Reach parity with FSDP / Megatron but with a Tinker endpoint.
This way we can use SkyRL's frontend code (e.g. training loop, sync or fully async; agent integration) but with a hosted service. Either the Tinker product, or a tinker implementation (e.g. SkyRL's own), so that we can dog food our tinker implementation against realistic workloads, while offering users a unified entry point for both local GPUs and serverless training endpints.
Reach parity with FSDP / Megatron but with a Tinker endpoint.
This way we can use SkyRL's frontend code (e.g. training loop, sync or fully async; agent integration) but with a hosted service. Either the Tinker product, or a tinker implementation (e.g. SkyRL's own), so that we can dog food our tinker implementation against realistic workloads, while offering users a unified entry point for both local GPUs and serverless training endpints.