Quanto adds PyTorch quantization backend for Optimum — int8/int4/float8 support

AI Impact Summary

Quanto introduces a PyTorch quantization backend for Optimum that lets you quantize weights and activations to int8/int4/float8 across CPU, CUDA, and MPS, cutting memory usage and enabling edge deployment. It plugs into the Transformers workflow via QuantoConfig and provides a full quantization lifecycle (quantize, calibrate, tune, freeze) with serialization through safetensors. The docs illustrate usage with models like facebook/opt-125m and openai/whisper-large-v3, and reference larger targets such as meta-llama/Meta-Llama-3.1-8B, signaling broad applicability across LLMs. Expect memory and latency benefits, but plan for validation of accuracy depending on the chosen quantization scheme.