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
Commits on main
fix(grpo): attribute non-finite parameters in watchdog warning and trainer state (Refs #342) (#2060)
4a0c314 Pernav Jain committed 27 days ago
fix(config): gate reasoning_effort and train_on_eot on the tasks that read them (#806) (#1587)
304e62e Dipak Chaudhari committed 27 days ago
fix(serve): require the tool token on the adapter routes; let soup loop send it (#1139) (#1184)
46b65d0 Dipak Chaudhari committed 27 days ago
fix(doctor): validate config before printing environment success summary (#2100) (#2106)
7a6edd9 Nariman Abdikarim committed 27 days ago
feat(stream): training.stream_recompute keeps streamed activations instead of recomputing (#2152)
2d56210 Alpamys Makazhan committed 27 days ago
bench: publish layer count reconstruction validator (#1289)
e23db1b Samarth Mukhija committed 27 days ago
fix(simpo): cap the prompt on the transformers backend, as dpo and orpo do (#2112) (#2116)
2586427 dryn85-max committed 27 days ago
docs(data): document accepted and refused preference row layouts for DPO and KTO (#2086) (#2120)
d6a4535 Areen Agrawal committed 27 days ago
feat(provenance): coordinate distributed base revisions (#2127)
a7d31d6 Nurkhan Esenbek committed 27 days ago
fix(ui): open a loopback URL when the server binds to all interfaces (#2054) (#2110)
bb3694d Rahul Seervi committed 27 days ago
feat(grpo): add the judge_rank reward, one LLM-judge call per group (#2047) (#2117)
4956cae Yang Bai committed 27 days ago
fix(adapters): read bfloat16 adapters in adapters check and adapters diff (#2058) (#2063)
a1921ee Younes Beriane committed 27 days ago
ci: run the remaining MLX test files in the macOS mlx-smoke job (#2147)
2032ca7 Alpamys Makazhan committed 27 days ago
fix(config): refuse lora rank_pattern/alpha_pattern with backend: mlx (they were ignored) (#2150)
e05b6f6 Alpamys Makazhan committed 28 days ago
feat(recipes): add Qwen3.5 4B text DPO variant (#2129)
a82528c Nurkhan Esenbek committed 28 days ago