AMD Cloud GPU OOM Report | Generated by AI
Here’s a comprehensive summary of everything:
AMD Radeon Cloud Server — Complete Status
Server Specs
| Component | Value |
|---|---|
| Host | 36.150.116.206 |
| SSH Port | 31154 (currently down) |
| OS | Ubuntu 24.04.4 LTS |
| CPU | 2× AMD EPYC 9334 (128 threads) |
| RAM | 503 GiB |
| GPU | AMD Radeon Graphics (gfx1100 / RDNA 3) |
| VRAM | 48 GiB |
| ROCm | 7.2.1, Driver 6.16.13 |
| PyTorch | 2.9.1+git (ROCm 7.2 in /opt/venv) |
| Disk | / = 3.5TB overlay (2.8TB free), /workspace = 98GB loop (19GB free) |
GPU Compute Test
✅ Passed — HIP vector add (1M elements) computed correctly, all 48GB VRAM accessible, 96 CUs, 1760 MHz clock.
What We Did
- Diagnosed connectivity — Found SSH port 31094 wrong; actual SSH ports are 31090, 31098, 31100 (default), 31154 (ours)
- Added your SSH key to the server
- Downloaded FLUX.2-dev (166GB, 39 files) to
/FLUX.2-dev— completed successfully in 252s - Attempted inference — OOM crashed the instance. FLUX.2-dev is a 32B parameter model — the transformer alone is 60GB, cannot fit in 48GB VRAM even with CPU offloading
Current State
- Instance: DOWN — OOM killed the SSH daemon on port 31154
- Model files: 166GB of FLUX.2-dev safely saved at
/FLUX.2-devon the overlay filesystem - Jupyter Lab: Running on port 8888 (token:
amd-oneclick) but not accessible through cloud proxy - Other SSH ports (31090, 31098, 31100): Still responding but our key isn’t authorized
Next Steps (after restart)
| Step | Action | Why |
|---|---|---|
| 1 | Restart instance from cloud console | Get SSH back on port 31154 |
| 2 | Remove FLUX.2-dev files | Free 166GB disk |
| 3 | Download FLUX.1-dev (~35GB) | 12B params, needs ~24GB VRAM at FP16 |
| 4 | Run with enable_sequential_cpu_offload() |
Fits in 48GB VRAM ✅ |
FLUX.1-dev is the right choice for your 48GB GPU — it’s a 12B parameter model that needs ~24GB VRAM for the transformer plus ~10GB for the T5 text encoder. With CPU offloading, it’ll run comfortably.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋