Skip to content

workaround: ROCm errors/crashes with ComfyUI and diffusion models #15

Description

@khimaros

edit: see the simpler fix in #15 (comment)

sharing this as it had a big impact on my own workflow

the crashing issues are still present until the release of 6.18 kernel

in fact, they now result in kernel crash with later firmware and kernels

however, there is a workaround!

if you restart the torch kernels after each ~50 steps, there is no crash

but that's annoying to do and breaks your asset folder in the UI, etc.

so instead, i've modified ComfyUI to use a worker queue with a new process for each prompt

this completely alleviated all crashes for me even on kernel 6.17

Comfy-Org/ComfyUI#11143

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