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): #2007 [skip ci]
Alpamysd2c07e964447
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.mdindex b60df88b..71272331 100644--- a/CONTRIBUTORS.md+++ b/CONTRIBUTORS.md@@ -786,6 +786,7 @@ Listed by first contribution. PR numbers link the work.- `soup ab` exited on a bare `ZeroDivisionError` or `OverflowError`, or printed `continue` with a `nan` ratio, for a column whose spread or size a float cannot carry (`0` and `1e-161`, `0` and `1e200`, 40 rows of `1e308`). They refused all three where they arise, with a message naming the column, and squared as `d * d` so an overflowing deviation is caught instead of raised. 11 tests, 8 failing on `main`; 18 of 19 mutations caught after two review rounds ([#1384](https://github.com/MakazhanAlpamys/Soup/issues/1384), [#1392](https://github.com/MakazhanAlpamys/Soup/pull/1392))- The #724 wiring scan only counted `get_peft_model(` and a trl `peft_config=` as building a LoRA adapter, so `mlx_sft.py` (mlx-lm's `linear_to_lora_layers`) was neither wired nor exempt, which is how `backend: mlx` ignored `loraplus_lr_ratio` unseen. After holding their earlier #1385 while #1383 merged, they came back with the follow-up the review named: the scan now counts the MLX builder and `mlx_sft.py` is exempt with a reason, **a new test fails when an exemption reason names a test file that does not exist, so the pointer to the merged file cannot go stale silently**, and the #1324 test file gained the cases #1383 lacked (every ratio, direct `SoupConfig(...)`, explicit `null`, wired-backend controls, plain `soup train`, `soup doctor --config` exit 2) plus a docs line for `unsloth`. 135 passed; 6 of 6 mutations fail a test (1 to 9 tests each) ([#1324](https://github.com/MakazhanAlpamys/Soup/issues/1324), [#1516](https://github.com/MakazhanAlpamys/Soup/pull/1516))- `soup local-rl train` without `--once` rendered the nightly systemd `.service` / launchd `.plist` with `--db`, `--model` and `--train-method` only, so the scheduled job trained on as few as 10 pairs into `local_rl_adapter/` whatever `--min-pairs` and `-o` the user had passed. They claimed the slice on the umbrella issue, rendered both flags into both unit files, and moved the runner's validators out of `run_nightly_train` so the scaffold refuses a value the nightly run would refuse (`--min-pairs 0` exits 2) **before the scheduler directory is created, which a test pins**. The tests parse the `.service` with `shlex` and the `.plist` with `plistlib` and run the rendered argv through the real CLI. 14 new tests, 12 fail on `main`; 15 of 16 mutations fail a test, the survivor being the unpinned docs line ([#1448](https://github.com/MakazhanAlpamys/Soup/issues/1448), [#1517](https://github.com/MakazhanAlpamys/Soup/pull/1517))+ - `soup sweep --strategy random` hung forever, with no output and `--dry-run` included, when a value list named the same value twice (`lr=1e-5,2e-5,1e-5`, the same value spelled two ways, or a repeated entry in a `--sweep-config` file): it asked the sampler for as many distinct arms as the product of the list lengths, which did not exist. They deduplicated each list in first-seen order before counting, so the sampler draws only distinct arms and the run-them-all branch no longer repeats one, and added a draw-capped fixture so the regression fails fast instead of hanging CI. After the review they read the CLI check through the repo's ANSI-strip helper and added a test that records what the sampler sees, so a repeated value in a later list and the first-seen order are both pinned (the three variants, no dedup, `set()` and first-list-only, each fail a test) ([#2007](https://github.com/MakazhanAlpamys/Soup/pull/2007))- **Krishna** ([@Ru0k3](https://github.com/Ru0k3))- First contribution: `soup registry diff`, `soup eval compare` and `soup eval leaderboard` kept the oldest eval score per benchmark. They overwrote rows in the tracker's newest-first order, so a re-measurement was ignored: on the issue's example both commands reported +0.20 and no regression, where comparing the newest scores gives a -0.05 regression. **They found that the leaderboard had the same bug, which the issue did not name**, and routed all three commands and the eval gate's `registry://` baseline through one helper that picks the newest `created_at` and then the latest insert. Their tests list rows newest first, as the tracker returns them, so restoring the old code at any call site fails a test. The second round added a two-model leaderboard test for the one mutation that had survived, and skips empty benchmark names as #1260 did. 13 of 15 mutations fail a test, and the other two would need a NULL value that the schema forbids ([#1270](https://github.com/MakazhanAlpamys/Soup/issues/1270), [#1291](https://github.com/MakazhanAlpamys/Soup/pull/1291))