llama.cpp NUMA mirroring boosts dual-EPYC inference by up to 137%
mattescala · reddit · 2026-08-30
Addressing the dual-socket bottleneck where cores read weights across the slow interconnect, the author enabled the unused NUMAMIRROR strategy in ggml-cpu.h. This keeps a full copy of weights on each NUMA node, doubling RAM usage but allowing local access. Benchmarks show 64-71% speedups for DeepSeek-V4 and GLM-5.2, and a massive 137% boost for gemma-4-31B dense decoding. A PR has been submitted to llama.cpp.
More from Infra
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16
- Google open-sources AX, a Kubernetes-style orchestrator for billions of AI agent tasks — thisdudelikesAI · 2026-10-16