PyTorch 官方的量化与稀疏化工具库，支持训练和推理阶段的低精度计算，覆盖 float8、MX 等数据类型，可加速 Transformer 和 Llama 等大模型。
…nd apply adascale (00c6089, c548be0) Enable FP8 ( float8e4m3fn , float8e5m2 ) quantization simulation on both CPU and CUDA, completing FP8 support alongside ONNX QDQ expo
…b.com/pytorch/ao/pull/4205 Rename Sparse2x4CUTLASSFloat8Tensor to Float8Sparse2x4 2DData2DMetadataTensor ( 4343) by @bbeckca in https://github.com/pytorch/ao/pull/4343 Mo
…tra-node networking. Deprecations Deprecate v1 of Float8WeightOnlyConfig , Float8DynamicActivationFloat8WeightConfig , Int8DynamicActivationIntxWeightConfig , IntxWeightO
…e currently support the following stable configs: Float8DynamicActivationFloat8WeightConfig Int4WeightOnlyConfig IntxWeightOnlyConfig Int8DynamicActivationIntxWeightConfi
…htOnlyConfig()) Full list of deprecated functions float8 dynamic activation float8 weight float8 static activation float8 weight float8 weight only fpx weight only gemlit
…ging Face hub 263 by @sayakpaul - Add support for float8 e4f3mnuz 310 (from 281) by @maktukmak - Faster and less memory-intensive requantization 290 by @latentCall145 - S