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): #1288 [skip ci]
Alpamys9b2f4030a020
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.mdindex 4f318e3d..a83bed0d 100644--- a/CONTRIBUTORS.md+++ b/CONTRIBUTORS.md@@ -480,6 +480,7 @@ Listed by first contribution. PR numbers link the work.- Made mixed direct-I/O stripes visible through per-root runtime stats and a `k/N` ready line, named each page-cache fallback and its operating-system refusal reason, and limited disk read-ahead guidance to disk-tier runs ([#1615](https://github.com/MakazhanAlpamys/Soup/issues/1615), [#1962](https://github.com/MakazhanAlpamys/Soup/pull/1962)).- Published the STEP 8 pinning-cost measurement (`benchmarks/harness/pincost.py`): pinned versus pageable streamed NF4 throughput under the historical control, n=3 per arm against one resident reference. **It reports the historically wrong pinned arm as WRONG next to its tok/s rather than refusing to time it, and it decides from the measurement**: the ratio must reach 1.5x and exceed the arms' own spread, otherwise exit 1. It took five review rounds; the last one closed an empty LoRA-gradient set that read as a pass (0 of 0). ([#379](https://github.com/MakazhanAlpamys/Soup/issues/379), [#1172](https://github.com/MakazhanAlpamys/Soup/pull/1172))- Published the STEP 6 depth-versus-bytes reconstruction (`benchmarks/harness/depth_vs_bytes.py`): the record's synthetic 48-layer Llama sweep with three backward reads per row, and **a verdict that fails any sweep contradicting the record**, namely every NF4 read at or below 163.8 MiB per layer exact, every read at or above 171.5 MiB wrong, and a bf16 control above the bracket exact (the version first submitted accepted a bf16 "control" at 0.05 MiB, which cannot rule bytes out). Also added the `--hidden`, `--intermediate`, `--heads` and `--kv-heads` flags to `synth_checkpoint.py`. ([#379](https://github.com/MakazhanAlpamys/Soup/issues/379), [#1287](https://github.com/MakazhanAlpamys/Soup/pull/1287))+ - Published the STEP 2b repeated-backward reconstruction (`benchmarks/harness/repeat_backward.py`): five streamed backwards under the historical control against one persistent resident NF4 reference. **It does not require the first backward to be exact**, because the record has it at 256/256 in one run and 20/256 in another; it requires later backwards to be wrong while every loss stays finite and equal to the resident loss, and refuses empty, mismatched and non-finite gradient sets. ([#379](https://github.com/MakazhanAlpamys/Soup/issues/379), [#1288](https://github.com/MakazhanAlpamys/Soup/pull/1288))- **chadetov** ([@chadetov](https://github.com/chadetov))- First contribution, and it is the kind that is easy to file as a one-line fixture tweak and much harder to *diagnose*: `test_sft_smoke` had been red since v0.36.0, and they traced why nobody knew. `pyproject`'s `addopts` carry `-m 'not smoke'`, and the only smoke job in CI is `mlx-smoke`, which selects `-k mlx_sft_smoke` on an install that asserts the PyTorch stack is **absent** — so `test_sft_smoke` and `test_dpo_smoke` run in neither the default suite nor CI, which is how a fixture drifted out of sync with a hard error added five releases earlier. **The scope discipline is what got it merged unchanged**: the MLX fixture carries the same `format: chatml` line and they left it alone, because it reaches the MLX trainer rather than `sft_format.py` and is the one smoke test CI actually runs and passes — verified, not assumed. They also declined to wire the PyTorch smoke tests into CI in the same PR and said why (model downloads, runner minutes), leaving the call to the maintainer instead of taking it ([#594](https://github.com/MakazhanAlpamys/Soup/pull/594)). Separately, and on the same day, they measured `utils/log_level.py`'s branch coverage on #273 and filed [#595](https://github.com/MakazhanAlpamys/Soup/issues/595) — four code sites quoting a probe-denominated VRAM ratio under a "real peak" label, plus a schema field description still carrying a justification the gate record had explicitly withdrawn