Repurpose numpy-compiler compiler/runtime flow for PyTorch.

* A bit gross because I took the chance to upgrade all of the backend bits to the new MLIR Python bindings and we still co-mingle the old and new for now.
* Since the Python created PassManagers are configured for explicit nesting, I had to upgrade some of the pass pipelines to be explicit.
* The demo in mul_maximum_e2e.py now compiles, runs through PyTorch and through the JIT, prints and asserts the same results.
* I am not claiming that this is the prettiest API in this patch: consider that this is just directly using low-level APIs and there should be an intervening high level API.

external/llvm-project

@@ -1 +1 @@
-Subproject commit 53a0d45db6d0f33dfbb724c99ce2560ae25473c2
+Subproject commit 5fef6ce0cce05a9fb05f47c9d62f3724377ea076

frontends/pytorch/examples/mul_maximum_e2e.py

@@ -3,9 +3,16 @@
 # See frontends/pytorch/LICENSE for license information.
 
 import sys
+import numpy as np
 import torch
 import torch_mlir
 
+import npcomp
+from npcomp.compiler.pytorch.backend.refjit import *
+from npcomp.compiler.utils import logging
+
+logging.enable()
+
 lhs = torch.ones((4, 6, 1))
 rhs = torch.ones((1, 1, 3)) * 0.6
 bias = torch.ones((1, 1, 3)) * 0.2
@@ -17,8 +24,13 @@ with mb.capture_function("mul_maximum", [lhs, rhs, threshold, bias]) as f:
   result = result + bias
   f.returns([result])
 
-print(f"Result(f{result.size()}) = {result}", file=sys.stderr)
-# TODO: Currently need to route through:
-#   npcomp-opt -aten-recognize-kernels -convert-aten-to-tcf \
-#     -numpy-public-functions-to-tensor -canonicalize
-mb.module.operation.print()
+backend = CompilerBackend()
+jit_module = backend.load(backend.compile(mb.module))
+
+jit_result = jit_module.mul_maximum(lhs.numpy(), rhs.numpy(), threshold.numpy(),
+                                    bias.numpy())
+
+print(f"PyTorch Result = {result.numpy()}", file=sys.stderr)
+print(f"JIT Result = {jit_result}", file=sys.stderr)
+
+np.testing.assert_allclose(result.numpy(), jit_result)

include/npcomp/Backend/RefJIT/PythonModule.h

@@ -15,7 +15,7 @@ namespace npcomp {
 namespace python {
 
 /// Defines an "refjit" module with backend support definitions.
-void defineBackendRefJitModule(py::module m);
+void defineBackendRefJitModule(py::module &m);
 
 } // namespace python
 } // namespace npcomp

include/npcomp/Python/PybindUtils.h

@@ -15,6 +15,9 @@
 #include <pybind11/pytypes.h>
 #include <pybind11/stl.h>
 
+#include "mlir-c/Bindings/Python/Interop.h"
+#include "mlir-c/IR.h"
+#include "mlir-c/Pass.h"
 #include "llvm/ADT/Optional.h"
 
 namespace py = pybind11;
@@ -25,6 +28,45 @@ namespace detail {
 template <typename T>
 struct type_caster<llvm::Optional<T>> : optional_caster<llvm::Optional<T>> {};
 
+/// Casts object -> MlirContext.
+template <> struct type_caster<MlirContext> {
+  PYBIND11_TYPE_CASTER(MlirContext, _("MlirContext"));
+  bool load(handle src, bool) {
+    auto capsule = src.attr(MLIR_PYTHON_CAPI_PTR_ATTR);
+    value = mlirPythonCapsuleToContext(capsule.ptr());
+    if (mlirContextIsNull(value)) {
+      return false;
+    }
+    return true;
+  }
+};
+
+/// Casts object -> MlirModule.
+template <> struct type_caster<MlirModule> {
+  PYBIND11_TYPE_CASTER(MlirModule, _("MlirModule"));
+  bool load(handle src, bool) {
+    auto capsule = src.attr(MLIR_PYTHON_CAPI_PTR_ATTR);
+    value = mlirPythonCapsuleToModule(capsule.ptr());
+    if (mlirModuleIsNull(value)) {
+      return false;
+    }
+    return true;
+  }
+};
+
+/// Casts object -> MlirPassManager.
+template <> struct type_caster<MlirPassManager> {
+  PYBIND11_TYPE_CASTER(MlirPassManager, _("MlirPassManager"));
+  bool load(handle src, bool) {
+    auto capsule = src.attr(MLIR_PYTHON_CAPI_PTR_ATTR);
+    value = mlirPythonCapsuleToPassManager(capsule.ptr());
+    if (mlirPassManagerIsNull(value)) {
+      return false;
+    }
+    return true;
+  }
+};
+
 } // namespace detail
 } // namespace pybind11
 

lib/Backend/RefJIT/PythonModule.cpp

@@ -10,6 +10,8 @@
 
 #include "pybind11/numpy.h"
 
+#include "mlir/CAPI/IR.h"
+#include "mlir/CAPI/Pass.h"
 #include "npcomp/Python/MlirIr.h"
 #include "npcomp/Python/MlirPass.h"
 #include "npcomp/RefBackend/JITHelpers/JITModule.h"
@@ -84,20 +86,22 @@ py::array wrapTensorAsArray(Ref<Tensor> tensor) {
                    /*base=*/std::move(pyTensor));
 }
 
-void npcomp::python::defineBackendRefJitModule(py::module m) {
-  m.def("build_backend_compilation_pipeline", [](PyPassManager &pm) {
-    JITModule::buildBackendCompilationPipeline(pm.passManager);
+void npcomp::python::defineBackendRefJitModule(py::module &m) {
+  m.def("build_backend_compilation_pipeline", [](MlirPassManager capiPm) {
+    mlir::PassManager *pm = unwrap(capiPm);
+    JITModule::buildBackendCompilationPipeline(*pm);
   });
   py::class_<JITModule>(m, "JITModule")
       .def_static(
           "from_compiled_module",
-          [](PyModuleOp module, std::vector<std::string> pySharedLibs)
+          [](MlirModule capiModule, std::vector<std::string> pySharedLibs)
               -> std::unique_ptr<JITModule> {
             SmallVector<StringRef, 4> sharedLibs(pySharedLibs.begin(),
                                                  pySharedLibs.end());
-            auto jitModule = checkError(
-                JITModule::fromCompiledModule(module.moduleOp, sharedLibs),
-                "error creating JITModule: ");
+            auto module = unwrap(capiModule);
+            auto jitModule =
+                checkError(JITModule::fromCompiledModule(module, sharedLibs),
+                           "error creating JITModule: ");
             return jitModule;
           },
           py::arg("module"), py::arg("shared_libs"))

lib/Conversion/TCFToStd/TCFToStd.cpp

@@ -69,8 +69,8 @@ matchAndRewriteBinaryElementwise(Operation *op, PatternRewriter &rewriter) {
       loc, resultType, rhs, broadcastedShape);
   Value binaryOpResult;
   if (isa<tcf::AddOp>(op)) {
-    binaryOpResult = rewriter.create<AddFOp>(
-        loc, result.getType(), lhsBroadcasted, rhsBroadcasted);
+    binaryOpResult = rewriter.create<AddFOp>(loc, result.getType(),
+                                             lhsBroadcasted, rhsBroadcasted);
   } else if (isa<tcf::MaxOp>(op)) {
     // XXX: remove TCP dep
     // XXX: remove TCP ops from TCP
@@ -79,8 +79,8 @@ matchAndRewriteBinaryElementwise(Operation *op, PatternRewriter &rewriter) {
     binaryOpResult =
         rewriter.create<SelectOp>(loc, pred, lhsBroadcasted, rhsBroadcasted);
   } else if (isa<tcf::MulOp>(op)) {
-    binaryOpResult = rewriter.create<MulFOp>(
-        loc, result.getType(), lhsBroadcasted, rhsBroadcasted);
+    binaryOpResult = rewriter.create<MulFOp>(loc, result.getType(),
+                                             lhsBroadcasted, rhsBroadcasted);
   } else {
     op->dump();
     llvm::report_fatal_error(

lib/RefBackend/RefBackend.cpp

@@ -194,17 +194,17 @@ void mlir::NPCOMP::createRefBackendLoweringPipeline(
   //
   // Also, converting to linalg herevopens up a lot of optimization
   // opportunities.
-  pm.addPass(createConvertElementwiseToLinalgPass());
+  pm.addNestedPass<FuncOp>(createConvertElementwiseToLinalgPass());
 
   if (options.optimize) {
-    pm.addPass(createLinalgFusionOfTensorOpsPass());
-    pm.addPass(createCanonicalizerPass());
-    pm.addPass(createCSEPass());
+    pm.addNestedPass<FuncOp>(createLinalgFusionOfTensorOpsPass());
+    pm.addNestedPass<FuncOp>(createCanonicalizerPass());
+    pm.addNestedPass<FuncOp>(createCSEPass());
   }
 
   // Lower shape constraints before we enter tensor->memref conversion.
   // That is, we expand shape.cstr_* ops to eager error handling code.
-  pm.addPass(createConvertShapeConstraintsPass());
+  pm.addNestedPass<FuncOp>(createConvertShapeConstraintsPass());
   // Run shape canonicalizations. In particular, this erases shape.assuming,
   // now that we have converted shape constraints.
   // TODO: This is kind of ugly. Either we use pass options or a constructor
@@ -227,12 +227,12 @@ void mlir::NPCOMP::createRefBackendLoweringPipeline(
   // (tensor_load / tensor_to_memref) in the IR.
 
   // Bufferize the TCP dialect.
-  pm.addPass(createTCPBufferizePass());
+  pm.addNestedPass<FuncOp>(createTCPBufferizePass());
   // Lower tensor-valued constants to refback.global.
   pm.addPass(createLowerConstantTensorsToMemrefPass());
   // refback::AllocMemRefOp takes a shape (i.e. extent tensor) as an argument.
   // We need to resolve this to std.alloc which takes individual extents.
-  pm.addPass(createLowerAllocMemRefOpsPass());
+  pm.addNestedPass<FuncOp>(createLowerAllocMemRefOpsPass());
   // Lower shape ops to std.
   // TODO: This should in principle be moved before tensor->memref conversion.
   // But some of the tensor->memref lowerings above use shape.get_extent. For
@@ -241,8 +241,8 @@ void mlir::NPCOMP::createRefBackendLoweringPipeline(
   pm.addPass(createConvertShapeToStandardPass());
 
   // Run some upstream bufferization passes to finish bufferization.
-  pm.addPass(createStdBufferizePass());
-  pm.addPass(createSCFBufferizePass());
+  pm.addNestedPass<FuncOp>(createStdBufferizePass());
+  pm.addNestedPass<FuncOp>(createSCFBufferizePass());
   pm.addPass(createLinalgBufferizePass());
   pm.addPass(createFuncBufferizePass());
 
@@ -252,8 +252,8 @@ void mlir::NPCOMP::createRefBackendLoweringPipeline(
   // At this point, we have lots of loose stuff floating around from lowering,
   // so it's a good time to do some general cleanups.
   if (options.optimize) {
-    pm.addPass(createCanonicalizerPass());
-    pm.addPass(createCSEPass());
+    pm.addNestedPass<FuncOp>(createCanonicalizerPass());
+    pm.addNestedPass<FuncOp>(createCSEPass());
   }
 
   // --------------------------------------------------------------------------
@@ -264,12 +264,12 @@ void mlir::NPCOMP::createRefBackendLoweringPipeline(
 
   // Lower linalg ops to loops.
   // TODO: Do some linalg optimizations like tiling here.
-  pm.addPass(createConvertLinalgToLoopsPass());
+  pm.addNestedPass<FuncOp>(createConvertLinalgToLoopsPass());
 
   // Run a some cleanups.
   if (options.optimize) {
-    pm.addPass(createCanonicalizerPass());
-    pm.addPass(createCSEPass());
+    pm.addNestedPass<FuncOp>(createCanonicalizerPass());
+    pm.addNestedPass<FuncOp>(createCSEPass());
   }
 
   // --------------------------------------------------------------------------
@@ -277,7 +277,7 @@ void mlir::NPCOMP::createRefBackendLoweringPipeline(
   // --------------------------------------------------------------------------
 
   // Convert scf to std control flow in preparation for going to LLVM.
-  pm.addPass(createLowerToCFGPass());
+  pm.addNestedPass<FuncOp>(createLowerToCFGPass());
 
   // Convert functions signatures and other constructs that interface with the
   // runtime to the `refbackrt` dialect.
@@ -291,8 +291,8 @@ void mlir::NPCOMP::createRefBackendLoweringPipeline(
   // Although LLVM will clean everything up eventually, for the sake of IR
   // clarity while still in MLIR, run some cleanups.
   if (options.optimize) {
-    pm.addPass(createCanonicalizerPass());
-    pm.addPass(createCSEPass());
+    pm.addNestedPass<FuncOp>(createCanonicalizerPass());
+    pm.addNestedPass<FuncOp>(createCSEPass());
   }
 }
 
@@ -305,6 +305,7 @@ void mlir::NPCOMP::createTCFRefBackendLoweringPipeline(
   // case of invalid broadcasts.
   //
   // TCP does not. So we need to reify the broadcasting and error checking.
+  // These all run at the module level.
   pm.addPass(createConvertTCFToStdPass());
   pm.addPass(createConvertTCFToLinalgPass());
   pm.addPass(createConvertTCFToTCPPass());

python/CMakeLists.txt

@@ -9,7 +9,7 @@ add_custom_command(
   COMMAND ${CMAKE_COMMAND} -E copy
         # TODO: Make the runtime library work for windows.
         ${CMAKE_BINARY_DIR}/lib/libNPCOMPCompilerRuntimeShlib.so
-        ${CMAKE_CURRENT_BINARY_DIR}/npcomp/compiler/_refjit_resources/libNPCOMPCompilerRuntimeShlib.so
+        ${CMAKE_CURRENT_BINARY_DIR}/npcomp/compiler/generic/backend/libNPCOMPCompilerRuntimeShlib.so
 )
 add_dependencies(NPCOMPPythonResources
   NPCOMPCompilerRuntimeShlib

python/NpcompModule.cpp

@@ -12,6 +12,7 @@
 #include "npcomp/Python/MlirInit.h"
 #include "npcomp/Python/NpcompModule.h"
 #include "npcomp/Python/PybindUtils.h"
+#include "npcomp-c/Registration.h"
 #include "llvm/Support/CommandLine.h"
 
 #ifdef NPCOMP_ENABLE_REFJIT
@@ -26,7 +27,7 @@ namespace mlir {
 namespace npcomp {
 namespace python {
 
-void defineLLVMModule(pybind11::module m) {
+static void defineLLVMModule(pybind11::module m) {
   m.def("print_help_message", []() { llvm::cl::PrintHelpMessage(); });
   m.def("add_option",
         [](std::string name, llvm::Optional<std::string> value) {
@@ -56,6 +57,12 @@ void defineLLVMModule(pybind11::module m) {
         py::arg("name"));
 }
 
+static void defineGlobals(py::module &m) {
+  m.def("register_dialects", [](MlirContext context) {
+    npcompRegisterAllDialects(context);
+  });
+}
+
 PYBIND11_MODULE(_npcomp, m) {
   // Guard the once init to happen once per process (vs module, which in
   // mondo builds can happen multiple times).
@@ -64,6 +71,8 @@ PYBIND11_MODULE(_npcomp, m) {
 
   m.doc() = "Npcomp native python bindings";
 
+  // TODO: Retire the llvm, mlir, passes, and dialect modules in favor of
+  // upstream Python bindings.
   auto llvm_m = m.def_submodule("llvm", "LLVM interop");
   defineLLVMModule(llvm_m);
 
@@ -81,6 +90,9 @@ PYBIND11_MODULE(_npcomp, m) {
   auto npcomp_dialect = m.def_submodule("dialect", "NPComp custom dialects");
   defineNpcompDialect(npcomp_dialect);
 
+  // Globals.
+  defineGlobals(m);
+
   // Optional backend modules.
   auto backend_m = m.def_submodule("backend", "Backend support");
   (void)backend_m;

python/npcomp/compiler/generic/backend/refjit.py

@@ -0,0 +1,66 @@
+#  Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+#  See https://llvm.org/LICENSE.txt for license information.
+#  SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+import os
+
+_refjit = None
+
+BACKEND_PASSES = (
+    "func(convert-scf-to-std)",
+    "func(canonicalize)",
+    "func(tcf-shape-refinement)",
+)
+
+
+def get_refjit():
+  """Dynamically resolves the refjit backend native module."""
+  global _refjit
+  if _refjit is not None:
+    return _refjit
+  try:
+    from _npcomp.backend import refjit as imported_refjit
+  except ImportError:
+    raise ImportError(
+        "The npcomp native module was not compiled with refjit support")
+  _refjit = imported_refjit
+  return _refjit
+
+
+def is_enabled() -> bool:
+  """Returns whether the backend is enabled for the current build."""
+  try:
+    _get_refjit()
+    return True
+  except ImportError:
+    return False
+
+
+def get_runtime_libs():
+  # The _refjit_resources directory is at the npcomp.compiler level.
+  resources_dir = os.path.join(os.path.dirname(__file__))
+  return [os.path.join(resources_dir, "libNPCOMPCompilerRuntimeShlib.so")]
+
+
+class JitModuleInvoker:
+  """Wrapper around a native JitModule for calling functions."""
+
+  def __init__(self, jit_module):
+    super().__init__()
+    self._jit_module = jit_module
+
+  def __getattr__(self, function_name):
+    return self.__getitem__(function_name)
+
+  def __getitem__(self, function_name):
+
+    def invoke(*args):
+      results = self._jit_module.invoke(function_name, args)
+      if len(results) == 1:
+        # De-tuple.
+        return results[0]
+      else:
+        return tuple(results)
+
+    invoke.__isnpcomp__ = True
+    return invoke

python/npcomp/compiler/numpy/backend/refjit.py

@@ -4,7 +4,11 @@
 
 import os
 
-from _npcomp import mlir
+from mlir.ir import *
+from mlir.passmanager import *
+from _npcomp import register_dialects
+from _npcomp import mlir as legacy_mlir
+from npcomp.compiler.generic.backend import refjit as refjit_backend
 from npcomp.compiler.utils import logging
 
 __all__ = [
@@ -13,45 +17,16 @@ __all__ = [
 ]
 
 FRONTEND_PASSES = (
-    "npcomp-cpa-type-inference",
+    "func(npcomp-cpa-type-inference)",
     "numpy-public-functions-to-tensor",
-    "convert-numpy-to-tcf",
-    "convert-scf-to-std",
-    "canonicalize",
-    "tcf-shape-refinement",
+    "func(convert-numpy-to-tcf)",
+    "func(convert-scf-to-std)",
+    "func(canonicalize)",
+    "func(tcf-shape-refinement)",
 )
 
-_refjit = None
-
-
-def _get_refjit():
-  """Dynamically resolves the refjit backend native module."""
-  global _refjit
-  if _refjit is not None:
-    return _refjit
-  try:
-    from _npcomp.backend import refjit as imported_refjit
-  except ImportError:
-    raise ImportError(
-        "The npcomp native module was not compiled with refjit support")
-  _refjit = imported_refjit
-  return _refjit
-
-
-def is_enabled() -> bool:
-  """Returns whether the backend is enabled for the current build."""
-  try:
-    _get_refjit()
-    return True
-  except ImportError:
-    return False
-
-
-def get_runtime_libs():
-  # The _refjit_resources directory is at the npcomp.compiler level.
-  resources_dir = os.path.join(os.path.dirname(__file__), "..", "..",
-                               "_refjit_resources")
-  return [os.path.join(resources_dir, "libNPCOMPCompilerRuntimeShlib.so")]
+# Re-export.
+is_enabled = refjit_backend.is_enabled
 
 
 class CompilerBackend:
@@ -59,37 +34,41 @@ class CompilerBackend:
 
   def __init__(self):
     super().__init__()
-    self._refjit = _get_refjit()
+    self._refjit = refjit_backend.get_refjit()
     self._debug = logging.debug_enabled()
 
-  def compile(self, imported_ir_module: mlir.ir.ModuleOp):
+  def compile(self, legacy_imported_ir_module: legacy_mlir.ir.ModuleOp):
     """Compiles an imported module.
 
     Args:
-      imported_ir_module: The MLIR module as imported from the ImportFrontend.
+      legacy_imported_ir_module: The MLIR module as imported from the
+        ImportFrontend.
     Returns:
       An opaque, backend specific module object that can be passed to load.
       The object may actually be something more specific to the backend (i.e.
       for IREE, it is a serialized VM flatbuffer) but the contract is that
       it is operated on by methods on this class.
     """
-    # Frontend.
-    pm = mlir.passes.PassManager(imported_ir_module.context)
-    pm.addPassPipelines(*FRONTEND_PASSES)
-    pm.run(imported_ir_module)
-    if self._debug:
-      logging.debug("Frontend IR:{}", imported_ir_module.to_asm())
+    # TODO: Once transitioned to new Python API, don't reparse the module.
+    with Context() as context:
+      register_dialects(context)
+      imported_module = Module.parse(legacy_imported_ir_module.to_asm())
+      # Frontend.
+      pm = PassManager.parse(",".join(FRONTEND_PASSES))
+      pm.run(imported_module)
+      if self._debug:
+        logging.debug("Frontend IR:{}", imported_module)
 
-    # Backend.
-    # Note that this is a separate pass manager purely to aid in debugging.
-    pm = mlir.passes.PassManager(imported_ir_module.context)
-    self._refjit.build_backend_compilation_pipeline(pm)
-    pm.run(imported_ir_module)
-    if self._debug:
-      logging.debug("Backend IR:{}", imported_ir_module.to_asm())
+      # Backend.
+      # Note that this is a separate pass manager purely to aid in debugging.
+      pm = PassManager()
+      self._refjit.build_backend_compilation_pipeline(pm)
+      pm.run(imported_module)
+      if self._debug:
+        logging.debug("Backend IR:{}", imported_module)
 
     jit_module = self._refjit.JITModule.from_compiled_module(
-        imported_ir_module, get_runtime_libs())
+        imported_module, refjit_backend.get_runtime_libs())
     return jit_module
 
   def load(self, jit_module):
@@ -97,25 +76,4 @@ class CompilerBackend:
 
     Since this is a JIT instead of an AOT compiler,
     """
-    return JitModuleInvoker(jit_module)
-
-
-class JitModuleInvoker:
-  """Wrapper around a native JitModule for calling functions."""
-
-  def __init__(self, jit_module):
-    super().__init__()
-    self._jit_module = jit_module
-
-  def __getitem__(self, function_name):
-
-    def invoke(*args):
-      results = self._jit_module.invoke(function_name, args)
-      if len(results) == 1:
-        # De-tuple.
-        return results[0]
-      else:
-        return tuple(results)
-
-    invoke.__isnpcomp__ = True
-    return invoke
+    return refjit_backend.JitModuleInvoker(jit_module)

python/npcomp/compiler/pytorch/backend/refjit.py

@@ -0,0 +1,69 @@
+#  Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+#  See https://llvm.org/LICENSE.txt for license information.
+#  SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+import os
+
+from mlir.ir import *
+from mlir.passmanager import *
+from npcomp.compiler.generic.backend import refjit as refjit_backend
+from npcomp.compiler.utils import logging
+
+__all__ = [
+    "is_enabled",
+    "CompilerBackend",
+]
+
+FRONTEND_PASSES = ("func(aten-recognize-kernels)", "func(convert-aten-to-tcf)",
+                   "numpy-public-functions-to-tensor", "canonicalize")
+
+# Re-export.
+is_enabled = refjit_backend.is_enabled
+
+
+class CompilerBackend:
+  """Main entry-point for the backend."""
+
+  def __init__(self):
+    super().__init__()
+    self._refjit = refjit_backend.get_refjit()
+    self._debug = logging.debug_enabled()
+
+  def compile(self, imported_module: Module):
+    """Compiles an imported module.
+
+    Args:
+      imported_module: The MLIR module consisting of funcs in the torch
+        dialect.
+    Returns:
+      An opaque, backend specific module object that can be passed to load.
+      The object may actually be something more specific to the backend (i.e.
+      for IREE, it is a serialized VM flatbuffer) but the contract is that
+      it is operated on by methods on this class.
+    """
+    # TODO: Once transitioned to new Python API, don't reparse the module.
+    with Context() as context:
+      # Frontend.
+      pm = PassManager.parse(",".join(FRONTEND_PASSES))
+      pm.run(imported_module)
+      if self._debug:
+        logging.debug("Frontend IR:{}", imported_module)
+
+      # Backend.
+      # Note that this is a separate pass manager purely to aid in debugging.
+      pm = PassManager()
+      self._refjit.build_backend_compilation_pipeline(pm)
+      pm.run(imported_module)
+      if self._debug:
+        logging.debug("Backend IR:{}", imported_module)
+
+    jit_module = self._refjit.JITModule.from_compiled_module(
+        imported_module, refjit_backend.get_runtime_libs())
+    return jit_module
+
+  def load(self, jit_module):
+    """Loads a compiled artifact into the runtime.
+
+    Since this is a JIT instead of an AOT compiler,
+    """
+    return refjit_backend.JitModuleInvoker(jit_module)