Loading huggingface PEFT format LORA adapters in mlx

[nfergu]

Is it possible to load huggingface PEFT format LORA adapters in mlx?

I understand that it's possible to use LORA adapters in mlx by supplying the --adapter-path parameter when loading a model. However, as far as I can tell this only supports LORA adapters that have been created using mlx. Is it possible to load huggingface PEFT format adapters instead?

[awni]

No it's not yet supported. If the structure of those adapters is pretty similar to the MLX ones it should be fairly easy to implement.

Do you have a PEFT adapter you were interested in that we can look at the format of?

[nfergu]

No it's not yet supported. If the structure of those adapters is pretty similar to the MLX ones it should be fairly easy to implement.

Do you have a PEFT adapter you were interested in that we can look at the format of?

Sure ... here's an example: https://huggingface.co/tttx/sft-32b-020925-19k-5ep

It doesn't actually say that it's a PEFT adapter, but looking at adapter_config.json it mentions PEFT so I think it is:

{
  "alpha_pattern": {},
  "auto_mapping": null,
  "base_model_name_or_path": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
  "bias": "none",
  "fan_in_fan_out": false,
  "inference_mode": true,
  "init_lora_weights": true,
  "layer_replication": null,
  "layers_pattern": null,
  "layers_to_transform": null,
  "loftq_config": {},
  "lora_alpha": 64,
  "lora_dropout": 0.0,
  "megatron_config": null,
  "megatron_core": "megatron.core",
  "modules_to_save": null,
  "peft_type": "LORA",
  "r": 64,
  "rank_pattern": {},
  "revision": null,
  "target_modules": [
    "k_proj",
    "o_proj",
    "up_proj",
    "v_proj",
    "down_proj",
    "q_proj",
    "gate_proj"
  ],
  "task_type": "CAUSAL_LM",
  "use_dora": false,
  "use_rslora": false
}