Inkling 2 Mini
Harness adapters
Pi is the first harness. Training is planned on two HF H200s, with zero training updates so far. The learner exported all seven adapter projections. The next check is whether the reference vLLM server generates correctly with them on the released Inkling model. The scores below come from the earlier SkyRL runs.
Live GPU activity
Earlier SkyRL training
These runs used a synchronous SkyRL learner. The final Pi sampler snapshots showed 71.1% and 69.8% rolling prefix-cache hits. That learner recomputed shared prefixes across paths. Cache measurements · Original compatibility test