Merge pull request #18 from lemonade-sdk/gpu_device_selection

feat: integrate mesa-2404 content snap for Vulkan support and add gpu-device selection

Author: kenvandine

README.md

@@ -35,6 +35,16 @@ The snap automatically detects your GPU and selects the best backend:
 - **AMD GPUs** - Uses ROCm for optimal performance
 - **Other GPUs** - Uses Vulkan
 
+### GPU Drivers (Vulkan)
+
+Vulkan support is provided by the `mesa-2404` content snap, which is installed
+automatically. If it isn't connected, connect it manually:
+
+```bash
+sudo snap connect lemonade-server:gpu-2404 mesa-2404:gpu-2404
+sudo snap restart lemonade-server.daemon
+```
+
 ### Enable ROCm Support (AMD GPUs)
 
 For ROCm GPU acceleration on AMD hardware, connect the `process-control` interface:
@@ -113,6 +123,10 @@ sudo snap set lemonade-server ctx-size=32768
 
 # Set optional llamacpp-args
 sudo snap set lemonade-server llamacpp-args="--no-mmap --flash-attn on"
+
+# Select a specific GPU for Vulkan (default: auto-detect all)
+# Useful on multi-GPU systems to target a discrete GPU
+sudo snap set lemonade-server gpu-device=nvidia
 ```
 
 Changes take effect after the service restarts automatically.
@@ -124,6 +138,7 @@ Changes take effect after the service restarts automatically.
 | `log-level` | info | Log verbosity (debug, info, warn) |
 | `ctx-size` | 4096 | Context window size in tokens |
 | `llamacpp-args` | "" | Optional args to pass to llamacpp-server |
+| `gpu-device` | "" | Vulkan GPU filter: `nvidia`, `intel`, `amd`, or empty for all |
 
 ### Model cache location
 
@@ -180,13 +195,20 @@ sudo journalctl -u snap.lemonade-server.daemon.service --no-pager -n 50
 
 ### GPU not detected
 
-Ensure you have the proper GPU drivers installed:
+Ensure the `gpu-2404` interface is connected:
 ```bash
-# For AMD GPUs
-sudo apt install mesa-vulkan-drivers
+sudo snap connect lemonade-server:gpu-2404 mesa-2404:gpu-2404
+sudo snap restart lemonade-server.daemon
+```
 
-# Check Vulkan support
-vulkaninfo | head -20
+### Vulkan crashes on multi-GPU systems
+
+On systems with multiple GPUs (e.g. Intel integrated + NVIDIA discrete), the
+Vulkan backend may default to the integrated GPU. Use `gpu-device` to select
+the discrete GPU:
+
+```bash
+sudo snap set lemonade-server gpu-device=nvidia
 ```
 
 ### ROCm not working (AMD GPUs)

scripts/lemonade-server-wrapper

@@ -16,6 +16,28 @@ export LEMONADE_LOG_LEVEL=${LEMONADE_LOG_LEVEL:-info}
 export LEMONADE_CTX_SIZE=${LEMONADE_CTX_SIZE:-4096}
 export LEMONADE_LLAMACPP_ARGS=${LEMONADE_LLAMACPP_ARGS:-""}
 
+# GPU device selection via Vulkan loader driver filtering
+LEMONADE_GPU_DEVICE=$(snapctl get gpu-device)
+if [ -n "$LEMONADE_GPU_DEVICE" ]; then
+  case "$LEMONADE_GPU_DEVICE" in
+    nvidia)
+      export VK_LOADER_DRIVERS_SELECT='*nvidia*'
+      echo "[Lemonade Wrapper] GPU device filter: nvidia only" >&2
+      ;;
+    intel)
+      export VK_LOADER_DRIVERS_SELECT='*intel*'
+      echo "[Lemonade Wrapper] GPU device filter: intel only" >&2
+      ;;
+    amd)
+      export VK_LOADER_DRIVERS_SELECT='*radeon*'
+      echo "[Lemonade Wrapper] GPU device filter: amd only" >&2
+      ;;
+    *)
+      echo "[Lemonade Wrapper] Unknown gpu-device value: $LEMONADE_GPU_DEVICE (ignoring)" >&2
+      ;;
+  esac
+fi
+
 BACKEND="vulkan"  # default
 
 # Supported ROCm architectures (must have matching backend binaries)

snap/hooks/install

@@ -6,3 +6,4 @@ snapctl set ip-address=127.0.0.1
 snapctl set log-level=info
 snapctl set ctx-size=4096
 snapctl set llamacpp-args=""
+snapctl set gpu-device=""

snap/snapcraft.yaml

@@ -50,6 +50,12 @@ layout:
   /usr/share/lemonade-server:
     symlink: $SNAP/usr/share/lemonade-server
 
+plugs:
+  gpu-2404:
+    interface: content
+    target: $SNAP/gpu-2404
+    default-provider: mesa-2404
+
 slots:
   # Adds a pseudo content interface to allow clients to express their
   # runtime dependency on lemonade-server
@@ -62,30 +68,35 @@ slots:
 apps:
   metrics:
     command: bin/metrics
+    command-chain: [bin/gpu-2404-wrapper]
     plugs:
       - home
       - removable-media
       - network
       - network-bind
       - opengl
+      - gpu-2404
       - process-control
       - hardware-observe
       - system-observe
   lemonade-server:
     command: usr/bin/lemonade-server
+    command-chain: [bin/gpu-2404-wrapper]
     plugs:
       - home
       - removable-media
       - network
       - network-bind
       - opengl
+      - gpu-2404
       - hardware-observe
       - system-observe
     environment:
-      LD_LIBRARY_PATH: "$SNAP/usr/share/lemonade-server:$SNAP/usr/lib/llamacpp/vulkan:$SNAP/usr/lib/x86_64-linux-gnu:$SNAP/lib/x86_64-linux-gnu:/var/lib/snapd/lib/gl${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
+      LD_LIBRARY_PATH: "$SNAP/usr/share/lemonade-server:$SNAP_COMMON/cache/lemonade/bin/llamacpp/vulka    n:$SNAP_COMMON/cache/lemonade/bin/whispercpp/vulkan:$SNAP/usr/lib/x86_64-linux-gnu:$SNAP/lib/x86_64-linux-gnu:/var/lib/snapd/lib/gl${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
       PATH: "$SNAP/usr/bin:$PATH"
   daemon:
     command: bin/lemonade-server-wrapper serve
+    command-chain: [bin/gpu-2404-wrapper]
     daemon: simple
     install-mode: enable
     restart-condition: on-failure
@@ -96,11 +107,12 @@ apps:
       - network
       - network-bind
       - opengl
+      - gpu-2404
       - process-control
       - hardware-observe
       - system-observe
     environment:
-      LD_LIBRARY_PATH: "$SNAP/usr/share/lemonade-server:$SNAP/usr/lib/llamacpp/vulkan:$SNAP/usr/lib/x86_64-linux-gnu:$SNAP/lib/x86_64-linux-gnu:/var/lib/snapd/lib/gl${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
+      LD_LIBRARY_PATH: "$SNAP/usr/share/lemonade-server:$SNAP_COMMON/cache/lemonade/bin/llamacpp/vulka    n:$SNAP_COMMON/cache/lemonade/bin/whispercpp/vulkan:$SNAP/usr/lib/x86_64-linux-gnu:$SNAP/lib/x86_64-linux-gnu:/var/lib/snapd/lib/gl${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
       PATH: "$SNAP/usr/bin:$PATH"
       HOME: "$SNAP_COMMON"
       LEMONADE_CACHE_DIR: "$SNAP_COMMON/cache/lemonade"
@@ -119,7 +131,6 @@ parts:
       - unzip # For extracting backends at runtime
       - libdrm-amdgpu1 # SD backends need this at runtime
       - libnuma1 # SD backends need this at runtime
-      - libvulkan1
     override-build: |
       set +u # For unbound variables
       # workaround for build.snapcraft.io builds
@@ -160,6 +171,16 @@ parts:
     override-build: |
       craftctl default
 
+  gpu-snap:
+    after: [lemonade-server, deps]
+    source: https://github.com/canonical/gpu-snap.git
+    plugin: dump
+    override-prime: |
+      craftctl default
+      ${CRAFT_PART_SRC}/bin/gpu-2404-cleanup mesa-2404
+    prime:
+      - bin/gpu-2404-wrapper
+
   scripts:
     plugin: dump
     source: scripts/