mrleo1nid/swarm-assistent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Ship Assistent 0.13.0: real QLoRA pipeline and GGUF Ollama register.

Browse Source 
Replace fake train loop with TRL SFTTrainer, HF column mapping with fiction preset,
safetensors to GGUF conversion, and ollama create using ollama_base plus ADAPTER.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Leonid Pershin
2026-08-22 14:45:07 +03:00
co-authored by Cursor
parent 1a03c3178f
commit 474a674a35
12 changed files with 1002 additions and 133 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+12
View File
@@ -4,6 +4,8 @@ SwarmUI extension for **collaborative Krea 2** prompting via **Ollama**: chat +
**Turn model:** one user message is one *turn*. A turn may fan out into nested LLM *hops* — Krea prompt prep, empty-patch retry, vision, auto-critique. Hops share one `HOP_BUDGET`, never re-read the user's text (their prompt is client-authored), and pass the busy gate that blocks new user sends. What a reply does to generation state is decided once, in `resolveTurnIntent`: the model's `actions:["generate"]` / `look_at` win, RU intent heuristics only back it up when the model forgets, and an explicit «запомни, не генерируй» vetoes both.
**Version 0.13.0** — **Реальный QLoRA-пайплайн**: `train_qlora.py` (TRL SFTTrainer + PEFT), HF-датасеты с маппингом (preset fiction title/tags→text), `max_samples`, полный post-train: safetensors → GGUF (`convert_lora_to_gguf.py`) → `ollama create` с `FROM ollama_base` + `ADAPTER`. Раннер: `builtin` + `custom`. Зависимости: `scripts/requirements-train.txt`.
**Version 0.12.1** — **Услышанное → агент**: одобренные примеры датасета сразу попадают в vector memory (`kind=heard`) и в контекст чата как `heard_examples` (без QLoRA). На вкладке «Датасет»: авто-подключение при одобрении, синхронизация всех, per-sample 🔗. Агент может запросить `heard_search`. Настройки: `training-agent.json`.
**Version 0.12.0** — App-level tabs (Чат / Карточки / **Обучение** / Настройки), боковая панель истории чатов, вкладка обучения LLM: курирование диалогов, импорт JSONL/CSV, Hugging Face datasets (фильтр совместимости), быстрый Ollama Modelfile, опциональный QLoRA-раннер с локаутом VRAM. HF token из SwarmUI User Settings (`huggingface_api`).
@@ -217,6 +219,16 @@ Patch fence keys: single source `Config/_base/patch-keys.json` → C# + client v
| `AssistentGetDatasetAgentSettings` / `AssistentSaveDatasetAgentSettings` | «Услышанное» → agent RAG (`training-agent.json`) |
| `AssistentLinkTrainSampleToAgent` / `AssistentUnlinkTrainSampleFromAgent` / `AssistentSyncDatasetToAgent` | Embed approved samples as `heard` memory |
## QLoRA setup (0.13.0)
1. Python env with CUDA: `pip install -r scripts/requirements-train.txt`
2. SwarmUI **User Settings** → `huggingface_api` (for HF base model download)
3. **Настройки → Модели**: runner kind = **builtin**, paths to `convert_lora_to_gguf.py` and **GGUF base** (same arch as HF base)
4. **Обучение → QLoRA**: HF base id, **Ollama base** (existing tag), output name, optional HF dataset (`krplt/ru-fictext-nsfw` auto-maps fiction preset)
5. Pipeline: train → `adapter_model.safetensors` → GGUF → `ollama create` with `ADAPTER`
Manual test checklist: small JSONL (5 pairs); HF dataset with `max_samples=50`; cancel job; missing deps (exit 2); missing gguf script (completed with note).
## License
MIT