Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
docs(contributors): #1952 [skip ci]
Alpamys340061b8de2b
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.mdindex a3119769..6a276d88 100644--- a/CONTRIBUTORS.md+++ b/CONTRIBUTORS.md@@ -1006,6 +1006,7 @@ Listed by first contribution. PR numbers link the work.- **Shivansh Shukla** ([@ShivanshShukla](https://github.com/ShivanshShukla))- First contribution. Made `soup can run` remove the temporary directory it extracted a can into when env capture, deploy or the training subprocess fails, and check the `--env-capture` path before that directory is created; a caller-supplied `--extract-dir` is never removed ([#1464](https://github.com/MakazhanAlpamys/Soup/issues/1464), [#1960](https://github.com/MakazhanAlpamys/Soup/pull/1960))+ - Made the Web UI delete the temporary `soup_ui_config_*.yaml` it writes for every `/api/train/start` once the training process exits, is stopped, or fails to start, and gave the raw file descriptor a single owner so a reused descriptor number is never closed by mistake ([#1464](https://github.com/MakazhanAlpamys/Soup/issues/1464), [#1952](https://github.com/MakazhanAlpamys/Soup/pull/1952))- **dryn85-max** ([@dryn85-max](https://github.com/dryn85-max))- First contribution. Added native DPO on `backend: mlx` with Soup's own training loop (sigmoid loss, reference log-probs precomputed once from the base model, a chunked vocabulary head), reading rows through the same `format: dpo` converter as the transformers backend ([#1846](https://github.com/MakazhanAlpamys/Soup/pull/1846))