List view

Enable blockwise FP8 dense training kernels on ROCm

CLA SignedThis label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.

float8

topic: rocm

Status: Open (in progress).

pytorch/aonumber 4036

#4036 In pytorch/ao;

· brucechanglongxu opened on Mar 10, 2026

Blockwise FP8 training seems to have much lower peak memory usage than others

float8

module: trainingquantize_ api training flowquantize_ api training flow

Status: Open.

#4101 In pytorch/ao;

· CFC87 opened on Mar 17, 2026

fp8 blockwise test failure on ROCM in CI

module: trainingquantize_ api training flowquantize_ api training flow

topic: rocm

Status: Open.

Bug

#4104 In pytorch/ao;

· danielvegamyhre opened on Mar 18, 2026