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): #1951 [skip ci]
Alpamys8b73e35e2e3e
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.mdindex 963905f7..590651cc 100644--- a/CONTRIBUTORS.md+++ b/CONTRIBUTORS.md@@ -942,6 +942,7 @@ Listed by first contribution. PR numbers link the work.- Closed #2076, reported by @feloau (split from #1913): on Apple Silicon `soup train` died with no traceback while loading weights, because transformers' threaded loader aborted in `_materialize_copy`. They set `HF_DEACTIVATE_ASYNC_LOAD=1` through one helper, `configure_mps_async_load`, before the model is built for `soup train`, `soup sweep` and `soup bench train`, keep a value the user set (so `0` opts out), leave cuda, cpu and mlx alone, and made `soup doctor` show the variable's state without claiming it is set. Eleven mutation checks each fail a named test, and @feloau verified the branch on an M2 Max ([#2076](https://github.com/MakazhanAlpamys/Soup/issues/2076), [#2040](https://github.com/MakazhanAlpamys/Soup/pull/2040))- Closed #1423: `soup profile` and the MCP `tool_profile` budgeted every run as LoRA, so a full fine-tune (`lora.r: 0`, Spectrum, LISA, `prm`, classifier or ASR without LoRA) showed 0 trainable parameters and 0 GB of optimizer memory. They moved the train pre-flight's classification, branch for branch, into one `resolve_peft_mode` shared by all three callers and gave `estimate_total` a full-fine-tune path with a gradient term; when review found no test pinned the gradient budget, they asserted the exact values (14.9 GB gradients, 62.28 GB total) and the rendered `Gradients` row after collapsing the table padding, so 17 of 20 one-edit mutations of the changed lines now fail the new file. ([#1423](https://github.com/MakazhanAlpamys/Soup/issues/1423), [#2041](https://github.com/MakazhanAlpamys/Soup/pull/2041))- Closed #1810: `soup compile` and `soup compile-tools` did not use the eval suite's signal correctly for DSPy, TextGrad and GEPA, so the documented TextGrad `set_backward_engine` route returned the unedited prompt with `Score: 0.000` and `Converged: True`, and `compile-tools` had no flag to name a model. They wired the metric and iteration bound through, added `--task-lm`, `--reflection-lm` and `--engine`, and kept the provider-free tests honest by making their doubles enforce the published TextGrad 0.1.8 and GEPA guards (engine singleton exclusivity, no `score` attribute, `val_aggregate_scores`), so 51 of 54 one-edit mutations of the source lines now fail the tests and the other 3 are equivalent. ([#1810](https://github.com/MakazhanAlpamys/Soup/issues/1810), [#1956](https://github.com/MakazhanAlpamys/Soup/pull/1956))+ - 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))- **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))