Re: #3290 FP8 Blockwise Training Tracker, quantization benchmarks#3306

Open

agolajko wants to merge 19 commits intopytorch:mainfrom

agolajko:feat/fp8-quant-bench-3290

Commits on Nov 6, 2025

benchmark for triton_fp8_blockwise_act_quant_transposed_lhs against naive torch implementation
agolajko
committed
benchmark for triton_fp8_blockwise_act_quant_lhs against naive implementation torch_blockwise_scale_act_quant_lhs from existing blockwise_fp8_training/kernels
agolajko
committed
benchmark for triton_fp8_blockwise_act_quant_rhs against naive implementation
agolajko
committed
bench for triton_fp8_blockwise_weight_quant_rhs against naive torch implementation
agolajko
committed
bench for triton_fp8_blockwise_weight_quant_transposed_rhs against naive torch
agolajko
committed
removed extra space from file name
agolajko
committed
Flipped mem layout of scales to streamline the LHS activation quantization
agolajko
committed

Commits on Nov 7, 2025

Commits on Nov 9, 2025

Fixes to comments
agolajko
committed