Uploaded September 2026 | Updated November 2026, 2 days ago
A simpler setup delivered more tokens per second, with the first token arriving sooner for a single request, than the model card's reference configuration.
At Ray Summit 2026, Eyal Lantzman, VP of AI Services at Nscale, uses DeepSeek-V4-Flash, Ray, and vLLM to explore the practical tradeoffs between scaling strategies, including tensor and expert parallelism, prefill/decode disaggregation, and whole-replica scale-out, plus the production caveats that matter at scale, from GPU topology and cache pressure to long context and speculative decoding.
You'll leave with what worked, what didn't, and how to choose a scaling strategy for frontier open models.
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
A simpler setup delivered more tokens per second, with the first token arriving sooner for a single request, than the model card's reference configuration.
At Ray Summit 2026, Eyal Lantzman, VP of AI Services at Nscale, uses DeepSeek-V4-Flash, Ray, and vLLM to explore the practical tradeoffs between scaling strategies, including tensor and expert parallelism, prefill/decode disaggregation, and whole-replica scale-out, plus the production caveats that matter at scale, from GPU topology and cache pressure to long context and speculative decoding.
You'll leave with what worked, what didn't, and how to choose a scaling strategy for frontier open models.
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










