alt.freeweights.studio
Freeweights Studio
Studio fine-tunes and runs Freeweights models from one job file for every engine. Training runs on your own computer: Apple, NVIDIA, AMD, Intel or CPU. It arrives in Phase 3. Until then this page shows the honest status of each engine and each kind of hardware.
Fine-tune in Freeweights Studiocoming in Phase 3
Hardware
Before a run, Studio checks your computer and tells you in plain words if a method or output is not available on it. Speed, memory use and results depend on your hardware and the model.
| Hardware | Backend | Status |
|---|---|---|
| Apple silicon (Mac) | MLX for LoRA and QLoRA; PyTorch MPS for full fine-tune and DPO | Not yet tested by Allternit on this hardware |
| NVIDIA on Linux | PyTorch CUDA | Not yet tested by Allternit on this hardware |
| NVIDIA on Windows | PyTorch CUDA | Not yet tested by Allternit on this hardware |
| AMD on Linux | PyTorch ROCm | Not yet tested by Allternit on this hardware |
| Intel GPU on Linux or Windows | PyTorch XPU | Not yet tested by Allternit on this hardware |
| CPU (also AMD on Windows and Intel Macs) | PyTorch CPU: LoRA only, small models, slow | Not yet tested by Allternit on this hardware |
Engines
| Engine | Kind | Status |
|---|---|---|
| Allternit Tune (Apache-2.0) | training | Not started |
| Upstream Unsloth (unmodified) | training | Not started |
| Axolotl | training | Not started |
| TRL + PEFT | training | Not started |
| LLaMA-Factory | training | Not started |
| torchtune | training | Not started |
| mlx-lm | training on Mac | Not started |
| Allternit engine | local inference | Not started |
| llama.cpp | local inference | Not started |
| MLX | local inference | Not started |
| Ollama | local inference | Not started |
| vLLM | server inference | Not started |
| SGLang | server inference | Not started |