Skip to content

armgit / Soup

Public Imported from GitHub
0

Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.

docs(contributors): #1855 [skip ci]

Alpamys8e1551e39afb
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
index fb332610..849c5e1a 100644
--- a/CONTRIBUTORS.md
+++ b/CONTRIBUTORS.md
@@ -1038,6 +1038,7 @@ Listed by first contribution. PR numbers link the work.
- Made `backend: mlx` honour `training.eval_steps` for SFT, counted in completed optimizer updates (Soup runs the evaluation itself, over every validation row, once more after the last update) ([#1990](https://github.com/MakazhanAlpamys/Soup/issues/1990), [#2074](https://github.com/MakazhanAlpamys/Soup/pull/2074))
- Added native SimPO on `backend: mlx` (TRL 0.29.1's CPO `simpo` objective with the chosen-completion NLL, reference-free, beta fixed at TRL's default 0.1, TRL-identical tokens on rows that fit), sharing the MLX preference loop and ORPO's resume guard ([#1999](https://github.com/MakazhanAlpamys/Soup/issues/1999), [#2097](https://github.com/MakazhanAlpamys/Soup/pull/2097))
- Capped the prompt on the transformers SimPO path at `data.max_length // 2` (keeping its end) after TRL tokenizes, as DPO and ORPO do, and corrected the issue's own claim about the #1208 refusal against the TRL source, with tests on the real setup compared against TRL's tokenizer ([#2112](https://github.com/MakazhanAlpamys/Soup/issues/2112), [#2116](https://github.com/MakazhanAlpamys/Soup/pull/2116))
+ - MLX SFT runs are now reproducible: `training.seed` seeds MLX's own RNG before the adapter is built and `training.data_seed` (or the seed) seeds the numpy RNG behind mlx-lm's batch shuffle right before training, `0` counts as a seed, and the "MLX backend ignores" warning is gone; after review they added a portable test that pins the seeds and their order without Apple Silicon and moved the no-warning check into a file that runs without mlx ([#1850](https://github.com/MakazhanAlpamys/Soup/issues/1850), [#1855](https://github.com/MakazhanAlpamys/Soup/pull/1855))
- **Tanmay Garg** ([@Tanmay9223](https://github.com/Tanmay9223))
- First contribution. Moved the PRM, MoLE routing and PPO wrappers onto the shared warmup resolution, and made PPO honour `--deepspeed`, take `response_length` from `data.max_length` and keep 3 saves ([#805](https://github.com/MakazhanAlpamys/Soup/issues/805), [#1775](https://github.com/MakazhanAlpamys/Soup/pull/1775))