Commits

History for

Commits on Sep 7, 2026

fix(trainer): build a PEFT LoRA+ optimizer instead of forwarding loraplus_lr_ratio to TrainingArguments (#738)
abdulwaarith0
authored
d7aea47

Commits on Sep 6, 2026

fix(trainer): put SFT packing on SFTConfig for TRL 0.29 (#709)

jagadeepmamidi
and
claude
authored
991ab8f

Commits on Aug 27, 2026

Commits on Aug 26, 2026

fix(train): enable BF16 on validated MPS trainers (#564)
Amix29
authored
4b2947b

Commits on Aug 25, 2026

fix(sft): reject rows without causal targets (#535)
Amix29
authored
aefc2cb
feat(trainer): extend LISA to continued pre-training (#476)
ousamabenyounes
authored
2794043

Commits on Aug 24, 2026

Commits on Aug 23, 2026

fix(sft): collate SmolVLM vision batches end to end (#302)
Amix29
authored
664f5c9

Commits on Aug 19, 2026

fix(cce): resolve cut_ce architecture from config, not a name substring (#383)
AmirF194
authored
5705a6c

Commits on Aug 13, 2026

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

MakazhanAlpamys
and
claude
committed
5793e83

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 8, 2026

feat(train): #341 #340 - a real training seed, and full fine-tuning as lora.r=0
MakazhanAlpamys
committed
dcb5eb5

Commits on Aug 7, 2026

Commits on Aug 6, 2026

fix(trainers): device_map must pin one GPU per rank under a distributed launch
MakazhanAlpamys
committed
aa8e418

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

fix(sft): mirror vision-processor pad_token from nested tokenizer (#302)
MakazhanAlpamys
committed
481dc38

Commits on Jul 17, 2026

fix(cli): quote install hints so `pip install soup-cli[extra]` works on cmd.exe (v0.71.37)

MakazhanAlpamys
and
claude
committed
a95fede

Commits on Jul 15, 2026

feat(lisa): wire LISA into SFT trainer + attach helper (v0.71.34)
MakazhanAlpamys
committed
2948c1e

Commits on Jul 4, 2026

fix: reuse shared vocabulary expansion in vision/audio SFT (#291)
CODING-DARSH
authored
23da524

Commits on Jul 3, 2026

fix(sft): apply configured vocabulary expansion (#287)
CODING-DARSH
authored
abf8fef

Commits on Jul 2, 2026

fix: remediate the HIGH-tier code-review findings

MakazhanAlpamys
and
claude
committed
58d280c
fix: remediate 6 CRITICAL code-review findings

MakazhanAlpamys
and
claude
committed
b587dee

Commits on Jun 12, 2026

feat(train): native Spectrum targeted training — soup spectrum scan + training.unfrozen_parameters (v0.71.23)
MakazhanAlpamys
committed
fcf4b33

Commits on Jun 10, 2026

feat(trainer): live TTS / BitNet / MoE-expert-quant trainers (v0.71.20)
MakazhanAlpamys
committed
a4dfbb3

Commits on Jun 9, 2026

feat(quant,multipack): vision/audio Quant Menu + multipack FSDP sharding (v0.71.19)
MakazhanAlpamys
committed
f51331d