MI355X container setup (Podman + VS Code + Oh My Bash)

Canonical instructions: example_setup.md (manual steps + how automation maps to them).

Agents in this folder: see AGENTS.md — you will be asked on each new session whether to allocate a GPU node and create the vllm-dev container.

One command (allocate + setup):

./start_allocation.sh

Active allocation details (node, job ID, image) are written to connection_info.txt after ./start_allocation.sh.

# On your allocated node (after ssh):
podman ps --filter name=vllm-dev
/shared/amdgpu/home/fai_qle/bin/docker ps   # same, for VS Code

1. Reserve a GPU node

salloc --reservation=mi355-gpu-34_gpu-39_gpu-40_gpu-49_gpu-50_gpu-51_gpu-55_gpu-56_reservation --exclusive --account=wrh --mem=0
# OR
salloc --reservation=mi355-gpu-41_gpu-42_gpu-43_gpu-44_gpu-45_gpu-46_gpu-47_gpu-48_reservation --exclusive --account=wrh --mem=0

Run all podman commands on the allocated compute node (e.g. mi355-gpu-40), not the SLURM controller.

2. Create / refresh the container

cd /shared/amdgpu/home/fai_qle/prepare_env
./setup_container.sh

Or manually follow example_setup.md (pull image, podman run, fake docker shim).

setup_container.sh copies your cluster ~/.ssh into the container (/root and vscode) for private git clone. To refresh an existing container: ./copy_ssh_to_container.sh vllm-dev.

3. Connect with VS Code on your PC (Windows)

SSH config — add to C:\Users\fai\.ssh\config

Host amd-login
    HostName aac16.amd.com
    User fai_qle
    IdentityFile C:\Users\fai\.ssh\id_ed25519
    ServerAliveInterval 60

Host mi355-gpu-39
    HostName mi355-gpu-39
    User fai_qle
    IdentityFile C:\Users\fai\.ssh\id_ed25519
    ProxyJump amd-login
    ServerAliveInterval 60

Test in PowerShell: ssh mi355-gpu-39

VS Code extensions (local)

• Remote - SSH (ms-vscode-remote.remote-ssh)
• Dev Containers (ms-vscode-remote.remote-containers)

Attach to the running container

1. VS Code → Remote-SSH: Connect to Host... → your node (see connection_info.txt)
2. Command Palette → Dev Containers: Attach to Running Container... → vllm-dev
3. Remote user: root

VS Code should open /shared/amdgpu/home/fai_qle (not /root/vllm or /app/vllm). Edits under that path persist on shared NFS after the container is removed. If it opens an old folder, use Dev Containers: Open Named Configuration File → vllm-dev → set "workspaceFolder": "/shared/amdgpu/home/fai_qle" per attached-container.vllm-dev.json. vLLM sources in the image live under /app/vllm if you need them.

Machine settings on the node already point VS Code at the podman shim:

/shared/amdgpu/home/fai_qle/bin/docker → podman

If Dev Containers cannot find Docker, set in VS Code settings (Remote window):

"dev.containers.dockerPath": "/shared/amdgpu/home/fai_qle/bin/docker"

4. What is installed inside the container

Component	Users	Notes
Oh My Bash	root	ohmybash/oh-my-bash
Cursor Agent CLI	root	agent in ~/.local/bin
Claude Code	root	claude — run once to log in
OpenAI Codex	system	codex in /usr/local/bin — run once to log in
tmux	system	tmux via apt
jq	system	jq via apt
ripgrep	system	rg via apt
GitHub CLI	system	gh via cli.github.com apt repo
Host ROCm module	7.2.3	See rocm_stack.env
vLLM container image	—	vllm/vllm-openai-rocm:nightly (container name vllm-dev)

Port 8080 on the host maps to 8000 in the container (vLLM API default).

5. Useful commands

# Shell in container
podman exec -it -u vscode vllm-dev bash -l

# Cursor Agent
podman exec -u vscode vllm-dev bash -l -c 'agent --version'

# Stop / remove
podman stop vllm-dev
podman rm vllm-dev