A note of thanks · Scale boundary

Halo

Halo brings pre-training, SFT, preference work, and RL into one training framework that spans a GPU and clusters. For this lab it is a map of the single-Mac boundary, not a Mac runtime dependency.

01 · The idea

What stayed with us

HF-native training and portable checkpoints, independent expert/context/tensor/data parallelism, faithful multi-turn trajectories, operation-specific model support, and explicit resume/export/serving boundaries make the single-machine-to-cluster transition inspectable.

02 · The local translation

What we did with it

Halo informed completed Issues #159, #161 and #162: trajectory export preserves tool actions and observations with per-turn supervision; compatibility reports distinguish static predictions from matching device receipts; artifact lifecycle contracts bind identity, history, runtimes and legal next actions. The 2026-10-02 source-pinned review connects these implementations to the original contracts.

03 · The boundary

Where the comparison stops

The documented runtime requires NVIDIA/CUDA containers, not MLX or Metal. Halo has not been run on this Mac or adopted as a dependency, and its B300 benchmarks have not been reproduced. Sampled-token RL, asynchronous weight synchronization and distributed training were not implemented locally; its license includes supplemental commercial and attribution terms.