Feature/amd

[lukemarsden]

Note

Adds AMD/ROCm support with vendor-aware GPU detection, memory parsing, vLLM env selection, compose/install updates, and API/UI field changes.

• Runner/Inference:
  • Add vendor detection (nvidia/amd) and use nvidia-smi or rocm-smi accordingly for GPU count/memory; parse ROCm outputs; expose SDKVersion (CUDA/ROCm) replacing CUDAVersion.
  • Initialize/update per-GPU stats for AMD; update logs and error paths; adjust process matching for vLLM CUDA/ROCm venvs.
  • vLLM runtime chooses Python venv by vendor (/workspace/vllm-cuda or /workspace/vllm-rocm).
• Wolf/External Agent:
  • Build GOW_REQUIRED_DEVICES dynamically from GPU_VENDOR (adds /dev/kfd for AMD) in both lobbies/apps flows; neutralize GPU monitoring messages.
• Docker/Build:
  • Dockerfile: create separate CUDA and ROCm vLLM virtualenvs; for ROCm, install ROCm PyTorch and build vLLM from source; copy examples to both; add symlink for backward compat.
  • docker-compose: add wolf-amd service (profile code-amd); make moonlight-web GPU-agnostic.
• Installer:
  • Detect GPU vendor (NVIDIA/AMD/Intel), set GPU_VENDOR in .env; select compose profiles (code vs code-amd); conditionally install NVIDIA runtime; load uhid module; generate runner.sh with vendor-specific GPU flags (NVIDIA --gpus, AMD /dev/kfd + ROCR_VISIBLE_DEVICES).
• API/Types:
  • Change GPUStatus field from cuda_version to sdk_version; propagate in runner status and related structs; clarify GPU stats comments to include ROCm.
• Frontend:
  • Make GPU labels/vendor references generic (e.g., "GPU" and "SDK"); vendor-neutral model name parsing in Runner and Admin dashboards; update GPU query label text.

^{Written by Cursor Bugbot for commit ab9fa8c. This will update automatically on new commits. Configure here.}

[cursor]

This PR is being reviewed by Cursor Bugbot

Details

You are on the Bugbot Free tier. On this plan, Bugbot will review limited PRs each billing cycle.

To receive Bugbot reviews on all of your PRs, visit the Cursor dashboard to activate Pro and start your 14-day free trial.

Bug: AMD render node not detected for Wolf

The WOLF_RENDER_NODE auto-detection only checks for NVIDIA GPUs by matching driver name "nvidia", but never checks for AMD GPUs (driver name "amdgpu"). When GPU_VENDOR is set to "amd", the code falls back to the default /dev/dri/renderD128, which may be incorrect on systems with multiple GPUs or virtual GPUs. AMD systems need the same render node detection logic to find the correct AMD GPU render node.

install.sh#L1665-L1685

helix/install.sh

Lines 1665 to 1685 in ab9fa8c

	# Auto-detect first NVIDIA render node for Wolf
	# On some systems (Lambda Labs), renderD128 is virtio-gpu (virtual), NVIDIA starts at renderD129
	WOLF_RENDER_NODE="/dev/dri/renderD128" # Default
	if [ -d "/sys/class/drm" ]; then
	for render_node in /dev/dri/renderD*; do
	if [ -e "$render_node" ]; then
	# Check if this render node is NVIDIA by checking driver symlink
	node_name=$(basename "$render_node")
	driver_link="/sys/class/drm/$node_name/device/driver"
	if [ -L "$driver_link" ]; then
	driver=$(readlink "$driver_link" | grep -o '[^/]*$')
	if [[ "$driver" == "nvidia" ]]; then
	WOLF_RENDER_NODE="$render_node"
	echo "Auto-detected NVIDIA render node: $WOLF_RENDER_NODE"
	break
	fi
	fi
	fi
	done
	fi

 

---

[cursor]

Bug: Incorrect rocm-smi flag in updateGPUMemoryMap

The updateGPUMemoryMap function uses rocm-smi --showmeminfo vram --csv with the --csv flag, but the parsing logic at lines 950-1027 expects structured text format with patterns like "VRAM Total Used Memory (B):". This is inconsistent with other AMD memory queries in the same file (lines 218 and 403) which correctly use rocm-smi --showmeminfo vram without the --csv flag. The --csv flag will produce CSV output that won't match the structured text parser, causing AMD GPU memory updates to fail silently.

 

[lukemarsden]

but doesn't this just imply the amdgpu drivers weren't properly installed? ref https://learn.microsoft.com/en-us/azure/virtual-machines/linux/azure-n-series-amd-gpu-driver-linux-installation-guide#rocm-driver-installation

[lukemarsden]

superceded by #1379, I'll integrate any of the amd changes needed there, thanks!