AMD Dev Cloud Volume Management Steps

AMD Dev Cloud Volume Management Steps | Generated by AI

Home 2026.06

Comprehensive Summary: Volume Management on AMD Dev Cloud (DigitalOcean)

1. Investigating “200 is not ok” Volume Resize Error

Checked doctl CLI code at /Users/lzwjava/projects/doctl/commands/volumes.go:

--size flag requires unit suffix (e.g. 200GiB), bare 200 = 200 bytes → 0 GiB
But the UI error was different — HTML shows min="101" max="16384", current volume was 100 GiB

Tested resize via API — every size failed:

doctl compute volume-action resize 52743aec-... --size 101 --region atl1  # 422
doctl compute volume-action resize 52743aec-... --size 200 --region atl1  # 422
doctl compute volume-action resize 52743aec-... --size 500 --region atl1  # 422

Also tested via Python (curl), detaching first, different regions — all 422 "invalid size specified".

Conclusion: Resize API not supported for AMD Dev Cloud partner (GPU) volumes.

2. Moving Data to Volume (129.212.178.103)

Checked disk usage:

ssh root@129.212.178.103 'df -h && du -sh /root/ /var/ /opt/'

Result: /root/ 37G (llama models), /var/ 60G (containerd), /opt/ 22G (ROCm)

Mounted and moved llama models (37G):

ssh root@129.212.178.103 'mount /dev/sda /mnt/volume_atl1_1780280110689'
ssh root@129.212.178.103 'rsync -a --progress /root/llama.cpp/models/ /mnt/volume_atl1_1780280110689/llama-models/'
ssh root@129.212.178.103 'rm -rf /root/llama.cpp/models && ln -s /mnt/volume_atl1_1780280110689/llama-models /root/llama.cpp/models'

Investigated containerd (59G):

ssh root@129.212.178.103 'docker images -a && docker ps -a && docker system df'

Found: rocm:latest (36GB), ubuntu:24.04 (119MB), exited rocm container (Jupyter Lab), 36GB build cache.

Cleaned up Docker artifacts:

ssh root@129.212.178.103 'docker rm rocm'
ssh root@129.212.178.103 'docker rmi rocm:latest ubuntu:24.04'
ssh root@129.212.178.103 'docker builder prune --all -f'
ssh root@129.212.178.103 'docker system prune --all -f'

Made mount persistent:

ssh root@129.212.178.103 'echo "/dev/sda /mnt/volume_atl1_1780280110689 ext4 defaults,nofail 0 2" >> /etc/fstab'

Result: Root disk freed 95G (124G → 29G used).

3. Detaching Volume and Destroying GPU Droplet

Unmounted volume:

ssh root@129.212.178.103 'umount /mnt/volume_atl1_1780280110689'

Detached volume:

doctl compute volume-action detach 52743aec-5d63-11f1-a928-0a58ac126378 574422820 --wait

Powered on droplet (needed for snapshot):

doctl compute droplet-action power-on 574422820 --wait

Created snapshot:

doctl compute droplet-action snapshot 574422820 --snapshot-name "gpu-mi300x-snapshot-20260601" --wait

Destroyed GPU droplet:

doctl compute droplet delete 574422820 --force

Verified:

doctl compute snapshot list --resource droplet
doctl compute volume list

Final State

Resource	ID	Status	Cost/mo
GPU Droplet	574422820	DESTROYED	$0
Snapshot	230979911	gpu-mi300x-snapshot-20260601 (30.6 GiB)	~$1.53
Volume	52743aec-…	DETACHED (100 GiB, 37G llama-models)	~$10
Total			~$11.50/mo

Was: ~$2+/hr = ~$1,460/mo if left running.

4. Extra Snapshots (potential cleanup)

doctl compute snapshot list --resource droplet

Snapshot	Size	Cost/mo	Safe to delete?
`7.2-gpu-mi300x1-...-1779256307024`	132.77 GiB	$6.64	Yes (original AMD image, re-downloadable)
`snap-snapshots-gpu-mi300x1-...-0530-2331`	6.58 GiB	$0.33	Yes
`gpu-mi300x-snapshot-20260601`	30.62 GiB	$1.53	No (your snapshot)

# To delete old ones and save ~$7/mo:
doctl compute snapshot delete 229365612 --force
doctl compute snapshot delete 230740400 --force

5. Next Steps (when ready)

# Create cheap CPU droplet for data prep
doctl compute droplet create data-prep \
  --size s-2vcpu-4gb \
  --region atl1 \
  --image ubuntu-24-04-x64 \
  --ssh-keys <fingerprint>

# Attach volume
doctl compute volume-action attach 52743aec-... <new-droplet-id>

# SSH in, mount, download dataset, tokenize
ssh root@<new-ip>
mount /dev/sda /mnt/volume_atl1_1780280110689

# When data ready: detach, destroy CPU droplet
doctl compute volume-action detach 52743aec-... <new-droplet-id>
doctl compute droplet delete <new-id> --force

# Recreate GPU droplet from snapshot
doctl compute droplet create gpu-train \
  --size gpu-mi300x1-192gb \
  --region atl1 \
  --image 230979911 \
  --ssh-keys <fingerprint>

# Attach volume, train
doctl compute volume-action attach 52743aec-... <gpu-droplet-id>

Back Donate