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): #2160 [skip ci]

Alpamys51777d12077d
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
index b98ab9c6..92761ab5 100644
--- a/CONTRIBUTORS.md
+++ b/CONTRIBUTORS.md
@@ -261,6 +261,7 @@ Listed by first contribution. PR numbers link the work.
- Coordinated the base-model commit across distributed Hugging Face Transformers `soup train` workers: global rank zero resolves once, one nullable result is broadcast before any rank loads the base, and ranks compare receipts after setup and training. **Two review findings shaped it**: the first version created the process group with a 120 s timeout, which Trainer's later `ddp_timeout` (1800 s) cannot override, so every collective of a multi-GPU run would have timed out after two minutes; the fix names the 1800 s default and the test compares it with `TrainingArguments`' own default, so a one-second shortfall fails. Then the real two-process Gloo cases timed out on a loaded macOS runner, and they were moved to Linux only with a harness test that pins the platform gate and a failure report that prints both ranks' logs and reaps the children. The broadcast, rank-ownership, timeout, platform-gate and diagnostics boundaries were each mutated and killed by named tests ([#1902](https://github.com/MakazhanAlpamys/Soup/issues/1902), [#2127](https://github.com/MakazhanAlpamys/Soup/pull/2127))
- Refused the resident Hugging Face Kimi K2 bases at config validation (#2104 option 2) with the real import traceback and source hashes as evidence: `Kimi-K2-Base`, `-Instruct`, `-Instruct-0905` and `-Thinking` stop before model loading, while `stream_layers: true`, local directories, other hubs, Unsloth and K2.5/K2.6 stay as they were. **The part worth copying is the per-id guidance**: after review the Base/Instruct text names the guarded streamed route and Thinking/Instruct-0905 say "no supported route yet" without naming `stream_layers`, with each recipe carrying its own reason in the snapshot, and when #2107 merged they rewrote the wording to the merged state with its prerequisites and evidence limits; the tests assert the right text per id and the absence of the old wording, and swapping reasons between ids fails named tests. Thirteen first-round and five second-round mutations, all killed ([#2104](https://github.com/MakazhanAlpamys/Soup/issues/2104), [#2130](https://github.com/MakazhanAlpamys/Soup/pull/2130))
- Gave `soup data ingest` an installable reader: a `documents` extra with `pypdf>=6.19.0` and `python-docx>=1.2.0`, both missing-reader errors naming `pip install "soup-cli[documents]"`, and Markdown/TXT documented as needing no extra. **The part worth copying is that CI needed no new cell**: stub readers pin the contract while the real PDF and DOCX round trips run when the packages are installed, and the comment above the extra says so after the review asked for it to be true. A test keeps `documents` out of `[all]`; 7 mutations each fail a named test ([#2057](https://github.com/MakazhanAlpamys/Soup/issues/2057), [#2096](https://github.com/MakazhanAlpamys/Soup/pull/2096))
+ - Added the `qwen3.5-2b-dpo` recipe from #275: the inherited Qwen3.5 DPO settings on `Qwen/Qwen3.5-2B`, text only with an explicit `modality: text`, named as untuned in the docs and fragment. The model id is pinned on `RecipeMeta.model` and the YAML `base:` and each fails its own test; the recipe is registered in the Qwen3.5 family audit, the count moved 170 -> 171 at every site and the snapshot delta is purely additive. The PR text names the six default-equivalent lines that survive deletion instead of leaving the count unexplained. The CPU test builds a two-layer random-weight native Qwen3.5 decoder with one linear-attention and one full-attention block, attaches the recipe's own LoRA targets and checks non-zero gradients on all four projection kinds ([#275](https://github.com/MakazhanAlpamys/Soup/issues/275), [#2160](https://github.com/MakazhanAlpamys/Soup/pull/2160))
- **AmixDigital** ([@Amix29](https://github.com/Amix29))
- Added native Sesame CSM fine-tuning for transcript/audio pairs, training the backbone and depth decoder while keeping the Mimi codec frozen; documented supported inputs, refusals and the scoped Apple Silicon evidence ([#265](https://github.com/MakazhanAlpamys/Soup/issues/265), [#1769](https://github.com/MakazhanAlpamys/Soup/pull/1769)).