Skip to content

Tuning llama.cpp / Ollama for faster, larger local inference #18

Description

@rlnorthcutt

The question

What's the best way to tune local inference (llama.cpp and/or Ollama) for faster operation and larger models — including MoE offload — on the hardware a typical user's container will actually have?

Background

Ships Ollama in the container today, but first-class local inference (see that epic) needs local inference to actually be fast and capable enough to be worth using, not just present.

How we'll answer it

Compare llama.cpp against Ollama for efficiency; investigate MoE offload and other techniques for running larger models on constrained hardware. Likely output is a guide/video, and possibly a tool or skill that applies tuned settings automatically.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    roadmapTier 1 item — triggers auto-add to the org project

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions