Unlock Metal for a workload

Unlock a tested Metal capability profile for one workload inside a Lume macOS VM.

A stock macOS guest reports conservative Metal capabilities, so some apps skip GPU paths the paravirtualized device can run. A small MIT shim, injected into one process, raises two answers: supportsFamily: up to a configured Apple family, and the maximum threadgroup memory. It does not pass a physical GPU through, patch the host, or touch the guest kernel. Background and benchmarks: technical writeup.

Experimental. Tested on Lume 0.5.1, an M1 Ultra host on macOS 26.6.1, and the Tahoe guest at macOS 26.5.2, with a capability probe, two llama.cpp workloads and one MLX-LM run. Reported support is not proof every Metal API works: test each workload.

1. Build the shim#

On an Apple silicon Mac with Xcode Command Line Tools:

git clone https://github.com/trycua/cua.git
cd cua/libs/lume/metal-capability-shim
./Scripts/build.sh
./Scripts/verify.sh      # architectures, signatures, checksums

dist/
├── LumeMetalCapabilities-arm64.dylib
├── LumeMetalCapabilities-arm64e.dylib
├── SHA256SUMS
└── metal-capabilities

Most guest workloads are arm64; check yours with lipo -archs <executable>.

2. Enable the host feature level#

Applies to VMs your user launches; restart the VM so the device is recreated:

lume stop my-vm
defaults write com.apple.gpusw.ParavirtualizedGraphics \
  ForceUnrestrictedDeviceFeatureLevel -bool true
defaults read com.apple.gpusw.ParavirtualizedGraphics \
  ForceUnrestrictedDeviceFeatureLevel
lume run my-vm

defaults read prints 1. Undo with defaults delete of the same key and a VM restart.

3. Copy it into the guest#

lume ssh my-vm "mkdir -p '/Users/lume/.local/share/lume/metal-capabilities'"
 
VM_IP=$(lume get my-vm --format json | jq -r '.[0].ipAddress')
scp ./dist/LumeMetalCapabilities-arm64.dylib ./dist/metal-capabilities \
  "lume@${VM_IP}:/Users/lume/.local/share/lume/metal-capabilities/"

Compare shasum -a 256 of both files on the host and in the guest.

4. Verify with the probe#

lume ssh my-vm \
  "DYLD_INSERT_LIBRARIES='/Users/lume/.local/share/lume/metal-capabilities/LumeMetalCapabilities-arm64.dylib' \
   LUME_METAL_APPLE_FAMILY_MAX=1009 \
   '/Users/lume/.local/share/lume/metal-capabilities/metal-capabilities' 1009"

Run it once without the two variables to see the stock answers. On the tested guest supportsFamily:1009 went from false to true and threadgroup memory from 32,768 to 65,536 bytes.

Variable	Default	Effect
`LUME_METAL_APPLE_FAMILY_MAX`	required	Apple-family ceiling, `1001` to `1999`; tested `1009` (Apple 9). Missing or invalid means stock behavior.
`LUME_METAL_MAX_THREADGROUP_MEMORY`	`65536`	Minimum reported threadgroup memory.
`LUME_METAL_RECOMMENDED_WORKING_SET_SIZE`	unchanged	Reported working set, only when set.

5. Run your workload#

Set the same two variables on the one process (and its children) that needs them:

lume ssh my-vm \
  "DYLD_INSERT_LIBRARIES='/Users/lume/.local/share/lume/metal-capabilities/LumeMetalCapabilities-arm64.dylib' \
   LUME_METAL_APPLE_FAMILY_MAX=1009 \
   /absolute/path/to/your-workload first-argument"

Never use launchctl setenv DYLD_INSERT_LIBRARIES; it affects the whole login session. For a long-running service, use a per-user LaunchAgent in the guest with the variables in its EnvironmentVariables:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
  <key>Label</key>
  <string>com.trycua.lume.metal-workload</string>
  <key>ProgramArguments</key>
  <array>
    <string>/absolute/path/to/your-workload</string>
    <string>first-argument</string>
  </array>
  <key>EnvironmentVariables</key>
  <dict>
    <key>DYLD_INSERT_LIBRARIES</key>
    <string>/Users/lume/.local/share/lume/metal-capabilities/LumeMetalCapabilities-arm64.dylib</string>
    <key>LUME_METAL_APPLE_FAMILY_MAX</key>
    <string>1009</string>
  </dict>
  <key>StandardErrorPath</key>
  <string>/Users/lume/Library/Logs/Lume/metal-workload.stderr.log</string>
</dict>
</plist>

PLIST="$HOME/Library/LaunchAgents/com.trycua.lume.metal-workload.plist"
mkdir -p "$HOME/Library/Logs/Lume"
plutil -lint "$PLIST"
launchctl bootstrap "gui/$(id -u)" "$PLIST"
launchctl print "gui/$(id -u)/com.trycua.lume.metal-workload"
# remove: launchctl bootout "gui/$(id -u)/com.trycua.lume.metal-workload"

If the workload fails, run it once without the shim before debugging further.

In Cua Spaces#

The GPU acceleration checkbox in New Space (Resources), --gpu on cua spaces create, or gpu="paravirtual" in the SDK does step 2 for you: cua sets the feature level before each start of that VM and removes it when the last GPU VM is deleted, if cua set it. While set, it applies to every VM your user starts. Steps 1 and 3 to 5 stay per workload. See Add a Space.