Skip to content

Proposal: on-demand model sleep/wake — offload weights to host RAM to free VRAM #228

Description

@yadiel-tien

My setup

I run ninfer-serve (qwen3.8-27b NVFP4, ~29 GB VRAM on an RTX 5090 D) as the
local brain for an agent that also drives ComfyUI for image generation on
the same GPU. Only one of the two can hold the VRAM at a time.

The pain

Today the handoff means killing the server and restarting it: ~40–60 s cold
start each way (24 GB of weights from NVMe + re-init). vLLM's sleep mode
(weights to pinned host RAM, ~1–2 s each way) is exactly what I'd want.

Request

An on-demand unload/wake for the model weights — free the VRAM while the
process stays up, re-load on demand. KV/state can just be dropped, it is
rebuilt.

Happy to help test and provide details on the use case.

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

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions