tilelang.autotuner.grouped_compile¶
Grouped compilation helpers for autotuner.
This module isolates backend-aware grouped compilation logic from AutoTuner.run so tuner.py can stay focused on orchestration.
Attributes¶
Functions¶
|
Compile one grouped unit for CUDA+tvm_ffi backend. |
Module Contents¶
- tilelang.autotuner.grouped_compile.CompileUnitResult¶
- tilelang.autotuner.grouped_compile.compile_grouped_unit_tvm_ffi(unit_items, compile_args, elaborate_func)¶
Compile one grouped unit for CUDA+tvm_ffi backend.
Flow: 1. Elaborate each config into a PrimFunc. 2. Lower each PrimFunc and build its host module in a fresh Z3 context. 3. Merge all device IR and compile it once in a separate fresh Z3 context. 4. Import the shared device module into each host runtime module. 5. Construct per-config JITKernel objects that share the grouped device module.
- Parameters:
unit_items (list[tuple[int, dict[str, Any]]])
compile_args (tilelang.autotuner.param.CompileArgs)
elaborate_func (collections.abc.Callable[Ellipsis, tvm.tirx.PrimFunc])
- Return type:
list[CompileUnitResult]