ERNIE-Image-Turbo on NucBox EVO X2 (gfx1151)

Setup for running Baidu's ERNIE-Image-Turbo (text-to-image, 8B DiT) on AMD Ryzen AI MAX+ 395 (Strix Halo, gfx1151) via diffusers.

Hardware / Environment

Setup gotchas

1. The official PyTorch ROCm wheel does not work on gfx1151

The official PyTorch nightly installed via pip install torch --index-url https://download.pytorch.org/whl/nightly/rocm6.4 does not include gfx1151 in its arch list:

arch list: ['gfx900', 'gfx906', 'gfx908', 'gfx90a', 'gfx942',
            'gfx1030', 'gfx1100', 'gfx1101', 'gfx1102', 'gfx1200', 'gfx1201']
gcnArchName: gfx1151

Even with HSA_OVERRIDE_GFX_VERSION=11.5.1, the matching kernels are absent from the wheel, so it crashes at runtime with HIP error: invalid device function.

Solution: use AMD TheRock's gfx1151-native nightly wheel.

pip install --index-url https://rocm.nightlies.amd.com/v2/gfx1151/ torch torchvision torchaudio

This installs torch-2.10.0+rocm7.13.0 together with rocm-sdk-libraries-gfx1151, and runs natively with arch list: ['gfx1151'].

2. ErnieImagePipeline API change

pipe.enable_vae_slicing() / pipe.enable_vae_tiling(), mentioned in the initial CLAUDE.md, have been removed in recent diffusers and moved onto the VAE module:

# Old (no longer works)
pipe.enable_vae_slicing()
pipe.enable_vae_tiling()

# New
pipe.vae.enable_slicing()
pipe.vae.enable_tiling()

3. AOTriton Flash Attention is gated behind an experimental flag

Setting TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 enables a Flash Attention–equivalent path on AMD GPUs through SDPA. In this setup it gave about a 2.5× speedup.

4. cpu_offload is unnecessary

With 48GB of VRAM, ERNIE-Image-Turbo (peak ~37GB at 1024x1024) fits entirely on the GPU. Replacing enable_model_cpu_offload() with pipe.to("cuda") yields another ~2× speedup.

Install

bash install.sh

After it finishes, log into Hugging Face (in a separate terminal is fine):

source venv/bin/activate
hf auth login --token <HF_TOKEN>   # get one at https://huggingface.co/settings/tokens

Run

source venv/bin/activate
python run_ernie.py

The first run downloads the model (~80GB) from Hugging Face, so make sure you have the disk space and time (about 5 minutes on a fast connection).

Final configuration (run_ernie.py)

• gfx1151-native PyTorch 2.10 + ROCm 7.13 (TheRock)
• AOTriton Flash Attention enabled (TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1)
• Full model resident on GPU (pipe.to("cuda"))
• VAE slicing/tiling enabled
• bfloat16, 8 steps

Performance (640x640 vs 1024x1024)

Measured at 8 steps, guidance_scale=1.0, bf16.

End-to-end this is about a 4.3× speedup over the initial config. 31s per 1024x1024 image is well ahead of the 1–3 min/image estimate in CLAUDE.md.

Prompts and images

The prompt and image size are written directly in run_ernie.py.

Prompt 1

Zundamon walking through Akihabara's electric town, anime style, bright colors, evening.

Prompt 2

Kinkaku-ji glowing in front of Mt. Fuji with a pond in the foreground, photographic style, bright colors, daytime.

References

• AMD TheRock - gfx1151 PyTorch wheels
• ERNIE-Image-Turbo on Hugging Face
• Strix Halo gfx1151: 93 ML experiments (ROCm/ROCm Issue #6034)