issue merging lora weights into qwen3 moe with peft

[csmith23]

We finetuned Qwen/Qwen3-30B-A3B-Instruct-2507, and are running into issues using PEFT to merge in the weights.
Specifically, PEFT is unable to failing to find the adapter keys when merging weights with the base model. We’re using this pattern to merge the weights

    base_model_name = "Qwen/Qwen3-30B-A3B-Instruct-2507"   
 
    model = PeftModel.from_pretrained(base_model, adapter_path)
    merged_model = model.merge_and_unload()
    merged_model.save_pretrained(merged_path)
    tokenizer = AutoTokenizer.from_pretrained(base_model_name)
    tokenizer.save_pretrained(merged_path)
    finetuned_vol.commit()

where the adapter_path is the path to the un-tarred directory (containing adapter_config.json and adapter_model.safetensors files) that was downloaded following the docs at https://tinker-docs.thinkingmachines.ai/download-weights.

The error we get is

/usr/local/lib/python3.12/site-packages/peft/peft_model.py:585: UserWarning: Found missing adapter keys while loading the checkpoint: ['base_model.model.model.layers.0.self_attn.q_proj.lora_A.default.weight', 'base_model.model.model.layers.0.self_attn.q_proj.lora_B.default.weight',
....
]

Is there anything we're missing that would solve this issue?

[danobi]

(First posted to discord)

Here's the script to merge weights: merge_tinker_adapter_to_hf_model.py

At some point we'll offer an API for this.