Achieving 253 t/s on RTX 5090: Optimizing Muse Glimmer 30B Inference

patricious · reddit · 2026-08-11

The author benchmarked the Muse Glimmer 30B model (Q5KM quantization) on an RTX 5090 (32GB). By applying PR #26842, which moves the DFlash draft argmax from CPU to GPU, they successfully resolved a key bottleneck.

With this optimization, inference speed for code generation jumped from 78 t/s to 220-253 t/s, and mixed agent tasks reached 188-213 t/s. The post includes detailed launch flags and notes that reasoning budget flags currently do not work with this template.

Related event: RTX 5090 runs 30B model inference at over 233 tps in tests(3 posts)→

Original post →

More from Infra

Infra channel →