Commits

History for

Commits on Aug 14, 2026

feat(stream): measure the VRAM fit instead of predicting it (#349)
MakazhanAlpamys
committed
36e9f71
docs: three claims that outlived their evidence

MakazhanAlpamys
and
claude
committed
9f57b7c
fix(errors): name the fix when Colab's preinstalled torchao breaks peft (#389)

MakazhanAlpamys
and
claude
committed
8c04757

Commits on Aug 13, 2026

fix(stream): the panel blamed Windows on a Linux box

MakazhanAlpamys
and
claude
committed
213fff2
fix(gpu): "bf16 supported" includes EMULATION, so a T4 said yes and got bf16

MakazhanAlpamys
and
claude
committed
c5f80a5
feat(stream): let training.stream_vram_override replace the measured free VRAM (#386)
AmirF194
authored
d31c1e5
fix(gpu): a "cuda" string with no CUDA runtime asks for no mixed precision

MakazhanAlpamys
and
claude
committed
a3f88a1
fix(trainer): ask the card whether it has bf16, in all fourteen places (#385, #387)

MakazhanAlpamys
and
claude
committed
5793e83
fix(stream): canonicalise named_parameters() for streamed models (#384)
AmirF194
authored
6116c81

Commits on Aug 12, 2026

fix(sft): normalise every checkpoint-* adapter under torch.compile, not only the final save (#380)
wilyan09007
authored
55f583b

Commits on Aug 11, 2026

fix(trainer): thread training.seed into every task wrapper, and apply it before the adapter (#381)
wilyan09007
authored
e5e36d6

Commits on Aug 7, 2026

Commits on Aug 6, 2026

Commits on Aug 4, 2026

Commits on Aug 3, 2026

feat(streaming): preference losses over layer streaming (v0.72.4)
MakazhanAlpamys
committed
4fb2540

Commits on Jul 28, 2026

Commits on Jul 27, 2026

fix(train): streamed adapters were saved unloadable (v0.72.1)
MakazhanAlpamys
committed
af8683b

Commits on Jul 26, 2026

feat(train): layer streaming — fine-tune models larger than VRAM (v0.72.0 BETA)
MakazhanAlpamys
committed
00833ac

Commits on Jul 19, 2026