Can I use my own checkpoint built on a supported base model?

[zyy07]

Hi Team! 😊I’d like to know whether it’s possible to upload and use my own checkpoint as the starting point for fine-tuning through the Tinker-like API. What's your take on this issue?

[pcmoritz]

Hi @zyy07 yes you can do this with skyrl-tx, you can just run something like (in the https://github.com/NovaSky-AI/SkyRL/tree/main/skyrl-tx directory):

uv run --extra gpu --extra tinker -m tx.tinker.api  --base-model Qwen/Qwen3-4B-Instruct-2507 --backend-config '{"max_lora_adapters": 3, "max_lora_rank": 1, "tensor_parallel_size": 8, "train_micro_batch_size": 8}'

(the CLI options are documented here https://github.com/NovaSky-AI/SkyRL/blob/main/skyrl-tx/tx/tinker/config.py#L12 and the --backend-config options are documented here https://github.com/NovaSky-AI/SkyRL/blob/main/skyrl-tx/tx/tinker/backends/jax.py#L59) and --base-model can be any base model that supports the architectures we have implemented so far, currently Qwen3ForCausalLM, Qwen3MoeForCausalLM and LlamaForCausalLM / Llama3ForCausalLM, more architecture contributions are welcome). See more usage examples in https://github.com/NovaSky-AI/SkyRL/tree/main/skyrl-tx#-usage-examples and you can swap out the base models with your own huggingface checkpoints.

Then you can train a model e.g. with

export TINKER_API_KEY="dummy"
uv run --with wandb --with tinker==0.3.0 sl_loop.py     base_url=http://localhost:8000     model_name=Qwen/Qwen3-4B-Instruct-2507 lora_rank=1 max_length=512

from https://github.com/thinking-machines-lab/tinker-cookbook, or run RL training e.g. with rl_loop.py.

Let us know if you have more questions about this or run into problems with any particular model, happy to help!