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): #2115 [skip ci]

Alpamysf312fed99c14
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
index 7c11211c..9a42b6a0 100644
--- a/CONTRIBUTORS.md
+++ b/CONTRIBUTORS.md
@@ -329,6 +329,7 @@ Listed by first contribution. PR numbers link the work.
- Closed #1758, reported by @abdulwaarith0: `utils/live_eval` had no public way to pass chat-template options or to render and encode a prompt as `make_generator` does, so callers patched private names and a generation could change without any error. They made `encode_prompt` and `MAX_PROMPT_TOKENS` public, routed both generators through that one helper, added `chat_template_kwargs` to both and `soup ship --chat-template-kwargs`, and documented `loaded=(model, tokenizer, device)` reuse and caller-managed batch padding. **The part worth copying is how the review gap closed**: the refusal test pinned 3 of the 9 reserved names, so six removals and a shallow copy of the options kept the suite green; the next push parametrized the refusal over the whole set (now 11, with `chat_template` and `return_assistant_tokens_mask` added), pinned a nested value through a template that reads `settings.flags.enable_thinking`, kept the underlying render error in the message and added the `RecursionError` catch the review had marked optional. In a fresh run each of the 11 names removed one at a time fails its own case, a shallow copy fails 3 tests and 35 of 36 mutations are caught; the survivor is the `isinstance(prompt, str)` guard in the public `encode_prompt`. They say plainly that the end-to-end test is a tiny randomly initialized Llama on CPU, with no CUDA or pretrained-quality claim ([#1758](https://github.com/MakazhanAlpamys/Soup/issues/1758), [#1776](https://github.com/MakazhanAlpamys/Soup/pull/1776))
- Closed #1760 (reported by abdulwaarith0): a GGUF export printed nothing while it converted and could not skip the pip step for the convert dependencies. `soup export --format gguf` now reports its phases, `--verbose` streams the converter and quantizer output with the failure summary kept (bounded tail, child terminated and reaped on an output error), and `--no-install-deps` names missing `gguf` / `sentencepiece` / `protobuf` before any merge or toolchain work without calling pip. **The part worth copying is how they closed review findings with tests that fail for the right reason**: the dropped verbose forwarding on the direct f16/f32 convert call is pinned by eight cases, the `protobuf` probe by a test that fails when its module name is renamed, when presence ignores it or when the probe is dropped, and the phase-line assertions went through the ANSI-strip helper so they hold under `FORCE_COLOR=1`. The PR says plainly that no real llama.cpp conversion was run ([#1760](https://github.com/MakazhanAlpamys/Soup/issues/1760), [#1777](https://github.com/MakazhanAlpamys/Soup/pull/1777))
- Added native DoRA to MLX text SFT, which `backend: mlx` had refused: mlx-lm's DoRA layers on dense unquantized and pre-quantized 4-bit targets, `fine_tune_type: dora` written by both metadata sites so the adapter reloads and fuses natively, unsupported targets refused before any layer is converted, and a warm start that refuses an adapter with missing, extra or reshaped tensors before loading any of them. **Review of the first round asked for a test of the `task == "sft"` condition, and the new test pins the DoRA-specific message (grpo and pretrain), because the general MLX task validator would otherwise hide a removed condition.** `tests/test_mlx_dora.py` passes 14 (8 fail on the merge-base), 7 validator mutations and the `fine_tune_type` site are each caught, and the MLX-only lines run in the `mlx-smoke` job ([#1918](https://github.com/MakazhanAlpamys/Soup/issues/1918), [#1920](https://github.com/MakazhanAlpamys/Soup/pull/1920))
+ - Added the loss watchdog to native MLX text SFT: reports and the stop are counted in completed optimizer updates under gradient accumulation, the stop uses a signal only Soup catches, and the last adapter is saved with the stop reason and the completed work recorded; a non-finite loss is an error, documented as the difference from the Transformers watchdog ([#1989](https://github.com/MakazhanAlpamys/Soup/issues/1989), [#2115](https://github.com/MakazhanAlpamys/Soup/pull/2115))
- **Srinivasan R** ([@Srinivasan8888](https://github.com/Srinivasan8888))
- Connected recorded thumbs feedback to loop harvesting ([#1440](https://github.com/MakazhanAlpamys/Soup/issues/1440), [#1832](https://github.com/MakazhanAlpamys/Soup/pull/1832)).