Uploaded November 2025 | Updated November 2026, 50 minutes ago
At Ray Summit 2025, Anindya Saha from Zoox presents the following:
Post-training workflows for vision-language models processing large-scale image datasets present coordination challenges among multiple processes and stages that often lead to complex, difficult-to-maintain systems. Inspired from real-world experience, we will talk and demo through a use case of image captioning to demonstrate how Ray Data, Ray Data LLM and vLLM simplify building production-scale post training workflows while achieving high throughput and resource efficiency.
We'll explore practical patterns on:
- How do we develop a workflow from prototype to production scale gradually
- Use Ray Data's distributed processing for efficient data loading, batching and transformation
- Use Ray Data LLM's Processor abstraction for seamless vLLM integration with preprocessing and postprocessing steps
- Integrate vLLM and scaling up inference on multiple GPUs for high-throughput
- Fully customize preprocess, postprocess steps; Manage states with classes.
- Integrate Prometheus & Grafana for real time performance monitoring
- Optimize GPU resource utilization across distributed workloads
Liked this video? Check out other Ray Summit breakout session recordings youtube.com/playlist?list=PLzTswPQNepXllnU0C36WtkC0dqkAoDulh
Subscribe to our YouTube channel to stay up-to-date on the future of AI! youtube.com/c/anyscale
🔗 Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
At Ray Summit 2025, Anindya Saha from Zoox presents the following:
Post-training workflows for vision-language models processing large-scale image datasets present coordination challenges among multiple processes and stages that often lead to complex, difficult-to-maintain systems. Inspired from real-world experience, we will talk and demo through a use case of image captioning to demonstrate how Ray Data, Ray Data LLM and vLLM simplify building production-scale post training workflows while achieving high throughput and resource efficiency.
We'll explore practical patterns on:
- How do we develop a workflow from prototype to production scale gradually
- Use Ray Data's distributed processing for efficient data loading, batching and transformation
- Use Ray Data LLM's Processor abstraction for seamless vLLM integration with preprocessing and postprocessing steps
- Integrate vLLM and scaling up inference on multiple GPUs for high-throughput
- Fully customize preprocess, postprocess steps; Manage states with classes.
- Integrate Prometheus & Grafana for real time performance monitoring
- Optimize GPU resource utilization across distributed workloads
Liked this video? Check out other Ray Summit breakout session recordings youtube.com/playlist?list=PLzTswPQNepXllnU0C36WtkC0dqkAoDulh
Subscribe to our YouTube channel to stay up-to-date on the future of AI! youtube.com/c/anyscale
🔗 Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com










