Chenhao Xue

We're still working on it, and it'll be released along with quantized models' pth files later. If you only need quantized weights and bias, there's a way for quick start....

Currently that's not available, sorry for that. Some slight modification could get things work, if: 1. You have calibration dataset for the model and you know how to preprocess the...

我们这篇工作主要考虑Vision transformer，所以没有实现conv3D算子的量化和封装。ViT的embedding操作是使用conv2D实现的，这部分代码在`layers.conv`里，您可以考虑仿照conv2D实现conv3D，从代码的核心逻辑上二者没有什么区别，只要把相关操作的维度增加一下就好了。conv2d当中每一步操作的维度如何变换，我在注释中逐行写出来了，可以用作参考。如果您已经实现了conv3D算子，您还要在`utils.net_wrap`里加上对conv3D的封装操作。

The quantized models only contains NN parameters in int8, which cannot be loaded and put into use directly. I would suggest you follow the given example files and quantize a...

I've only taken notice of the issues recently. I'll start working on a more user-friendly interface ASAP, so that you don't have to quantize a model from scratch .

We import our ViT model from package `timm`, and this is how they store their weight tensor. Indeed, W_Q, W_K and W_V should be [2304, 768], but `timm` fuses the...

PTQ4ViT adopts two classes `SoSMatMul` and `PostGeluLinear` in `quant_layers/`. `SoSMatMul` adopts a variable `split` to quantize post-softmax values. It indicates the split point of two ranges [0, split] and [split,...

A_interval_candidates should be initialized with 2 ** m * initial a_neg_interval. Looks like a mistake pop up when merging the released version code, original code for experiment is kind of...