diffusers-rocm

Portable builds of HuggingFace diffusers with AMD ROCm 7.12 acceleration. Each release is a self-contained archive that bundles a relocatable CPython interpreter, diffusers + transformers + accelerate, PyTorch ROCm, the required ROCm user-space libraries (as pip packages), and a small OpenAI-compatible HTTP server (diffusers-server) — no system Python, PyTorch, or ROCm install required. Automated pipeline targets integration with Lemonade.

Important

Early Development: This project is in active development. ROCm support for consumer AMD GPUs (RDNA) is still maturing. We welcome issue reports and contributions.

Supported Devices

GPU Target	Architecture	Devices
gfx1151	STX Halo APU	Ryzen AI MAX+ Pro 395
gfx1150	STX Point APU	Ryzen AI 300
gfx120X	RDNA4 GPUs	RX 9070 XT, RX 9070, RX 9060 XT, RX 9060
gfx110X	RDNA3 GPUs	RX 7900 XTX/XT/GRE, RX 7800 XT, RX 7700 XT, RX 7600 XT/7600

All builds include ROCm 7.12 user-space built-in — no separate ROCm installation required. You still need a Linux kernel with a working amdgpu driver for your GPU; for gfx1151 specifically this means kernel 6.18.4+ (see Lemonade's gfx1151 notes).

Quick Start

1. Download the build for your GPU from the latest release. Larger releases are split into .partNN-of-MM.tar.gz parts because they exceed GitHub's 2 GB per-asset limit.
2. Extract:

   mkdir -p ~/diffusers-rocm
   # Single-archive releases:
   tar xzf diffusers0.35.0-rocm7.12.0-gfx1151-x64.tar.gz -C ~/diffusers-rocm
   # Multi-part releases:
   cat diffusers0.35.0-rocm7.12.0-gfx1151-x64.part01-of-02.tar.gz \
       diffusers0.35.0-rocm7.12.0-gfx1151-x64.part02-of-02.tar.gz \
     | tar xz -C ~/diffusers-rocm

3. Run the server:

   ~/diffusers-rocm/bin/diffusers-server \
     --model unsloth/FLUX.1-schnell-GGUF \
     --port 8000

4. Generate an image:

   curl http://localhost:8000/v1/images/generations \
     -H "Content-Type: application/json" \
     -d '{
       "model": "unsloth/FLUX.1-schnell-GGUF",
       "prompt": "a red apple on a wooden table",
       "n": 1,
       "size": "1024x1024"
     }'

Lemonade Integration: These builds are designed to work as a backend for Lemonade, which manages downloading, launching, and routing requests automatically.

What's Included

Each release archive extracts to a relocatable CPython 3.12 distribution:

bin/
  diffusers-server            # Launcher shim (sets LD_LIBRARY_PATH, execs the HTTP service)
  python3.12                  # Bundled CPython (python-build-standalone)
lib/
  libpython3.12.so
  python3.12/
    site-packages/
      diffusers/              # pip from PyPI
      transformers/           # pip from PyPI
      accelerate/, peft/, safetensors/, gguf/, ...
      torch/                  # pip from repo.amd.com/rocm/whl/<arch>/
      _rocm_sdk_core/lib/     # ROCm core user-space (hip, hsa, comgr, clang, llvm)
      _rocm_sdk_libraries_gfx<arch>/lib/
                              # Per-arch ROCm math libs (rocblas, hipblas, MIOpen, ...)
      diffusers_server/       # The OpenAI-compatible HTTP server in this repo

API

diffusers-server exposes OpenAI-compatible endpoints:

• POST /v1/images/generations — text-to-image (sync)
• GET /health — liveness probe
• (planned phase 2) POST /v1/videos/generations + GET /v1/videos/generations/{id} — async video gen
• (planned phase 3) POST /v1/images/edits, POST /v1/images/variations

Automated Builds

The GitHub Actions workflow:

• Downloads relocatable CPython 3.12 from astral-sh/python-build-standalone
• Installs PyTorch ROCm from AMD's per-arch pip index (https://repo.amd.com/rocm/whl/<target>/), which pulls rocm-sdk-core and rocm-sdk-libraries-gfx<target> as transitive deps
• Installs diffusers + transformers + accelerate + peft + safetensors + gguf + huggingface_hub + Pillow + fastapi + uvicorn from PyPI
• Installs this repo's diffusers-server package into site-packages
• Generates a bin/diffusers-server shim that wires up LD_LIBRARY_PATH at startup
• Tars the result, splits into < 2 GB parts, uploads to GitHub Releases

Dependencies

Runtime (bundled)

• diffusers — HuggingFace diffusion model library (pure Python)
• PyTorch — tensor compute (ROCm wheel)
• ROCm SDK wheels — AMD's pip-packaged ROCm user-space
• python-build-standalone — relocatable CPython 3.12

Build (CI only)

• Ubuntu 22.04 GitHub Actions runner
• pip (no compilation, no cmake, no patchelf — everything is pre-built wheels)

License

MIT — see LICENSE.