[Question]: Deployed worker using my own Docker Image gets stuck in queue on RunPod and never executes (Custom Nodes)

[billyberkouwer]

I am trying to use this worker with custom nodes. When I deploy using your docker image, all works smoothly on RunPod as expected. However, when I fork the repo, create a local image, then push and deploy from my own repo, the worker deploys but my requests get stuck in the queue and never execute.

I have followed the documentation for editing the docker image to use custom nodes. However, this issue occurs even when I am using an unedited version of the Dockerfile.

I am sure that I am missing something but I am unsure why even an unedited copy of this repo produces an image that doesn't work when the image you have supplied does.

My process:

1. Clone repo
2. Run command from directory: e.g. docker build --build-arg MODEL_TYPE=sdxl -t billymylesberkouwer/runpod-worker-comfy:dev-sdxl --platform linux/amd64 .
3. Push image to my hub
4. Create a new RunPod template using billymylesberkouwer/runpod-worker-comfy:dev-sdxl
5. Create a new Serverless using this template
6. Use contents of test_input.json to create a request using RunPod request dashboard

Result:

An endless delay, unhealthy workers, no execution

Expected result:

A success or failure message and successful execution

Any help would be greatly appreciated. Many thanks in advance!

[billyberkouwer]

I have found that using this approach suggested by ecornbower has successfully deployed and is returning succesfully using test_input.json. I would still appreciate it if you/anyone else can shed some light on why the approach I took following the documentation is failing it would still be much appreciated!

[TimPietrusky]

@billyberkouwer thanks for reporting this. We will have to dig a bit deeper to make sure to provide a valid response. And also there is also this PR #30 open which should make custom nodes working for everyone. The thing is just time, but we will make sure to include this, as this is super important for a lot of people.

[billyberkouwer]

@TimPietrusky Many thanks for getting back to me on this! This PR looks like a great solution. I'm currently working on another aspect of my project but will comment on the PR to feedback if I end up adopting it in production. For now, a reduction in the number of nodes and some additional installs and updates based on errors that were thrown seem to have provided a solution. But the ability to use ComfyUI manager snapshots and the full range of nodes sounds preferable