ML community, will take every effort to bring the library up-to-date with latest advancements, model support, and bug fixes.

We will backport bug fixes to AutoGPTQ on a case-by-case basis.

Major Changes (Advantages) vs AutoGPTQ

• 🚀 Extensive model support for: Llama 3.2 Vision, MiniCPM3, GRIN-Moe, Phi 3.5, EXAONE 3.0, InternLM 2.5, Gemma 2, DeepSeek-V2, DeepSeek-V2-Lite, ChatGLM, MiniCPM, Phi-3, Qwen2MoE, DBRX (Converted).
• 🚀 vLLM inference integration for quantized model where format = FORMAT.GPTQ
• 🚀 SGLang inference integration for quantized model where format = FORMAT.GPTQ
• 🚀 Intel/AutoRound QUANT_METHOD support added for a potentially higher quality quantization with lm_head module quantization support for even more vram reduction: format export to FORMAT.GPTQ for max inference compatibility.
• 🚀 Intel/QBits support added for [2,3,4,8] bit quantization/inference on CPU.
• 🚀 BITBLAS format/inference support from Microsoft
• 🚀Sym=False Support. AutoGPTQ has unusable sym=false. (Re-quant required)
• 🚀lm_head module quant inference support for further VRAM reduction.
• 🚀 Faster quantization: More than 50% faster for TinyLlama + 4090 with batching and large calibration dataset.
• 🚀 Better quality quants as measured by PPL. (Test config: defaults + sym=True + FORMAT.GPTQ, TinyLlama)
• 🚀 Model weights sharding support
• 🚀 Security: hash check of model weights on load
• 🚀 Over 50% faster PPL calculations (OPT)
• 🚀 Over 40% faster packing stage in quantization (Llama 3.1 8B)
• ✨ Alert users of sub-optimal calibration data. Most new users get this part horribly wrong.
• ✨ Increased compatibility with newest models with auto-padding of in/out-features for [ Exllama, Exllama V2 ] backends.
• 👾 Removed non-working, partially working, or fully deprecated features: Peft, ROCM, AWQ Gemm inference, Triton v1 (replaced by v2), Fused Attention (Replaced by Marlin/Exllama).
• 👾 Fixed packing Performance regression on high core-count systems. Backported to AutoGPTQ
• 👾 Fixed crash on H100. Backported to AutoGPTQ
• ✨ 10s of thousands of lines of refactor/cleanup.
• ✨ Over 8+ overly complex api args removed/merged into simple human-readable args.
• ✨ Added CI workflow for validation of future PRs and prevent code regressions.
• ✨ Added perplexity unit-test to prevent against model quant quality regressions.
• 👾 De-bloated 271K lines of which 250K was caused by a single dataset used only by an example.
• 👾 De-bloat the number of args presented in public .from_quantized()/.from_pretrained() api
• ✨ Shorter and more concise public api/internal vars. No need to mimic HF style for verbose class names.
• ✨ Everything that did not pass unit-tests have been removed from repo.

Model Support ( 🚀 GPTQModel only )

Ready to deply quantized models

Model
Baichuan	✅	EXAONE 3.0	🚀	InternLM 1/2.5	🚀	OPT	✅
Bloom	✅	Falon	✅	Llama 1/2/3	✅	Phi/Phi-3	🚀
ChatGLM	🚀	Gemma 2	🚀	Llama 3.2 Vision	🚀	Qwen	✅
CodeGen	✅	GPTBigCod	✅	LongLLaMA	✅	Qwen2MoE	🚀
Cohere	✅	GPTNeoX	✅	MiniCPM3	🚀	RefinedWeb	✅
DBRX Converted	🚀	GPT-2	✅	Mistral	✅	StableLM	✅
Deci	✅	GPT-J	✅	Mixtral	✅	StarCoder2	✅
DeepSeek-V2	🚀	Granite	🚀	MOSS	✅	XVERSE	✅
DeepSeek-V2-Lite	🚀	GRIN-MoE	🚀	MPT	✅	Yi	✅

Compatiblity

We aim for 100% compatibility with models quanted by AutoGPTQ <= 0.7.1 and will consider syncing future compatibilty on a case-by-case basis.

Platform/GPU Requirements

GPTQModel is currently Linux only and requires CUDA capability >= 6.0 Nvidia GPU.

WSL on Windows should work as well.

ROCM/AMD support will be re-added in a future version after everything on ROCM has been validated. Only fully validated features will be re-added from the original AutoGPTQ repo.

Install

PIP

# You can install optional modules like autoround, qbits, vllm, sglang, or bitblas. Example: pip install -v --no-build-isolation gptqmodel[auto_round,vllm,sglang,bitblas,qbits]
pip install -v gptqmodel --no-build-isolation

Install from source

# clone repo
git clone https://github.com/ModelCloud/GPTQModel.git && cd GPTQModel

# pip: compile and install
# You can install optional modules like autoround, qbits, vllm, sglang, or bitblas. Example: pip install -v --no-build-isolation .[auto_round,vllm,sglang,bitblas,qbits]
pip install -v . --no-build-isolation

# uv: compile and install 
uv pip install -v . --no-build-isolation

Script installation

# You can pass modules as arguments, e.g., --vllm --sglang --bitblas. Example: bash install.sh --vllm --sglang --bitblas
bash install.sh

Quantization and Inference

warning: this is just a showcase of the usage of basic apis in GPTQModel, which uses only one sample to quantize a much small model, quality of quantized model using such little samples may not good.

Below is an example for the simplest use of gptqmodel to quantize a model and inference after quantization:

from transformers import AutoTokenizer
from gptqmodel import GPTQModel, QuantizeConfig

pretrained_model_dir = "facebook/opt-125m"
quant_output_dir = "opt-125m-4bit"

tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True)
calibration_dataset = [
    tokenizer(
        "The world is a wonderful place full of beauty and love."
    )
]

quant_config = QuantizeConfig(
    bits=4,  # 4-bit
    group_size=128,  # 128 is good balance between quality and performance
)

# load un-quantized model, by default, the model will always be loaded into CPU memory
model = GPTQModel.from_pretrained(pretrained_model_dir, quant_config)

# quantize model, the calibration_dataset should be list of dict whose keys can only be "input_ids" and "attention_mask"
model.quantize(calibration_dataset)

# save quantized model
model.save_quantized(quant_output_dir)

# load quantized model to the first GPU
model = GPTQModel.from_quantized(quant_output_dir)

# inference with model.generate
print(tokenizer.decode(model.generate(**tokenizer("gptqmodel is", return_tensors="pt").to(model.device))[0]))

For more advanced features of model quantization, please reference to this script

How to Add Support for a New Model

Read the gptqmodel/models/llama.py code which explains in detail via comments how the model support is defined. Use it as guide to PR for to new models. Most models follow the same pattern.

Evaluation on Downstream Tasks

You can use tasks defined in gptqmodel.eval_tasks to evaluate model's performance on specific down-stream task before and after quantization.

The predefined tasks support all causal-language-models implemented in 🤗 transformers and in this project.

Below is an example to evaluate `EleutherAI/gpt-j-6b` on sequence-classification task using `cardiffnlp/tweet_sentiment_multilingual` dataset:

from functools import partial

import datasets
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
from gptqmodel import GPTQModel, QuantizeConfig
from gptqmodel.eval_tasks import SequenceClassificationTask

MODEL = "EleutherAI/gpt-j-6b"
DATASET = "cardiffnlp/tweet_sentiment_multilingual"
TEMPLATE = "Question:What's the sentiment of the given text? Choices are {labels}.\nText: {text}\nAnswer:"
ID2LABEL = {
    0: "negative",
    1: "neutral",
    2: "positive"
}
LABELS = list(ID2LABEL.values())


def ds_refactor_fn(samples):
    text_data = samples["text"]
    label_data = samples["label"]

    new_samples = {"prompt": [], "label": []}
    for text, label in zip(text_data, label_data):
        prompt = TEMPLATE.format(labels=LABELS, text=text)
        new_samples["prompt"].append(prompt)
        new_samples["label"].append(ID2LABEL[label])

    return new_samples


#  model = AutoModelForCausalLM.from_pretrained(MODEL).eval().half().to("cuda:0")
model = GPTQModel.from_pretrained(MODEL, QuantizeConfig())
tokenizer = AutoTokenizer.from_pretrained(MODEL)

task = SequenceClassificationTask(
    model=model,
    tokenizer=tokenizer,
    classes=LABELS,
    data_name_or_path=DATASET,
    prompt_col_name="prompt",
    label_col_name="label",
    **{
        "num_samples": 1000,  # how many samples will be sampled to evaluation
        "sample_max_len": 1024,  # max tokens for each sample
        "block_max_len": 2048,  # max tokens for each data block
        # function to load dataset, one must only accept data_name_or_path as input
        # and return datasets.Dataset
        "load_fn": partial(datasets.load_dataset, name="english"),
        # function to preprocess dataset, which is used for datasets.Dataset.map,
        # must return Dict[str, list] with only two keys: [prompt_col_name, label_col_name]
        "preprocess_fn": ds_refactor_fn,
        # truncate label when sample's length exceed sample_max_len
        "truncate_prompt": False
    }
)

# note that max_new_tokens will be automatically specified internally based on given classes
print(task.run())

# self-consistency
print(
    task.run(
        generation_config=GenerationConfig(
            num_beams=3,
            num_return_sequences=3,
            do_sample=True
        )
    )
)

Learn More

tutorials provide step-by-step guidance to integrate gptqmodel with your own project and some best practice principles.

examples provide plenty of example scripts to use gptqmodel in different ways.

Supported Evaluation Tasks

Currently, gptqmodel supports: LanguageModelingTask, SequenceClassificationTask and TextSummarizationTask; more Tasks will come soon!

Which kernel is used by default?

GPTQModel will use Marlin, Exllama v2, Triton kernels in that order for maximum inference performance.

Acknowledgements

• Elias Frantar, Saleh Ashkboos, Torsten Hoefler and Dan Alistarh: for creating GPTQ and Marlin.
• PanQiWei: for creation of AutoGPTQ which this project code is based upon.
• FXMarty: for maintaining and support of AutoGPTQ.
• Qwopqwop200: for quantization code used in this project adapted from GPTQ-for-LLaMa.
• Turboderp: for releasing Exllama v1 and Exllama v2 kernels adapted for use in this project.
• FpgaMiner: for GPTQ-Triton kernels used in GPTQ-for-LLaMa which is adapted into this project.