tilelang.jit.adapter.tvm_ffi¶

Utilities to adapt TVM-FFI kernels to Torch tensors.

TVM-FFI obtains the active work stream through Torch’s DLPack Exchange API. The Ascend adapter installs TileLang’s Torch NPU callback before the first tensor reaches an executable.

Attributes¶

COMPILE_ARGS

Classes¶

TVMFFIKernelAdapter

Adapter that runs a TVM runtime.Executable with Torch tensors.

Module Contents¶

tilelang.jit.adapter.tvm_ffi.COMPILE_ARGS¶
class tilelang.jit.adapter.tvm_ffi.TVMFFIKernelAdapter(params, result_idx, target, func_or_mod, host_mod=None, device_mod=None, rt_mod=None, host_kernel_source=None, device_kernel_source=None, verbose=False, pass_configs=None, compile_flags=None)¶

Bases: tilelang.jit.adapter.base.BaseKernelAdapter

Adapter that runs a TVM runtime.Executable with Torch tensors.

Notes - Torch tensors use TVM-FFI’s zero-copy DLPack Exchange API conversion. - Ascend execution installs a Cython callback that reads Torch’s current

NPU stream for every invocation.

Parameters:

  • params (list[tilelang.engine.param.KernelParam])

  • result_idx (list[int])

  • target (str | tvm.target.Target)

  • func_or_mod (tvm.tirx.PrimFunc | tilelang.tvm.IRModule)

  • host_mod (tilelang.tvm.IRModule | None)

  • device_mod (tilelang.tvm.IRModule | None)

  • rt_mod (tilelang.tvm.runtime.Module | None)

  • host_kernel_source (str | None)

  • device_kernel_source (str | None)

  • verbose (bool)

  • pass_configs (dict[str, Any] | None)

  • compile_flags (list[str] | None)

target: str | tvm.target.Target = 'cuda'¶
ir_module: tilelang.tvm.IRModule | None = None¶
host_kernel_source: str | None = None¶
device_kernel_source: str | None = None¶
executable: tilelang.tvm.runtime.Executable | tilelang.tvm.runtime.Module | None = None¶
pass_configs: dict[str, Any] | None = None¶
host_mod: tilelang.tvm.IRModule | None = None¶
device_mod: tilelang.tvm.IRModule | None = None¶
rt_mod: tilelang.tvm.runtime.Module | None = None¶
dynamic_symbolic_map: dict[tvm.tirx.Var, tuple[int, int, int, int]] | None = None¶
params¶
result_idx¶
verbose = False¶
compile_flags = None¶
kernel_global_source = None¶
get_exportable_executable()¶

Return the lazy executable, or the runnable module loaded from disk cache.

Return type:

tilelang.tvm.runtime.Executable | tilelang.tvm.runtime.Module

classmethod from_database(params, result_idx, target, func_or_mod, host_kernel_source, device_kernel_source, kernel_lib_path, verbose=False, pass_configs=None, compile_flags=None)¶
Parameters:
get_host_source()¶

Returns the source code of the host module.

Return type:

str | None

get_device_source()¶

Returns the source code of the device module.

Return type:

str | None

get_kernel_source(kernel_only=False)¶

Returns the source code of the compiled kernel.

Parameters:

kernel_only (bool)

property prim_func: tvm.tirx.PrimFunc¶

Returns the primary TIR function from the IR module.

Return type:

tvm.tirx.PrimFunc