Release v0.7.0 · ggml-org/llama.cpp
github.com
Overview llama.cpp v0.7.0 adds the LiliCorr speculative decoding technique and MTP for GLM5-Next, a new GPU cache for MoE experts kept in host memory (--moe-cache-mib, up to ~2.2x decode speedup wh...

--moe-cache-mib doesn’t seem optimized for tight VRAM situations or if your PCIe link gets saturated by it, but it improves performance on some setups.

  • 9tr6gyp3@lemmy.worldEnglish
    1·
    6 days ago

    Ah, i must have missed it. Thanks for the info 🙂

    Hmm, not seeing it directly in the llama.cpp 0.6.0 version. Where were you seeing it?

    • trewq@lemmy.worldEnglish
      2·
      6 days ago

      Should be added sometime early this year. I think it was mentioned in HF blog/article when i was looking for way to use multiple models without restarting llama