Skip to content

TVM Compile

Use the compile action to compile a Relax IR model for a target architecture and produce a deployable model.so.

What the action does

Select a target CPU profile and optional compiler and TVM pipeline settings to produce the compiled artifact. These settings belong to the Run; reusable Tasks contain only common infrastructure settings. CORE records the result in function.so_model.

Target architecture

Profiles come from CORE's runtime-tvm.yaml; the SDK accepts configured names rather than a fixed enumeration. The bundled profiles are auto, amd64, amd64_v3, amd64_native, arm64, arm64_pi5, and arm32.

amd64 means Intel/AMD 64-bit (x86-64-v2); arm64 means ARM 64-bit (AArch64); arm32 requires ARMv7 hard-float with NEON. auto uses the compile node's architecture. Older aliases cpu/llvm, x86, x86_v3, x86_native, and armv7l map to auto, amd64, amd64_v3, amd64_native, and arm32.

Advanced runs can provide an explicit LLVM target instead of selecting a profile.

  • Create


    Use the TVM Function created for the build action.

  • Run


    Compile Relax IR and configure the target architecture.

  • Use


    Pass the compiled model artifact to the serve action.