pytorch - 💡(How to fix) Fix deprecation of derived quantized dtypes (quint8, qint8, qint32) [1 pull requests]

We are deprecating the quantized tensor dtypes (torch.quint8, torch.qint8, torch.qint32) and the functions that produce them (torch.quantize_per_tensor, torch.quantize_per_channel, etc.).

Motivation

The quantized dtypes bundle scale, zero_point, and quantization scheme metadata directly into the tensor. This was an early design choice that has not scaled well — it creates a parallel type system within PyTorch that requires dedicated operator implementations for every quantized op, limits composability with the rest of the ecosystem, and makes it difficult to support new quantization schemes.

The expected path forward is to use basic dtypes from core (e.g. torch.int8, torch.uint8, torch.float8_e4m3fn) and handle how scale, zero_point, and other quantization parameters are calculated and tracked in the modeling code. This approach gives users full control over quantization scheme and parameter management, without requiring modifications to core.

For reference implementations of this pattern, see https://github.com/pytorch/ao.