These efficiency gains enable the Miles platform by radixark to execute full reinforcement learning loops on a single 4-GPU tray using Megatron training and SGLang rollouts in one container. This setup increased the GSM8K reward for Qwen3-30B-A3B from roughly 45% to 95% over 50 rollouts. The Vera CPU also supports 64 concurrent mini-SWE-agent sandboxes for agentic RL with Qwen3.5-35B-A3B, while the system supports end-to-end rollout and training for DeepSeek-V4-Flash.
Tuesday, October 20, 2026
Story
SGLang Speeds Kimi K3 Inference by 20% on NVIDIA Vera Rubin
AI11 days ago
Kernel optimization for MoE tail fusion removed 276 kernel launches per decode step, creating a 5.9% speedup in end-to-end inference on next-generation chips. The SGLang project, working with NVIDIA on early-access Vera Rubin hardware, achieved up to 20% faster FP8 MLA at batch 1 with a 128K context for the Kimi K3 model. Other gains include 20% faster KDA verification with bit-identical output and a 16% speedup for batch 16, made possible by Rubin's 327 KiB of shared memory per CTA.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.