PyTorch-compatible backward API - #7665

Merged

tohtana merged 39 commits into

deepspeedai:masterfrom

tohtana:tohtana/backward_non_scalar

Nov 19, 2025

Commits on Oct 17, 2025

rename backward prologue method
tohtana
committed
8a27283

Commits on Oct 18, 2025

refactor loss scaling
tohtana
committed
d85cfd9
refactor backward
tohtana
committed
bded5c8
fix for bf16 optimizer
tohtana
committed
1f413d6
simplify preprocess/postprocess of backward
tohtana
committed
cc87977

Commits on Oct 19, 2025

fix order of backward postprocess
tohtana
committed
95018a3
enable non-scalar backward only for ZeROOptimizer
tohtana
committed
80d0e7d
fix zero+fp16 case
tohtana
committed
db70476
add config to enable allow_user_backward
tohtana
committed
50b29d8
fix flag for error handling
tohtana
committed
076b187

Commits on Nov 3, 2025

resolve conflict
tohtana
committed
f6748d1

Commits on Nov 7, 2025

add test cases
tohtana
committed
280b1fa
Merge branch 'master' into tohtana/backward_non_scalar
tohtana
authored
5d5e64e
fix format
tohtana
committed
6ce26f3
return scaled loss from engine's backward
tohtana
committed
0c579d5

Commits on Nov 10, 2025

Merge branch 'master' into tohtana/backward_non_scalar
tohtana
authored
1615036

Commits on Nov 11, 2025

remove option to enable user backward
tohtana
committed
c8758f7

Commits on Nov 12, 2025

add hook utility
tohtana
committed
9962f2c
fix for z2
tohtana
committed
a8f15a0
fix scaling
tohtana
committed
1d0a721

Commits on Nov 13, 2025

exclude unused params from counter
tohtana
committed
39372ac
set default flag
tohtana
committed
7eacbc7
handle non-zero optimizer
tohtana
committed
6cce937
call epilogue in engine's backward
tohtana
committed
b72b5a7
prevent hooks from being called from nested backward
tohtana
committed
1307a87
run post hook fo rz3
tohtana
committed
98cc865
Merge branch 'master' into tohtana/backward_non_scalar
tohtana
authored
adb6990
added comments
tohtana
committed
9328dfa
remove hard-coded tolerances
tohtana
committed
01b3251
add test for multiple engines
tohtana
committed
78f7ad4

Commits on Nov 14, 2025

update document
tohtana
committed
73f7ff1

Commits on Nov 17, 2025

remove deprecated comment
tohtana
committed
26308cd
simplify utility func to count effective grad nodes
tohtana
committed
9963546
fix combination with leaf module
tohtana
committed
b730f46
refactor tests
tohtana
committed
08b1599

Commits on Nov 18, 2025

refactor tests
tohtana
committed
92d3068
fix loss scaling
tohtana
committed
ebac40b
Merge branch 'master' into tohtana/backward_non_scalar
tohtana
authored
fcf7c8c
Merge branch 'master' into tohtana/backward_non_scalar
tohtana
authored
90e1b7d