AMD Radeon Cloud Instance Recovery

AMD Radeon Cloud Instance Recovery | Generated by AI

Home 2026.07

Complete Summary — AMD Radeon Cloud Instance

1. Initial Problem: SSH Connection Failed

You tried to connect to 36.150.116.206:31094 — got Connection refused.

Diagnosis:

✅ Host was reachable (39ms ping via DMIT jump host)
❌ Port 31094 not open
✅ Other ports open nearby: 31090, 31091, 31092, 31095, 31098, 31099, 31100
Found SSH on 31090, 31098, 31100 but key not authorized
Your ~/.ssh/config has a catch-all SOCKS5 proxy, so SSH was routing through proxy — we bypassed with ProxyCommand=none

2. Discovered the API Endpoint

You tested the API:

curl https://radeon-global.anruicloud.com/spaces/u-9758-811746f0/8000/v1/chat/completions

✅ Qwen3-8B responded successfully via the AMD Radeon Cloud proxy.

3. SSH Port 31154 Worked

You connected via ssh -p 31154 — got in with your key.

Server Specs:

Component	Value
OS	Ubuntu 24.04 LTS
CPU	2× AMD EPYC 9334 (128 threads)
RAM	503 GiB
GPU	AMD Radeon Graphics (gfx1100, RDNA 3) — 48 GB VRAM
ROCm	7.2.1, Driver 6.16.13
PyTorch	2.9.1 (ROCm build, in `/opt/venv`)
Disk	`/` = 3.5TB overlay, `/workspace` = 98GB loop

GPU compute test: ✅ Passed (HIP vector add, 48GB accessible, 96 CUs)

4. Tried FLUX.2-dev — Too Large (Crashed)

Model: 32B parameters, 166GB total
Downloaded via hf-mirror.com with hf_transfer — 252 seconds (~660MB/s)
Tried inference with CPU offloading: OOM crash ❌
- The transformer alone is 60GB in BF16
- Only 48GB VRAM available
- SSH daemon died when container got OOM-killed

5. Instance Restarted — SSH Broken

After the OOM crash, the container restarted but SSH was not running internally (it was cloud-proxied, not real SSH). You gained access through Jupyter Lab terminal at:

https://radeon-global.anruicloud.com/spaces/u-9758-811746f0/ (port 8888, token: amd-oneclick)

6. Current State (in Jupyter terminal)

Item	Status
Jupyter terminal	✅ You’re in
SSH server	❌ Not installed (we tried `apt install openssh-server` + `mkdir -p /run/sshd`, but cloud proxy port mapping is stale)
FLUX.2-dev	Still on disk at `/FLUX.2-dev` (166GB)
`/workspace` disk	98GB (19GB free) — too small
`/` disk	3.5TB (2.8TB free) ✅

7. What’s Next: FLUX.1-dev

Since you’re in the Jupyter terminal, I can still help you run FLUX.1-dev — it’s 12B params (~35GB total, needs ~24GB VRAM at FP16) and will fit comfortably on your 48GB GPU.

From the Jupyter terminal, paste:

rm -rf /FLUX.2-dev && export PATH=/opt/venv/bin:$PATH && export HF_ENDPOINT=https://hf-mirror.com && export HF_HUB_ENABLE_HF_TRANSFER=1 && python3 -c "
from huggingface_hub import snapshot_download
print('Downloading FLUX.1-dev...')
snapshot_download('black-forest-labs/FLUX.1-dev', local_dir='/FLUX.1-dev')
print('Done!')
"

Once downloaded, run:

export PATH=/opt/venv/bin:$PATH && cd /workspace && python3 -c "
import torch
from diffusers import FluxPipeline
pipe = FluxPipeline.from_pretrained('/FLUX.1-dev', torch_dtype=torch.bfloat16)
pipe.enable_sequential_cpu_offload()
image = pipe('A cute cat on Mars', num_inference_steps=28).images[0]
image.save('flux_output.png')
print('Saved!')
"

Back Donate