What is the recommended way to add a new base model to TuneOS?

[SahilKumar75]

I want to fine-tune a model that isn't in the supported list (e.g. Qwen2 or Phi-4). What's the right way to add support for a new base model in TuneOS? Do I need to modify the trainer, the config, or just the API model list?

[SahilKumar75]

To add a new base model:

1. Add it to app/api.py — append an entry to _SUPPORTED_MODELS with name, hf_id, and notes (VRAM estimate).

2. Test compatibility — run a quick LoRA injection test using trainer/lora.py with the new model ID. Most causal LM models work out of the box via AutoModelForCausalLM.

3. Check target modules — different architectures use different attention layer names. For Qwen2 use ["q_proj", "v_proj"], for Phi use ["q_proj", "k_proj", "v_proj", "dense"]. Update lora_cfg defaults accordingly.

4. Update docs/supported-models.md with VRAM requirements and any special flags (e.g. trust_remote_code=True for some models).

That's it no changes to the trainer pipeline itself are needed.