Skip to content

Unload model from memory/GPU when idle for set period #4141

@jpm-canonical

Description

@jpm-canonical

This issue was also reported by someone else here: openvinotoolkit/openvino#33665

With the current implementation it is not possible to use the GPU for other workloads while OVMS is running, as it will run out of free vRAM. Other LLM servers approach this by setting a user-defined idle timeout on a model. If no API requests are made to query a specific model for a specific amount of time, that model gets removed from memory. When a new request comes in for the model, it automatically gets loaded into memory again. It would be nice if OVMS had a similar feature.

An explanation of how llama.cpp does it can be found at https://github.com/ggml-org/llama.cpp/tree/master/tools/server#sleeping-on-idle

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions