Skip to content

Commit

Permalink
quant pt 4
Browse files Browse the repository at this point in the history
  • Loading branch information
@stevhliu
stevhliu committed Dec 21, 2024
1 parent ed42966 commit 2670156
Show file tree
Hide file tree
Showing 2 changed files with 24 additions and 61 deletions.
2 changes: 2 additions & 0 deletions docs/source/en/_toctree.yml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -182,6 +182,8 @@
title: Quanto
- local: quantization/torchao
title: torchao
- local: quantization/vptq
title: VPTQ
- local: quantization/contribute
title: Contribute
- title: Deploy to production
Expand Down
83 changes: 22 additions & 61 deletions docs/source/en/quantization/vptq.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,53 +14,47 @@ rendered properly in your Markdown viewer.
-->

# VPTQ
# VPTQ

> [!TIP]
> Try VPTQ on [Hugging Face](https://huggingface.co/spaces/microsoft/VPTQ)!
> Try VPTQ on [Google Colab](https://colab.research.google.com/github/microsoft/VPTQ/blob/main/notebooks/vptq_example.ipynb)!
> Know more about VPTQ on [ArXiv](https://arxiv.org/pdf/2409.17066)!
[Vector Post-Training Quantization (VPTQ)](https://github.com/microsoft/VPTQ) is a Post-Training Quantization (PTQ) method that leverages vector quantization to quantize LLMs at an extremely low bit-width (<2-bit). VPTQ can compress a 70B, even a 405B model, to 1-2 bits without retraining and still maintain a high-degree of accuracy. It is a lightweight quantization algorithm that takes ~17 hours to quantize a 405B model. VPTQ features agile quantization inference with low decoding overhead and high throughput and Time To First Token (TTFT).

Vector Post-Training Quantization ([VPTQ](https://github.com/microsoft/VPTQ)) is a novel Post-Training Quantization method that leverages Vector Quantization to high accuracy on LLMs at an extremely low bit-width (<2-bit). VPTQ can compress 70B, even the 405B model, to 1-2 bits without retraining and maintain high accuracy.
Run the command below to install VPTQ which provides efficient kernels for inference on NVIDIA and AMD GPUs.

- Better Accuracy on 1-2 bits, (405B @ <2bit, 70B @ 2bit)
- Lightweight Quantization Algorithm: only cost ~17 hours to quantize 405B Llama-3.1
- Agile Quantization Inference: low decode overhead, best throughput, and TTFT

Inference support for VPTQ is released in the `vptq` library. Make sure to install it to run the models:
```bash
pip install vptq
```

The library provides efficient kernels for NVIDIA/AMD GPU inference.
The [VPTQ-community](https://huggingface.co/VPTQ-community) provides a collection of VPTQ-quantized models. The model name contains information about its bitwidth (excluding cookbook, parameter, and padding overhead). Consider the [Meta-Llama-3.1-70B-Instruct-v8-k65536-256-woft] model as an example.

- The model name is Meta-Llama-3.1-70B-Instruct.
- The number of centroids is given by 65536 (2^16).
- The number of residual centroids is given by 256 (2^8).

To run VPTQ models simply load a model that has been quantized with VPTQ:
The equivalent bit-width calculation is given by the following.

## Inference example
**Run Llama 3.1 70b on RTX4090 (24G @ ~2bits) in real time**
![Llama3 1-70b-prompt](https://github.com/user-attachments/assets/d8729aca-4e1d-4fe1-ac71-c14da4bdd97f)
- index: log2(65536) = 16 / 8 = 2-bits
- residual index: log2(256) = 8 / 8 = 1-bit
- total bit-width: 2 + 1 = 3-bits

From here, estimate the model size by multiplying 70B * 3-bits / 8-bits/byte for a total of 26.25GB.

```python
Load a VPTQ quantized model with [`~PreTrainedModel.from_pretrained`].

```py
from transformers import AutoTokenizer, AutoModelForCausalLM

quantized_model = AutoModelForCausalLM.from_pretrained(
"VPTQ-community/Meta-Llama-3.1-70B-Instruct-v16-k65536-65536-woft",
torch_dtype="auto",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("VPTQ-community/Meta-Llama-3.1-70B-Instruct-v16-k65536-65536-woft")
input_ids = tokenizer("hello, it's me", return_tensors="pt").to("cuda")
out = model.generate(**input_ids, max_new_tokens=32, do_sample=False)
```

## Quantize your own model
VPTQ algorithm early-released at [VPTQ ](https://github.com/microsoft/VPTQ/tree/algorithm),
and checkout the [tutorial](https://github.com/microsoft/VPTQ/blob/algorithm/algorithm.md).
To quantize your own model, refer to the [VPTQ Quantization Algorithm Tutorial](https://github.com/microsoft/VPTQ/blob/algorithm/algorithm.md) tutorial.

## Early Results from Tech Report
VPTQ achieves better accuracy and higher throughput with lower quantization overhead across models of different sizes. The following experimental results are for reference only; VPTQ can achieve better outcomes under reasonable parameters, especially in terms of model accuracy and inference speed.
## Benchmarks

VPTQ achieves better accuracy and higher throughput with lower quantization overhead across models of different sizes. The following experimental results are for reference only; VPTQ can achieve better outcomes under reasonable parameters, especially in terms of model accuracy and inference speed.

| Model | bitwidth | W2↓ | C4↓ | AvgQA↑ | tok/s↑ | mem(GB) | cost/h↓ |
| ----------- | -------- | ---- | ---- | ------ | ------ | ------- | ------- |
Expand All @@ -71,41 +65,8 @@ VPTQ achieves better accuracy and higher throughput with lower quantization over
| LLaMA-2 70B | 2.07 | 3.93 | 5.72 | 68.6 | 9.7 | 19.54 | 19 |
| | 2.11 | 3.92 | 5.71 | 68.7 | 9.7 | 20.01 | 19 |

## Resources

See an example demo of VPTQ on the VPTQ Online Demo [Space](https://huggingface.co/spaces/microsoft/VPTQ) or try running the VPTQ inference [notebook](https://colab.research.google.com/github/microsoft/VPTQ/blob/main/notebooks/vptq_example.ipynb).

## More Models in [VPTQ-community](https://huggingface.co/VPTQ-community)

⚠️ The repository only provides a method of model quantization algorithm.

⚠️ The open-source community VPTQ-community provides models based on the technical report and quantization algorithm.



**Quick Estimation of Model Bitwidth (Excluding Codebook Overhead)**:

- **Model Naming Convention**: The model's name includes the **vector length** $v$, **codebook (lookup table) size**, and **residual codebook size**. For example, "Meta-Llama-3.1-70B-Instruct-v8-k65536-256-woft" is "Meta-Llama-3.1-70B-Instruct", where:
- **Vector Length**: 8
- **Number of Centroids**: 65536 (2^16)
- **Number of Residual Centroids**: 256 (2^8)
- **Equivalent Bitwidth Calculation**:
- **Index**: log2(65536) = 16 / 8 = 2 bits
- **Residual Index**: log2(256) = 8 / 8 = 1 bit
- **Total Bitwidth**: 2 + 1 = 3 bits
- **Model Size Estimation**: 70B * 3 bits / 8 bits per Byte = 26.25 GB

- **Note**: This estimate does not include the size of the codebook (lookup table), other parameter overheads, and the padding overhead for storing indices. For the detailed calculation method, please refer to **Tech Report Appendix C.2**.


| Model Series | Collections | (Estimated) Bit per weight |
| :--------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------: | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Llama 3.1 Nemotron 70B Instruct HF | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-llama-31-nemotron-70b-instruct-hf-without-finetune-671730b96f16208d0b3fe942) | [4 bits](https://huggingface.co/VPTQ-community/Llama-3.1-Nemotron-70B-Instruct-HF-v8-k65536-65536-woft) [3 bits](https://huggingface.co/VPTQ-community/Llama-3.1-Nemotron-70B-Instruct-HF-v8-k65536-256-woft) [2 bits (1)](https://huggingface.co/VPTQ-community/Llama-3.1-Nemotron-70B-Instruct-HF-v16-k65536-65536-woft) [2 bits (2)](https://huggingface.co/VPTQ-community/Llama-3.1-Nemotron-70B-Instruct-HF-v8-k65536-0-woft) [1.875 bits](https://huggingface.co/VPTQ-community/Llama-3.1-Nemotron-70B-Instruct-HF-v16-k65536-16384-woft) [1.625 bits](https://huggingface.co/VPTQ-community/Llama-3.1-Nemotron-70B-Instruct-HF-v16-k65536-1024-woft) [1.5 bits](https://huggingface.co/VPTQ-community/Llama-3.1-Nemotron-70B-Instruct-HF-v16-k65536-256-woft) |
| Llama 3.1 8B Instruct | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-llama-31-8b-instruct-without-finetune-66f2b70b1d002ceedef02d2e) | [4 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-8B-Instruct-v8-k65536-65536-woft) [3.5 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-8B-Instruct-v8-k65536-4096-woft) [3 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-8B-Instruct-v8-k65536-256-woft) [2.3 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-8B-Instruct-v12-k65536-4096-woft) |
| Llama 3.1 70B Instruct | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-llama-31-70b-instruct-without-finetune-66f2bf454d3dd78dfee2ff11) | [4 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-70B-Instruct-v8-k65536-65536-woft) [3 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-70B-Instruct-v8-k65536-256-woft) [2.25 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-70B-Instruct-v8-k65536-4-woft) [2 bits (1)](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-70B-Instruct-v16-k65536-65536-woft) [2 bits (2)](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-70B-Instruct-v8-k65536-0-woft) [1.93 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-70B-Instruct-v16-k65536-32768-woft) [1.875 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-70B-Instruct-v8-k32768-0-woft) [1.75 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-70B-Instruct-v8-k16384-0-woft) |
| Llama 3.1 405B Instruct | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-llama-31-405b-instruct-without-finetune-66f4413f9ba55e1a9e52cfb0) | [4 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v8-k65536-65536-woft) [3 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v8-k65536-256-woft) [2 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v16-k65536-65536-woft) [1.875 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v16-k32768-32768-woft) [1.625 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v16-k65536-1024-woft) [1.5 bits (1)](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v8-k4096-0-woft) [1.5 bits (2)](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v16-k65536-256-woft) [1.43 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v16-k65536-128-woft) [1.375 bits](https://huggingface.co/VPTQ-community/Meta-Llama-3.1-405B-Instruct-v16-k65536-64-woft) |
| Mistral Large Instruct 2407 (123B) | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-mistral-large-instruct-2407-without-finetune-6711ebfb7faf85eed9cceb16) | [4 bits](https://huggingface.co/VPTQ-community/Mistral-Large-Instruct-2407-v8-k65536-65536-woft) [3 bits](https://huggingface.co/VPTQ-community/Mistral-Large-Instruct-2407-v8-k65536-256-woft) [2 bits (1)](https://huggingface.co/VPTQ-community/Mistral-Large-Instruct-2407-v16-k65536-65536-woft) [2 bits (2)](https://huggingface.co/VPTQ-community/Mistral-Large-Instruct-2407-v8-k65536-0-woft) [1.875 bits](https://huggingface.co/VPTQ-community/Mistral-Large-Instruct-2407-v16-k65536-16384-woft) [1.75 bits](https://huggingface.co/VPTQ-community/Mistral-Large-Instruct-2407-v16-k65536-4096-woft) [1.625 bits](https://huggingface.co/VPTQ-community/Mistral-Large-Instruct-2407-v16-k65536-1024-woft) [1.5 bits](https://huggingface.co/VPTQ-community/Mistral-Large-Instruct-2407-v16-k65536-256-woft) |
| Qwen 2.5 7B Instruct | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-qwen-25-7b-instruct-without-finetune-66f3e9866d3167cc05ce954a) | [4 bits](https://huggingface.co/VPTQ-community/Qwen2.5-7B-Instruct-v8-k65536-65536-woft) [3 bits](https://huggingface.co/VPTQ-community/Qwen2.5-7B-Instruct-v8-k65536-256-woft) [2 bits (1)](https://huggingface.co/VPTQ-community/Qwen2.5-7B-Instruct-v8-k256-256-woft) [2 bits (2)](https://huggingface.co/VPTQ-community/Qwen2.5-7B-Instruct-v8-k65536-0-woft) [2 bits (3)](https://huggingface.co/VPTQ-community/Qwen2.5-7B-Instruct-v16-k65536-65536-woft) |
| Qwen 2.5 14B Instruct | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-qwen-25-14b-instruct-without-finetune-66f827f83c7ffa7931b8376c) | [4 bits](https://huggingface.co/VPTQ-community/Qwen2.5-14B-Instruct-v8-k65536-65536-woft) [3 bits](https://huggingface.co/VPTQ-community/Qwen2.5-14B-Instruct-v8-k65536-256-woft) [2 bits (1)](https://huggingface.co/VPTQ-community/Qwen2.5-14B-Instruct-v8-k256-256-woft) [2 bits (2)](https://huggingface.co/VPTQ-community/Qwen2.5-14B-Instruct-v8-k65536-0-woft) [2 bits (3)](https://huggingface.co/VPTQ-community/Qwen2.5-14B-Instruct-v16-k65536-65536-woft) |
| Qwen 2.5 32B Instruct | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-qwen-25-32b-instruct-without-finetune-66fe77173bf7d64139f0f613) | [4 bits](https://huggingface.co/VPTQ-community/Qwen2.5-32B-Instruct-v8-k65536-65536-woft) [3 bits](https://huggingface.co/VPTQ-community/Qwen2.5-32B-Instruct-v8-k65536-256-woft) [2 bits (1)](https://huggingface.co/VPTQ-community/Qwen2.5-32B-Instruct-v16-k65536-65536-woft) [2 bits (2)](https://huggingface.co/VPTQ-community/Qwen2.5-32B-Instruct-v8-k65536-0-woft) [2 bits (3)](https://huggingface.co/VPTQ-community/Qwen2.5-32B-Instruct-v8-k256-256-woft) |
| Qwen 2.5 72B Instruct | [HF 🤗](https://huggingface.co/collections/VPTQ-community/vptq-qwen-25-72b-instruct-without-finetune-66f3bf1b3757dfa1ecb481c0) | [4 bits](https://huggingface.co/VPTQ-community/Qwen2.5-72B-Instruct-v8-k65536-65536-woft) [3 bits](https://huggingface.co/VPTQ-community/Qwen2.5-72B-Instruct-v8-k65536-256-woft) [2.38 bits](https://huggingface.co/VPTQ-community/Qwen2.5-72B-Instruct-v8-k1024-512-woft) [2.25 bits (1)](https://huggingface.co/VPTQ-community/Qwen2.5-72B-Instruct-v8-k512-512-woft) [2.25 bits (2)](https://huggingface.co/VPTQ-community/Qwen2.5-72B-Instruct-v8-k65536-4-woft) [2 bits (1)](https://huggingface.co/VPTQ-community/Qwen2.5-72B-Instruct-v8-k65536-0-woft) [2 bits (2)](https://huggingface.co/VPTQ-community/Qwen2.5-72B-Instruct-v16-k65536-65536-woft) [1.94 bits](https://huggingface.co/VPTQ-community/Qwen2.5-72B-Instruct-v16-k65536-32768-woft) |
| Reproduced from the tech report | [HF 🤗](https://huggingface.co/collections/VPTQ-community/reproduced-vptq-tech-report-baseline-66fbf1dffe741cc9e93ecf04) | Results from the open source community for reference only, please use them responsibly. |
| Hessian and Inverse Hessian Matrix | [HF 🤗](https://huggingface.co/collections/VPTQ-community/hessian-and-invhessian-checkpoints-66fd249a104850d17b23fd8b) | Collected from RedPajama-Data-1T-Sample, following [Quip#](https://github.com/Cornell-RelaxML/quip-sharp/blob/main/quantize_llama/hessian_offline_llama.py)
For more information, read the VPTQ [paper](https://arxiv.org/pdf/2409.17066).

0 comments on commit 2670156

Please sign in to comment.