Miles

license

Apache-2.0(OSI)

source

public(radixark/miles)

core-gated

ungated(no enterprise or ee path in the repository root and no paid tier described in the README)

Apache-2.0, an OSI licence, with the full source public and the shipped code in this repository. The LICENSE body was read in full: the stock Apache-2.0 text, its core byte-identical to the canonical template, with nothing after END OF TERMS AND CONDITIONS but the standard appendix. No enterprise or ee path in the repository root and no paid tier described in the readme.

• https://raw.githubusercontent.com/radixark/miles/main/LICENSE recorded 2026-08-31

  LICENSE body, read in full: the stock Apache License 2.0 text, carrying no appended term or additional condition.

• https://raw.githubusercontent.com/radixark/miles/main/README.md recorded 2026-08-31

  README, read for the distribution and gating question: it describes installing and running the code from this repository and names no enterprise, paid or licence-gated build.

No package registry carries this product, so stars are the only available signal and the band is capped at 3 by the instrument. 2,293 stargazers places it at level 2. Stars measure attention rather than use, which is why the confidence is low. A PyPI project named `miles` exists but belongs to an unrelated package - version 0.1, no summary and no project URLs pointing at this repository - so it is not declared here.

• https://api.github.com/repos/radixark/miles recorded 2026-08-31

  Repo metadata for radixark/miles: stargazers_count 2293, archived false, pushed_at 2026-08-31.

Miles runs RL post-training at trillion-parameter scale on a Megatron-LM training backend with SGLang rollout, and its async design and P2P RDMA weight transfer are the same class of infrastructure verl provides. It sits one band below verl because verl publishes a substantially wider algorithm catalogue (PPO, GRPO, RLOO, ReMax, REINFORCE++, GSPO, PRIME, DAPO, DrGRPO plus SFT/DPO) while Miles documents PPO and GRPO; method breadth is what separates the top two bands in this category.

• https://raw.githubusercontent.com/radixark/miles/main/README.md recorded 2026-08-31

  About and Performance sections describe pairing SGLang for rollout with Megatron-LM for scalable training, a PyTorch FSDP2 backend alternative, fully async RL with decoupled rollout/training workers, and in-loop weight updates over P2P RDMA on a trillion-parameter model (Kimi-K2.6).