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): #2106 [skip ci]

Alpamys274f6f60b444
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
index 6cf9da2f..cc772fc6 100644
--- a/CONTRIBUTORS.md
+++ b/CONTRIBUTORS.md
@@ -955,6 +955,7 @@ Listed by first contribution. PR numbers link the work.
- Removed the deprecated `awq` and `gptq` export and quantize formats, the `[awq]` / `[gptq]` extras, the `--bits` / `--group-size` / `--calibration-samples` options and the export helpers behind them, as announced for v0.77, and updated the docs, README and its translations ([#1710](https://github.com/MakazhanAlpamys/Soup/issues/1710), [#1951](https://github.com/MakazhanAlpamys/Soup/pull/1951))
- `training.unlearn_alpha: 0.0` is documented as pure forget loss, but the unlearn trainer read it as `unlearn_alpha or 1.0`, so an explicit 0.0 silently kept the retain term at full weight. They replaced the truthiness check with an explicit `None` check at both sites (NPO/SimNPO and RMU) and wrote nine CPU tests that run the real wrapper methods for 0.0, unset and 0.5; putting `or 1.0` back at either site fails a named test ([#2108](https://github.com/MakazhanAlpamys/Soup/issues/2108), [#2114](https://github.com/MakazhanAlpamys/Soup/pull/2114))
- Added the eager `soup --version` option, which prints the same text as `soup version`, exits 0, wins over `--help` and a bad `--log-level`, and stays torch-free. Review found that nothing pinned `is_eager=True` (deleting just that line left every test green and made `soup --version --help` print the usage text); they added a test that invokes `["--version", "--help"]`, and removing the line now fails it ([#2101](https://github.com/MakazhanAlpamys/Soup/issues/2101), [#2105](https://github.com/MakazhanAlpamys/Soup/pull/2105))
+ - Closed #2100: `soup doctor --config <a config that is refused at load>` printed "All checks passed!" and then refused the config with exit 2, so the summary contradicted the result. They moved the config load ahead of the summary and the first version then hid the environment problems behind the refusal; review found it, and the second version printed the issue list and the fix commands first, then the refusal, with the green line only when nothing was wrong. The last round pinned the one condition no test covered (the green line with a missing package and a loadable config) and the two error exits the move had left untested. ([#2100](https://github.com/MakazhanAlpamys/Soup/issues/2100), [#2106](https://github.com/MakazhanAlpamys/Soup/pull/2106))
- **John Coplin** ([@johncoplin](https://github.com/johncoplin))
- First contribution: the friendly-error formatter's `HTTP 401`/`HTTP 403` text patterns let two whitespace groups compete for the same run, so an exception message with `HTTP` followed by a long whitespace run and no status took about 13 s to print at 50,000 spaces. Since #1366, that check runs for every exception in the chain. They claimed the issue, rewrote both patterns so the optional whitespace is consumed only after a `:`/`=` separator, and added a timing test that fails on `main`. **The part worth copying is the equivalence test**: all 15 phrasings from the issue run against the old and the new pattern for both 401 and 403, so the speed-up provably changes nothing about what gets classified. 31 tests (1 fails on main), 7 of 7 mutations killed ([#1367](https://github.com/MakazhanAlpamys/Soup/issues/1367), [#1594](https://github.com/MakazhanAlpamys/Soup/pull/1594))