Our Q4 multi-hardware roadmap is out. SGLang now runs on 10+ accelerator families, and the plugin system makes adding a new backend out-of-tree a first-class path.
In-tree backends, Q4 focus:
🔸 NVIDIA: collaboration roadmap refresh for Q4
🔸 AMD: MI45x, Helios rack-scale and Ryzen AI Halo/Radeon enablement; monokernels/megakernels, distributed serving, speculative decoding, HiCache, diffusion/Omni and Miles RL support.
🔸 Google TPU: Day-0 frontier models, model optimization on agentic workload, diffusion serving, RL rollout on TPU, feature compatibility, production reliability; the plugin runs on upstream SGLang
🔸 Ascend NPU: closing out the H2 milestones
🔸 Intel Xeon CPU and XPU: backend optimization tracks for Q4
🔸 T-Head PPU / Cambricon MLU: backend bring-up and roadmap tracking
New backends coming through the plugin:
🔸 AWS Trainium / MetaX
Plugin and platform work:
🔸 Capability-based device gating for out-of-tree platforms
🔸 Robust speculative decoding and broader quantization support honored by plugins
🔸 Production serving audit across backends