verl-mint¶

verl-mint is the open-source Mint Recursive training runtime built on veRL and Ray. It preserves the meaning and behavior of sessions, checkpoints, samplers, and futures in the Mint Recursive API. Your client follows the same training workflow, while your own GPUs, Ray or veRL cluster, and storage systems handle execution and storage.

The same SDK script that runs against the Mint Recursive hosted service runs unchanged against a local verl-mint deployment.

Algorithms¶

Algorithm	Status	Notes
SFT	Available	LoRA SFT, checkpoint save/load, resume, sampler handoff
GRPO	Available	Group Relative Policy Optimization; group-relative advantages without a value critic
DPO	Planned	Preference optimization track

Supported models¶

Family	Size	Status
Qwen	0.6B	Available
Qwen	4B	Available
Qwen	30B	Available
Qwen	8B	Planned

Install¶

git clone https://github.com/verl-project/verl-mint.git
cd verl-mint
uv sync --extra smoke

The smoke extra installs the official Mint Recursive SDK (mindlab-toolkit) from the MindLab repository.

Prerequisites¶

Set up Ray, GPUs, and worker images before using verl-mint. You provide this execution environment; the runtime does not provision it.

Prerequisite	Requirement
Ray	`ray[default]>=2.46.0,<3` on the API node and every worker

See the verl-mint repository for installation steps, algorithm details, and cluster configuration.