Currently we are missing example startup scripts for the tinker engine + skyrl-train other than the canonical examples in: https://docs.skyrl.ai/docs/tinker/overview.
We should populate https://github.com/NovaSky-AI/SkyRL/tree/main/examples/tinker with configs (max_tokens_per_microbatch/max context, model size, parallelism strategies) on fixed hardware configurations (B200/B300 + H100/H200) + model families (Qwen3.6-35B, Nemotron, etc..). This can largely be generated based off of existing configs for the python entrypoint.
Currently we are missing example startup scripts for the tinker engine + skyrl-train other than the canonical examples in: https://docs.skyrl.ai/docs/tinker/overview.
We should populate https://github.com/NovaSky-AI/SkyRL/tree/main/examples/tinker with configs (max_tokens_per_microbatch/max context, model size, parallelism strategies) on fixed hardware configurations (B200/B300 + H100/H200) + model families (Qwen3.6-35B, Nemotron, etc..). This can largely be generated based off of existing configs for the python entrypoint.