killown/f5rocm repository overview

⁠killown/f5rocm: F5-TTS for AMD ROCm

A high-performance, containerized environment for F5-TTS optimized for AMD Instinct and Radeon GPUs. This image handles the hardware-level complexity of ROCm and bypasses Python 3.12's "externally-managed" restrictions by injecting dependencies into the system-wide site-packages.

⁠Quick Start (RDNA 4 / 9060 XT)

Run this from your local project root. This command pulls the image from Docker Hub, maps your current directory to /app, and provides full GPU access using the gfx1201 override.

docker run -d
--name f5-tts-rocm
--network host
--privileged
--device=/dev/kfd
--device=/dev/dri
--shm-size=8gb
-v $(pwd):/app
-e HSA_OVERRIDE_GFX_VERSION=12.0.1
-e PYTORCH_ROCM_ARCH=gfx1201
killown/f5rocm:latest

⁠GPU Compatibility Matrix

ROCm 7.0+ natively supports RDNA 4 architectures. Use the following overrides for optimal performance.

⁠Technical Architecture

• System-Path Injection: Dependencies (click, gradio, cached_path) are installed in /usr/lib/python3.12/. This prevents Volume Shadowing, ensuring that mounting your local source code over /app does not hide the required Python modules.
• Full Stack OS Deps: Includes build-essential, openssl, libssl-dev, and the complete RDMA/InfiniBand suite for low-latency hardware communication.
• Audio Processing: Pre-configured with ffmpeg and sox (including libsox-fmt-mp3) for high-fidelity inference and reference audio handling.

⁠Verification

Check if the GPU is correctly mapped and visible to the PyTorch ROCm bridge:

docker exec -it f5-tts-rocm rocm-smi docker exec -it f5-tts-rocm python3 -c "import torch; print(f'ROCm version: {torch.version.hip}'); print(f'GPU Available: {torch.cuda.is_available()}')"

⁠License

This Docker image is provided for technical and research purposes. Check the F5-TTS repository for the base model license.