Platform & Backend Matrix

This page combines platform support and backend-module configuration for llamadart.

The native-assets hook currently pins llamadart-native tag b8638 (hook/build.dart). Module availability below is for that pinned tag.

Platform/architecture coverage#

All iOS targets above require the consuming Flutter/Xcode project to use a minimum deployment target of 16.4 or newer (for example platform :ios, '16.4').

Current module availability by bundle (b8638)#

Selector names and aliases#

llamadart_native_backends values are matched against modules discovered in the selected bundle. Current configurable-bundle module names are:

• cpu
• vulkan
• opencl
• cuda
• blas
• hip

Aliases:

• vk -> vulkan
• ocl -> opencl
• open-cl -> opencl

GpuBackend.metal remains valid as a runtime backend preference on Apple targets, but Apple targets are non-configurable in llamadart_native_backends.

Configuring native backend modules#

Use hooks.user_defines.llamadart.llamadart_native_backends:

hooks:
  user_defines:
    llamadart:
      llamadart_native_backends:
        platforms:
          android-arm64:
            backends: [vulkan]
            cpu_profile: full # default; use compact for baseline-only
          linux-x64: [vulkan, cuda]
          windows-x64:
            backends: [vulkan, cuda, blas]

Android arm64 CPU policy keys (platforms.android-arm64):

• cpu_profile: full (default): include all Android ARM CPU variant modules.
• cpu_profile: compact: include baseline CPU variant module only.
• cpu_variants: [...] (advanced): explicit CPU variant list, overrides cpu_profile.

Supported canonical cpu_variants values:

• android_armv8.0_1 (baseline)
• android_armv8.2_1
• android_armv8.2_2
• android_armv8.6_1
• android_armv9.0_1
• android_armv9.2_1
• android_armv9.2_2

Variant feature differences:

Accepted cpu_variants input forms are normalized, for example:

• baseline
• armv8_6_1
• v9_0_1
• android-armv9.2_2
• libggml-cpu-android_armv8.2_2.so

If cpu_variants contains unknown entries, they are ignored with warnings. If no valid entries remain, selection falls back to cpu_profile (or default full).

Selection and fallback behavior#

• Configurable targets start from defaults (cpu, vulkan) if available.
• cpu is auto-added as fallback when present in the bundle.
• Android arm64 defaults to cpu_profile: full.
• cpu_variants (if provided) takes precedence over cpu_profile for Android arm64.
• If requested modules are unavailable for a bundle, the hook warns and falls back to defaults.
• If defaults are also unavailable, all available modules in that bundle are used as fallback.
• Apple targets (ios-*, macos-*) support cpu + metal, but ignore per-backend module config in this hook path because runtime libraries are consolidated.
• windows-x64 performs extra runtime dependency validation:
  • cuda requires cudart and cublas DLLs.
  • blas requires OpenBLAS DLL.
• If you change llamadart_native_backends, run flutter clean once to clear stale native-asset outputs.

Related docs#

• Native Build Hooks
• Linux Prerequisites
• WebGPU Bridge