Uploaded September 2026 | Updated November 2026, 2 days ago
As models and data grow, the multimodal training bottleneck increasingly comes from the surrounding input work rather than the training step itself: decoding, transformation, and tokenization contending with gradient computation for accelerator time.
At Ray Summit 2026, Minghao Xie and Haoran Cai from Luma AI present Koba, their multimodal data-loading framework, and Kojo, a Ray-based substrate that disaggregates stateless preprocessing stages onto managed actor pools while preserving the existing training loop. The result: compute that scales on its own pool independently of the trainer, fault isolation so a preprocessing failure no longer cascades into training, and reclaimed accelerator capacity returned to the gradient path.
You'll leave with a reusable pattern for relocating stateless, out-of-step computation off the training GPUs and onto Ray, plus the operational lessons of running it at scale.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
As models and data grow, the multimodal training bottleneck increasingly comes from the surrounding input work rather than the training step itself: decoding, transformation, and tokenization contending with gradient computation for accelerator time.
At Ray Summit 2026, Minghao Xie and Haoran Cai from Luma AI present Koba, their multimodal data-loading framework, and Kojo, a Ray-based substrate that disaggregates stateless preprocessing stages onto managed actor pools while preserving the existing training loop. The result: compute that scales on its own pool independently of the trainer, fault isolation so a preprocessing failure no longer cascades into training, and reclaimed accelerator capacity returned to the gradient path.
You'll leave with a reusable pattern for relocating stateless, out-of-step computation off the training GPUs and onto Ray, plus the operational lessons of running it at scale.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com










