[bug]: ROCM Docker container not using AMD GPU

[Malbeth81]

Is there an existing issue for this problem?

• I have searched the existing issues

Install method

Manual

Operating system

Linux

GPU vendor

AMD (ROCm)

GPU model

RX 7900 XTX

GPU VRAM

24GB

Version number

main-rocm

Browser

No response

System Information

No response

What happened

I've been trying to install the application on TrueNAS Scale via the Docker container, following the steps from https://github.com/invoke-ai/InvokeAI/blob/main/docker/README.md, but it will always only use the CPU and the the GPU, which is passed through (kfd & dri devices), which works fine for other apps.

These are the logs generated by the Docker application:

2026-08-10 14:00:44.212740+00:00Could not detect ROCm GPU architecture: [Errno 2] No such file or directory: 'rocminfo'
2026-08-10 14:00:44.213089+00:00Could not detect ROCm warp size: [Errno 2] No such file or directory: 'rocminfo'. Defaulting to 64. (some 4-bit functions may not work!)
2026-08-10 14:00:44.215226+00:00Failed to load CPU gemm_4bit_forward from kernels-community: No module named 'kernels'. Please make sure you already pip install kernels and the kernels >= 0.11.1
m[2026-08-10 10:00:47,901]::[InvokeAI]::INFO --> Using torch device: CPU
m[2026-08-10 10:00:47,902]::[InvokeAI]::INFO --> cuDNN version: 3005001
m[2026-08-10 10:00:48,285]::[InvokeAI]::INFO --> Patchmatch initialized
2026-08-10 14:00:48.553379+00:00Siglip2ImageProcessorFast is deprecated. The Fast suffix for image processors has been removed; use Siglip2ImageProcessor instead.
[38;20m[2026-08-10 10:00:50,230]::[InvokeAI]::INFO --> InvokeAI version 6.14.0-alpha
[38;20m[2026-08-10 10:00:50,230]::[InvokeAI]::INFO --> Root directory = /invokeai
[38;20m[2026-08-10 10:00:50,231]::[InvokeAI]::INFO --> Initializing database at /invokeai/databases/invokeai.db
[38;20m[2026-08-10 10:00:50,249]::[InvokeAI]::INFO --> JWT secret loaded from database
[38;20m[2026-08-10 10:00:50,251]::[ModelManagerService]::INFO --> [MODEL CACHE] Calculated model RAM cache size: 12817.33 MB. Heuristics applied: [1].
[38;20m[2026-08-10 10:00:50,251]::[ModelManagerService]::INFO --> Model cache global RAM budget: 12.52 GB across 1 device cache(s).
[38;20m[2026-08-10 10:00:50,252]::[ModelInstallService]::INFO --> Restoring incomplete installs
[38;20m[2026-08-10 10:00:50,253]::[ModelInstallService]::INFO --> Finished restoring incomplete installs
[38;20m[2026-08-10 10:00:50,315]::[InvokeAI]::INFO --> Invoke running on http://0.0.0.0:9090 (Press CTRL+C to quit)

It seems the Dockerfile no longer installs rocm like it used to https://github.com/invoke-ai/InvokeAI/blob/v5.2.0/docker/Dockerfile#L41, so I don't get how it is supposed to function without it? Why keep publishing an image that is essentially broken and useless?

What you expected to happen

I expect the application to run in Docker using the image with the main-rocm tag and detect and use the GPU to run the models.

How to reproduce the problem

Simply run docker run --device /dev/kfd --device /dev/dri --publish 9090:9090 ghcr.io/invoke-ai/invokeai:main-rocm on a machine with an AMD GPU.

Additional context

No response

Discord username

Malbeth

[Pfannkuchensack]

The error Shows that is cant find anything rcom related so it fallbacks to CPU. The docker file uses https://github.com/invoke-ai/InvokeAI/blob/main/docker%2FDockerfile#L80 to do the selection of the right packages.

[Malbeth81]

I do have GPU_DRIVER set to "rocm" in my Docker container environment (compose), but that line you mentioned is only used when building the image from the Dockerfile, when using the pre-built image published to Github, I would expect those dependencies to be already within it, as you shown in the Dockerfile.

[Pfannkuchensack]

If you are using ROCM, you will need to ensure that the render group within the container and the host system use the same group ID. > To obtain the group ID of the render group on the host system, run getent group render and grab the number.

is that something you can check?

[Malbeth81]

Sure. 👍

$ getent group render
render:x:107:

Guessing this is not the one used in the Docker image, will try to set it later today and update you on the results. Thanks! 😄

[alex-mashin]

It seems that, with very exceptions, any application claiming to support ROCM, is bugged and unusable.

[alex-mashin]

Sure. 👍

$ getent group render
render:x:107:

Guessing this is not the one used in the Docker image, will try to set it later today and update you on the results. Thanks! 😄

Setting RENDER_GROUP_ID to correct id of render group does not help, even with group_add: -video -render.

It seems that the Docker image does not contain the program rocminfo. Neither Dockerfile nor .github/workflows/build-container.yml do mention it. Is GitHub supposed to provide it?

[Malbeth81]

It worked for me, InvokeAI now sees and uses the GPU!

2026-08-12 12:54:23.668017+00:00Could not detect ROCm GPU architecture: [Errno 2] No such file or directory: 'rocminfo'
2026-08-12 12:54:23.668071+00:002026-08-12T12:54:23.668071296Z
2026-08-12 12:54:23.668083+00:00ROCm GPU architecture detection failed despite ROCm being available.
2026-08-12 12:54:23.668093+00:002026-08-12T12:54:23.668093376Z
2026-08-12 12:54:23.668591+00:00Could not detect ROCm warp size: [Errno 2] No such file or directory: 'rocminfo'. Defaulting to 64. (some 4-bit functions may not work!)
2026-08-12 12:54:23.668624+00:002026-08-12T12:54:23.668624613Z
2026-08-12 12:54:23.668636+00:00ROCm warp size detection failed despite ROCm being available.
2026-08-12 12:54:23.668647+00:002026-08-12T12:54:23.668647583Z
[2026-08-12 08:54:28,009]::[InvokeAI]::INFO --> Using torch device: [AMD Radeon RX 7900 XTX (cuda:0), AMD Radeon 760M Graphics (cuda:1)]
[2026-08-12 08:54:28,011]::[InvokeAI]::INFO --> cuDNN version: 3005001
[2026-08-12 08:54:28,633]::[InvokeAI]::INFO --> Patchmatch initialized
Siglip2ImageProcessorFast is deprecated. The Fast suffix for image processors has been removed; use Siglip2ImageProcessor instead.
[2026-08-12 08:54:30,912]::[InvokeAI]::INFO --> InvokeAI version 6.14.0-alpha
[2026-08-12 08:54:30,912]::[InvokeAI]::INFO --> Root directory = /invokeai
[2026-08-12 08:54:30,913]::[InvokeAI]::INFO --> Initializing database at /invokeai/databases/invokeai.db
[2026-08-12 08:54:30,939]::[InvokeAI]::INFO --> JWT secret loaded from database
[2026-08-12 08:54:30,940]::[ModelManagerService]::INFO --> [MODEL CACHE] Calculated model RAM cache size: 12817.33 MB. Heuristics applied: [1].
[2026-08-12 08:54:30,940]::[ModelManagerService]::INFO --> [MODEL CACHE] Calculated model RAM cache size: 11793.33 MB. Heuristics applied: [1, 2].
[2026-08-12 08:54:30,940]::[ModelManagerService]::INFO --> Capping model cache RAM budget at 12.52 GB to leave system headroom (sum of per-device caches was 24.03 GB).
[2026-08-12 08:54:30,940]::[ModelManagerService]::INFO --> Model cache global RAM budget: 12.52 GB across 2 device cache(s).
[2026-08-12 08:54:30,942]::[ModelInstallService]::INFO --> Restoring incomplete installs
[2026-08-12 08:54:30,943]::[ModelInstallService]::INFO --> Finished restoring incomplete installs
[2026-08-12 08:54:30,943]::[InvokeAI]::INFO --> Starting session processor with 2 parallel workers on devices: cuda:0, cuda:1
[2026-08-12 08:54:30,997]::[InvokeAI]::INFO --> Invoke running on http://0.0.0.0:9090 (Press CTRL+C to quit)
[2026-08-12 08:54:51,305]::[InvokeAI]::INFO --> Executing queue item 3, session 8a06df1b-b36f-4d4c-8158-4baf276536b0 on cuda:0
Loading weights: 0%| | 0/398 [00:00<?, ?it/s]
Loading weights: 100%|██████████| 398/398 [00:00<00:00, 5614.20it/s]

However trying to get it to render something just causes the memory usage to spike to over 24GB and the app container crashes once it runs out...

Why is it using so much system memory if it runs on the GPU? Shouldn't it use the VRAM instead? My GPU has 24GB, but my TrueNAS host only has ~18GB free out of its 32GB memory with all the other services running.

Looking at the batch info, I notice it says "device": "cuda:0" but that seems to just be the identifier given to the AMD GPU. Still, here is the full batch info:

invokeai_batch.json

@alex-mashin Here is my Docker Compose YML file:

services:
  invokeai:
    devices:
      - /dev/dri:/dev/dri
      - /dev/kfd:/dev/kfd
    environment:
      AMD_VISIBLE_DEVICES: all
      GPU_DRIVER: rocm
      INVOKEAI_ROOT: /invokeai
      NVIDIA_VISIBLE_DEVICES: void
      RENDER_GROUP_ID: 107
      TZ: America/Montreal
    group_add:
      - 44
      - 107
      - 568
    hostname: invokeai
    image: ghcr.io/invoke-ai/invokeai:main-rocm
    platform: linux/amd64
    ports:
      - mode: ingress
        protocol: tcp
        published: 9090
        target: 9090
    restart: 'no'
    security_opt:
      - no-new-privileges=true
    volumes:
      - bind:
          create_host_path: False
          propagation: rprivate
        read_only: False
        source: /mnt/.ix-apps/app_mounts/invokeai/storage_entry
        target: /invokeai
        type: bind
volumes: {}
x-action-required: False
x-portals:
  - host: 0.0.0.0
    name: Web UI
    path: /
    port: 9090
    scheme: http

[alex-mashin]

Thank you for your input, @Malbeth81. I have also seen

invokeai-1  | Could not detect ROCm GPU architecture: [Errno 2] No such file or directory: 'rocminfo'
invokeai-1  | 
invokeai-1  | ROCm GPU architecture detection failed despite ROCm being available.
invokeai-1  |                 
invokeai-1  | Could not detect ROCm warp size: [Errno 2] No such file or directory: 'rocminfo'. Defaulting to 64. (some 4-bit functions may not work!)
invokeai-1  | 
invokeai-1  | ROCm warp size detection failed despite ROCm being available.

in my log.

By not working I meant that the server was not available at http://...:9090, but it appeared after I fixed a typo in my networks: clause.

Now I can download models from HuggingFace, he model seems to loaf fully into RAM, but image generation fails with [uvicorn.error]::ERROR --> Exception in ASGI application and KeyError: 'REQUEST_METHOD' in the logs.

[Malbeth81]

For the memory, I dunno why it doesn't use the VRAM, but by changing the resource limit for the Docker container and lowering the steps I was able to get it to generate an image, so the base issue has been fixed by setting the RENDER_GROUP_ID.

Maybe this is something you might want to consider adding to https://github.com/invoke-ai/InvokeAI/blob/main/docker/README.md 😉

Thank you @Pfannkuchensack !

[Pfannkuchensack]

The info is in the https://github.com/invoke-ai/InvokeAI/blob/main/docker/.env.sample file. But i dont really do docker stuff at all.