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): #2052 [skip ci]
Alpamys6f3a7ca72f5d
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.mdindex 9e0a34e7..a3119769 100644--- a/CONTRIBUTORS.md+++ b/CONTRIBUTORS.md@@ -893,6 +893,7 @@ Listed by first contribution. PR numbers link the work.- Every `SOUP_LAYER_STREAM_STRIPE_DIRS` refusal printed the entry with `repr`, so a Windows path came out with every backslash doubled and could not be copied back. They added `quote_path()`, which keeps a path as typed inside single quotes and escapes exactly the control characters the validation refuses, and used it at every site in the issue's table, including four Windows API errors that only fire on Windows. **They listed the spellings first (drive letter, UNC, spaces, trailing separator, POSIX, relative) and tested each one.** After review they added the changelog fragment, moved the helper to a neutral module and added a UNC case. 22 new tests (the file does not collect on `main` without the helper), 11 of 11 mutations killed ([#1611](https://github.com/MakazhanAlpamys/Soup/issues/1611), [#1626](https://github.com/MakazhanAlpamys/Soup/pull/1626))- Improved loopback Web UI host validation and IPv6 controls ([#1829](https://github.com/MakazhanAlpamys/Soup/issues/1829), [#1839](https://github.com/MakazhanAlpamys/Soup/pull/1839))- `backend: mlx` accepted `data.chat_template` and never applied it, so a prefix-stable override changed nothing and a Qwen3.5 run still died in the response-mask check after the model had loaded. They confirmed on `main` that the same config trains on the transformers backend, then refused the combination at config load with a message that names the field, the backend and both ways out, parametrised the tests over two registered names and a raw Jinja string on sft, dpo and orpo with the empty-string and unsloth controls, and removed the MLX DPO/ORPO warning and registry rows the refusal made unreachable ([#2075](https://github.com/MakazhanAlpamys/Soup/issues/2075), [#2051](https://github.com/MakazhanAlpamys/Soup/pull/2051))+ - `soup train` without `--yes` printed `Start training? [Y/n]:` even when nothing could answer it, so a script or a service looked hung and the message never mentioned `--yes`. They took the ruled shape (stop before the dataset loads, exit 3, a message that names `--yes`), said openly in the description that piping `y` stops working, put the check behind a `_stdin_isatty()` seam like autopilot's, wrote the `--yes` help sentence and regenerated the CLI reference, and extended the tests with controls for `--yes`, `--dry-run`, the loader and trainer never being reached, and a faked TTY ([#2077](https://github.com/MakazhanAlpamys/Soup/issues/2077), [#2052](https://github.com/MakazhanAlpamys/Soup/pull/2052))- **Akshat Shah** ([@akshatshahh](https://github.com/akshatshahh))- First contribution: `http://[::1]:<port>` was accepted by the judge API-base and vLLM checks but refused by eval-gate suites, `soup ship --judge-model` and the online-DPO trainer, because three gates kept their own `("localhost", "127.0.0.1")` literal and two more kept a third copy. They made all five import `net_guard.LOOPBACK_HOSTS` and added an AST ratchet that fails on any rebuilt loopback literal, including single-quoted and multi-line spellings. Their 27 tests: 11 fail on main, 27 pass here; 7 mutations, all killed ([#1548](https://github.com/MakazhanAlpamys/Soup/issues/1548), [#1554](https://github.com/MakazhanAlpamys/Soup/pull/1554))