FLUX.2 AMD GPU Server Setup | Generated by AI

Home 2026.07 English 中文

---

✅ Comprehensive Summary

Server Setup

Step	Status
SSH connection test	✅ Connected
GPU detection (ROCm)	✅ AMD Radeon GPU, 48GB VRAM
tmux installed	✅
Python venv created (/root/flux_env)	✅

GPU Specs

• GPU: AMD RDNA 3 (gfx1100) — 48GB VRAM (Samsung)
• ROCm: v7.2.1, HIP 7.2.53211
• CPU: AMD EPYC 9334, 128 threads
• RAM: 503 GB
• Disk: 3.5T (2.9T free)

Software Installed

Package	Version	Notes
PyTorch	2.13.0+rocm7.2	ROCm 7.2 compatible
diffusers	0.39.0	Latest
transformers	4.57.6	Downgraded from 5.x for compatibility
accelerate	1.14.0	For CPU offloading
sentencepiece	0.2.2	Tokenizer
huggingface-hub	logged in with token	HF mirror configured

Issues Encountered & Fixes

1. No network to huggingface.co → Used HF_ENDPOINT=https://hf-mirror.com (Tsinghua mirror)

2. Transformers 5.x incompatible with diffusers → Downgraded to transformers==4.57.6

3. Wrong pipeline class → FLUX.2-dev uses Flux2Pipeline (not FluxPipeline)

4. OOM on pipe.to("cuda") → Model ~35GB in bf16, tried to load all at once into 48GB VRAM → Fixed with pipe.enable_model_cpu_offload() (keeps only one component on GPU at a time)

Final Script (/root/run_flux.py)

• Loads FLUX.2-dev with CPU offloading
• Generates “A cute cat holding a sign that says hello world” at 1024×1024
• 28 inference steps
• Saves result to test_flux.png

How to Run

ssh -p 31145 root@36.150.116.206
tmux new -s flux
source /root/flux_env/bin/activate
python3 /root/run_flux.py

Detach with Ctrl+B, D · Re-attach with tmux attach -t flux

---

Generated by AI. Curating and sharing still takes effort. If you find it useful, feel free to donate. WeChat: @lzwjava · X: @lzwjava · Say hi 👋

Back Donate