llama.cpp PR Caches Hot MoE Experts on GPU, Doubling Inference Speed on 8GB VRAM

BTA_Labs · reddit · 2026-08-05

A new llama.cpp PR (#26563) introduces a heatmap to track the most frequently used MoE experts. Instead of keeping all experts on the GPU or offloading them entirely to the CPU, it caches these "hot" experts in VRAM while cold experts continue running on the CPU.

The author's tests on Qwen3.6-35B-A3B with 8GB VRAM showed significant improvements:

However, this isn't a universal fix: models with low expert reuse like Qwen3.5-122B-A10B were actually slower due to tracking and cache-management overhead. Currently CUDA-only and active only during single-token decoding, this approach offers a promising direction for running large MoE models on consumer GPUs without relying on extreme quantization.

Original post →

More from Infra

Infra channel →