v0.7.0

What's Changed

• fix(train,pi07): stabilize multi-rank in-training sim eval (OOM + desync) by @shuheng-liu in #409
• test(ci): quarantine live-Hub tests off the gating CPU run by @shuheng-liu in #410
• chore(claude): learn from #410 by @claude[bot] in #411
• fix(envs): silence per-worker robosuite import banner in sim-eval logs by @shuheng-liu in #412
• feat(train): add running-best checkpoint retention by @shuheng-liu in #413
• perf(datasets): read v3.0 columns via arrow formatter, not row-wise numpy by @shuheng-liu in #415
• fix(datasets): drop non-tensorizable columns in hf_transform_to_torch by @shuheng-liu in #416
• perf(datasets): columnar v3.0 meta read + per-repo dedup cache by @shuheng-liu in #417
• fix(datasets): val-split RAM + v3.0 video frame-index clamp + load retry by @shuheng-liu in #418
• fix(train): reset policy after in-training eval to free eval-batch buffers by @shuheng-liu in #419
• feat(eval): shrink wandb eval-video uploads via CRF + frame subsampling by @shuheng-liu in #420
• feat(train): fork the wandb run on resume by default by @shuheng-liu in #421
• Adding async high level planner and vla code to server.py by @akshay18iitg in #422
• perf(flash_cuda): head-parallel dK/dV + vectorized loads + cp.async by @WilliamYue37 in #376
• chore: bump version to 0.7.0 by @WilliamYue37 in #423

Full Changelog: v0.6.0...v0.7.0