Offloading only "hot" MoE experts to VRAM boosts llama.cpp throughput 50%
nbvehrfr · reddit · 2026-08-29
A developer forked llama.cpp with a twist: when a MoE model doesn't fit entirely in VRAM (here Qwen 3.8 Flash Next), instead of offloading whole layers, only the "hot" experts go to VRAM—raising throughput from 20 t/s to 30 t/s (+50%).
Key observation: certain groups of experts stay relatively stable across coding, refactoring and code-review workloads, making them worth pinning. PR: github.com/timadinorth/llama.cpp/pull/1.
Caveats: tested only on coding workloads, and only useful when the full model exceeds VRAM. The author doubts upstream will accept it—the low-level implementation was written by Opus, and he doesn't feel like defending every line.
More from Infra
- Next-gen GraalVM benchmarks: 4x faster startup, 2x less memory, no build step needed — RealGeneKim · 2026-10-15
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark AI laptops — tomwarren · 2026-10-15
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark laptops — The Verge AI · 2026-10-15
- Global smartphone shipments plunge 22% in worst quarterly decline on record as AI demand fuels memory costs — Polymarket · 2026-10-15
- SK Hynix drops 11% again, now below 3x FY2027 earnings as AI memory selloff deepens — JOBhakdi · 2026-10-15
- ASML Sells Zero Lithography Machines into Europe as US Hosts 75%+ of AI Compute — NinaDSchick · 2026-10-15