Issue getting imagilux/fish-speech running on a 9070 XT

[pete-h]

Hi,

I'm try to get imagilux/fish-speech to run on my 9070 XT.

I could not find fishaudio/s2-pro-int8 so I used https://huggingface.co/drbaph/s2-pro-fp8/tree/main instead (I hope that's fine).

I have the docker containers up and running, but I try to generate even a single word it doesn't work.

I hope somebody has an idea :)

Thanks
Pete

[solitario001]

Hi,

I'm try to get imagilux/fish-speech to run on my 9070 XT.

I could not find fishaudio/s2-pro-int8 so I used https://huggingface.co/drbaph/s2-pro-fp8/tree/main instead (I hope that's fine).

I have the docker containers up and running, but I try to generate even a single word it doesn't work.

I hope somebody has an idea :)

Thanks Pete

Hi Pete! I have the exact same GPU (9070 XT) and managed to fully solve this deadlock and the Out of Memory (OOM) crash inside the Docker setup.

The issue is that the default .env template doesn't expose the correct low-level render permissions to the container, forcing PyTorch into weird fallbacks and severe memory fragmentation on RDNA 4 architecture.

Here is the exact step-by-step fix:

1. Fix Host Permissions inside .env

Open your .env file and look for RENDER_GID. It is usually commented out by default. You need to uncomment it and set it to your host's actual render group ID:

• On your Linux host terminal, run: stat -c '%g' /dev/kfd
• Take that number and update your .env along with these memory stability variables:

RENDER_GID=YOUR_HOST_NUMBER_HERE
MAX_SEQ_LEN=1024
VRAM_FRACTION=0
PYTORCH_HIP_ALLOC_CONF="expandable_segments:True,garbage_collection_threshold:0.7,max_split_size_mb:2048"

Setting expandable_segments:True is critical to prevent the exact 1GB block fragmentation crash shown in your logs.

2. Isolate your Device (Optional - Multi-GPU setups only)

The default mapping (- /dev/kfd:/dev/kfd and - /dev/dri:/dev/dri) works fine out of the box. However, if you are running a multi-GPU system and experience bus conflicts or freezes during decode, go to your compose.rocm.yml and explicitly isolate your primary card path under services -> server -> devices:

    devices:
      - /dev/kfd:/dev/kfd
      - /dev/dri/card1:/dev/dri/card1
      - /dev/dri/renderD128:/dev/dri/renderD128

3. Clear Cache and Rebuild

Run a clean deployment forcing Docker to read the new system tables:

sudo docker compose -f compose.yml -f compose.rocm.yml --profile webui down -v
sudo docker compose -f compose.yml -f compose.rocm.yml --profile webui up --build --force-recreate

Once it boots, your 9070 XT will be fully recognized, allocating ~6.5GB to 6.9GB of VRAM safely without dropping dead on generation. Hope this helps!