vLLM ships Day-0 RL support for Beam in SkyRL, +62% LoRA decode throughput
vllm_project · x · 2026-10-16
vLLM announced Day-0 RL support for Reflection's Beam via SkyRL: LoRA, BF16 full fine-tuning and MXFP8 RL with vLLM rollouts on Reflection's Megatron-Bridge. Right-sizing expert adapters in vLLM lifts LoRA decode throughput from 276 to 447 tok/s/GPU (+62%) at 32 concurrent requests, PP2×TP8 (decode-only metric). Initial DAPO validation shows AIME25 avg@8 rising 61%→79% in 50 steps (LoRA: 4×8-B200; full FT: 8×8-B200). R3 works with pipeline parallelism; FP32 router weights preserved during weight sync.
Related event: vLLM and SkyRL Deliver Day-0 Support for Reflection Beam(4 posts)→
More from coding & agent
- open-science: open-source local-first AI research workbench hits 5k GitHub stars — rohanpaul_ai · 2026-10-17
- As agents get more capable, should the harness take on more or less? — sarahcat21 · 2026-10-17
- Intent Mobile brings large-scale agent orchestration to the Duo foldable: manage agents across laptops, desktops and VMs — LukeW · 2026-10-17
- PSMcpServer: Drop a PowerShell Script in a Folder and It Becomes a Live MCP Tool — dfinke · 2026-10-17
- Claude controls a Second Life avatar via MCP, working out joint poses itself — slhamlet · 2026-10-16
- CogniCore MCP server gives ChatGPT and Claude one shared memory — Neither-Witness-6010 · 2026-10-16