COMMITS
/ torchao/float8/float8_utils.py February 22, 2025
V
delete delayed scaling from torchao.float8 (#1753)
Vasiliy Kuznetsov committed
February 10, 2025
D
Support power of 2 scaling factors in float8 training and use e4m3 everywhere (#1670)
Daniel Vega-Myhre committed
December 19, 2024
V
refactor: move delayed and static scaling to StatefulFloat8Linear (#1439)
Vasiliy Kuznetsov committed
December 4, 2024
L
[float8] Allow specifying arbitrary dtype for each tensor (#1378)
Luca Wehrstedt committed
November 25, 2024
V
fix torchtitan + float8 + delayed + compile (#1334)
Vasiliy Kuznetsov committed
November 22, 2024
V
float8 training: fix bug with AC + compile (#1329)
Vasiliy Kuznetsov committed
October 23, 2024
J
[reland][ROCm] use dataclass for fnuz type setting (#1150)
Jeff Daily committed
M
Revert "[ROCm] use dataclass for fnuz type setting" (#1148)
Mark Saroufim committed
J
[ROCm] use dataclass for fnuz type setting (#1142)
Jeff Daily committed
October 7, 2024
V
add axiswise granularity to Float8Tensor (#919)
Vasiliy Kuznetsov committed
October 1, 2024
W
[float8] improve eager numerics for dynamic scales and gets on par with torch.compile (#904)
Wei (Will) Feng committed
J
add ruff lint rule to remove unused imports via ruff (#969)
Jiayu Liu committed
September 26, 2024
W
[float8] all-reduce amax on dp mesh instead of global pg (#933)
Wei (Will) Feng committed
September 25, 2024
S
Remove two if statements in fp8 padding (#935)
Shuqi Yang committed
August 23, 2024
S
Fix docstring args names (#735)
Sergii Dymchenko committed
August 7, 2024
V
float8: remove unneeded kernel for scale generation (#616)
Vasiliy Kuznetsov committed
July 30, 2024
V
move float8_experimental to torchao/float8
Vasiliy Kuznetsov committed