[LTC] Tensor[]? support operands type support using partial codegen (#2410)

* Tensor[]? support operands type support using partial codegen

* aten.index.Tensor support via partial codegen

* Add torch.index_put tracing support

* Added optional tensor list type support for LTC/TorchMLIR lowering

* Added comments

Co-authored-by: Gleb Kazantaev <gleb.kazantaev@cerebras.net>

build_tools/autogen_ltc_backend.yaml

@@ -1,11 +1,7 @@
 blacklist:
-# List of unsupported ops in LTC autogen because of some error
-- _index_put_impl_  # Error: TODO not sure if there are other valid types to handle here
-- _index_put_impl # Error: TODO not sure if there are other valid types to handle here
-- empty_like  # Error: TODO add support for type BaseType(name=<BaseTy.MemoryFormat: 12>)
-- index.Tensor  # Error: TODO not sure if there are other valid types to handle here
-- index_put  # Error: TODO not sure if there are other valid types to handle here
-- index_put_  # Error: TODO not sure if there are other valid types to handle here
+# Disabled in favour of `aten::index_put` which supports optional indices via `hacked_twin` JIT hack.
+# It also doesn't have confusing `unsafe` argument.
+- _index_put_impl
 
 # Ops with list of tensors output
 - split.Tensor
@@ -61,6 +57,8 @@ supported:
 - unbind_copy.int
 - split_copy.Tensor
 - split_with_sizes_copy
+- index.Tensor
+- index_put
 
 # ops required for functionalization
 - lift

e2e_testing/main.py

@@ -34,6 +34,7 @@ from .xfail_sets import (
     STABLEHLO_CRASHING_SET,
     TOSA_PASS_SET,
     LTC_XFAIL_SET,
+    LTC_CRASHING_SET,
     TORCHDYNAMO_XFAIL_SET,
     TORCHDYNAMO_CRASHING_SET
 )
@@ -114,7 +115,7 @@ def main():
     elif args.config == "lazy_tensor_core":
         config = LazyTensorCoreTestConfig()
         xfail_set = LTC_XFAIL_SET
-        crashing_set = set()
+        crashing_set = LTC_CRASHING_SET
     elif args.config == "torchdynamo":
         config = TorchDynamoTestConfig(RefBackendLinalgOnTensorsBackend())
         xfail_set = TORCHDYNAMO_XFAIL_SET

e2e_testing/xfail_sets.py

@@ -1263,6 +1263,12 @@ if torch_version_for_comparison() < version.parse("2.1.0.dev"):
         "ReshapeCollapseModule_basic",
     }
 
+LTC_CRASHING_SET = {
+    # TODO: update test to move all inputs to the lazy device. Otherwise test fails with:
+    # Check failed: lazy_tensor Input tensor is not a lazy tensor: CPUBoolType.
+    "HBC_basic",
+}
+
 LTC_XFAIL_SET = {
     "_Convolution2DAllFalseModule_basic",
     "_Convolution2DBenchmarkModule_basic",
@@ -1295,31 +1301,6 @@ LTC_XFAIL_SET = {
     "GeIntModule_basic",
     "GtFloatIntModule_basic",
     "GtIntModule_basic",
-    "HBC_basic",
-    "IndexPut1DFloatAccumulateModule_basic",
-    "IndexPut1DFloatNonAccumulateModule_basic",
-    "IndexPut1DIntAccumulateModule_basic",
-    "IndexPut1DIntNonAccumulateModule_basic",
-    "IndexPut2DFloatAccumulateModule_basic",
-    "IndexPut2DFloatNonAccumulateModule_basic",
-    "IndexPut2DIntAccumulateModule_basic",
-    "IndexPut2DIntNonAccumulateModule_basic",
-    "IndexPut3DFloatAccumulateModule_basic",
-    "IndexPut3DFloatNonAccumulateModule_basic",
-    "IndexPut3DIntAccumulateModule_basic",
-    "IndexPut3DIntNonAccumulateModule_basic",
-    "IndexPutHackedTwin1DFloatAccumulateModule_basic",
-    "IndexPutHackedTwin1DFloatNonAccumulateModule_basic",
-    "IndexPutHackedTwin1DIntAccumulateModule_basic",
-    "IndexPutHackedTwin1DIntNonAccumulateModule_basic",
-    "IndexPutHackedTwin2DFloatAccumulateModule_basic",
-    "IndexPutHackedTwin2DFloatNonAccumulateModule_basic",
-    "IndexPutHackedTwin2DIntAccumulateModule_basic",
-    "IndexPutHackedTwin2DIntNonAccumulateModule_basic",
-    "IndexPutHackedTwin3DFloatAccumulateModule_basic",
-    "IndexPutHackedTwin3DFloatNonAccumulateModule_basic",
-    "IndexPutHackedTwin3DIntAccumulateModule_basic",
-    "IndexPutHackedTwin3DIntNonAccumulateModule_basic",
     "IndexPutImpl1DFloatAccumulateModule_basic",
     "IndexPutImpl1DFloatNonAccumulateModule_basic",
     "IndexPutImpl1DIntAccumulateModule_basic",
@@ -1331,27 +1312,6 @@ LTC_XFAIL_SET = {
     "IndexPutImpl3DFloatAccumulateModule_basic",
     "IndexPutImpl3DFloatNonAccumulateModule_basic",
     "IndexPutImplIndexWithNoneModule_basic",
-    "IndexTensorModule3dInput_basic",
-    "IndexTensorModule_basic",
-    "IndexTensorStaticModule_basic",
-    "IndexTensorMultiIndexStaticModule_basic",
-    "IndexTensorMultiInputContiguousCenter_basic",
-    "IndexTensorMultiInputNonContiguous_basic",
-    "IndexTensorMultiInputOneDim_basic",
-    "IndexTensorMultiInputThreeIndexers_basic",
-    "IndexTensorMultiInput_basic",
-    "IndexTensorSelectDimModule_basic",
-    "IndexTensorMultiInputContiguousOneDimDynamic_basic",
-    "IndexTensorMultiInputNonContiguousOneDimDynamic_basic",
-    "IndexTensorMultiInputNonContiguousDynamic_basic",
-    "IndexTensorMultiInputNonContiguousMultipleStaticDims_basic",
-    "IndexTensorStaticContiguousWithNoneModule_basic",
-    "IndexTensorStaticNonContiguousWithNoneModule_basic",
-    "IndexTensorHackedTwinModule_basic",
-    "IndexTensorHackedTwinModule3dInput_basic",
-    "IndexTensorHackedTwinMultiInputNonContiguousMultipleStaticDims_basic",
-    "IndexTensorDyanmicInputContiguousWithNoneModule_basic",
-    "IndexTensorDyanmicInputNonContiguousWithNoneModule_basic",
     "LiftFreshCopyModule_basic",
     "Matmul_dot",
     "MulIntModule_basic",
@@ -1421,7 +1381,5 @@ LTC_XFAIL_SET = {
     "AtenComplexViewModule_basic",
     "ScatterValueFloatModule_basic",
     "ScatterValueIntModule_basic",
-    "IndexTensorNegativeIndexModule_basic",
     "UniformStaticShapeModule_basic",
-    "UnsafeIndexPutHackedTwin1DFloatNonAccumulateModule_basic",
 }

python/torch_mlir/csrc/base_lazy_backend/CMakeLists.txt

@@ -71,6 +71,8 @@ add_library(torch_mlir_ltc_backend SHARED
   mlir_node.cpp
   ops/device_data.cpp
   ops/generic.cpp
+  ops/index.cpp
+  ops/ivalue.cpp
   ops/split.cpp
   ops/unbind_int.cpp
   utils/jit_utils.cpp

python/torch_mlir/csrc/base_lazy_backend/mlir_native_functions.cpp

@@ -11,6 +11,7 @@
 //===----------------------------------------------------------------------===//
 
 #include <ATen/CompositeExplicitAutogradNonFunctionalFunctions.h>
+#include <ATen/CompositeExplicitAutogradFunctions.h>
 #include <ATen/FunctionalTensorWrapper.h>
 #include <ATen/InferSize.h>
 #include <ATen/MetaFunctions.h>
@@ -34,6 +35,8 @@
 #include "ops/to_copy.h"
 #include "ops/unbind_int.h"
 #include "ops/split.h"
+#include "ops/index.h"
+#include "ops/ivalue.h"
 #include "utils/exception.h"
 #include "utils/sys_utils.h"
 
@@ -71,6 +74,15 @@ std::vector<at::Tensor> to_meta(at::ITensorListRef t_list) {
   }
   return outs;
 }
+
+c10::List<c10::optional<at::Tensor>> to_meta(const c10::List<c10::optional<at::Tensor>>& t_list) {
+  c10::List<c10::optional<at::Tensor>> outs;
+  outs.reserve(t_list.size());
+  for (const auto& tensor : t_list) {
+    outs.push_back(to_meta(tensor));
+  }
+  return outs;
+}
 } // namespace
 
 namespace torch {
@@ -517,6 +529,89 @@ std::vector<at::Tensor> LazyNativeFunctions::split_copy_symint(const at::Tensor
   return result;
 }
 
+at::Tensor LazyNativeFunctions::index(const at::Tensor & self, const c10::List<c10::optional<at::Tensor>> & indices) {
+  TORCH_LAZY_FN_COUNTER("lazy::");
+  auto common_device = torch::lazy::GetBackendDevice(self);
+  TORCH_INTERNAL_ASSERT(common_device);
+  LazyTensorPtr lazy_self = torch::lazy::GetLtcTensorOrCreateForWrappedNumber(self, *common_device);
+
+  std::vector<torch::lazy::Value> values;
+  for (const auto & it : indices) {
+    c10::optional<at::Tensor> tensor = it;
+    LazyTensorPtr lazy_tensor = torch::lazy::TryGetLtcTensor(tensor.value_or(at::Tensor()));
+    values.push_back(lazy_tensor ? lazy_tensor->GetIrValue() : torch::lazy::Value(MakeNode<IValueConstant>(c10::IValue()), 0));
+  }
+
+  auto list = MakeNode<TorchMlirOptionalTensorList>(values);
+
+  torch::lazy::NodePtr node = torch::lazy::ReuseNode<IndexTensor>(lazy_self->GetIrValue(), list);
+
+  if (!node) {
+    auto self_meta = to_meta(self);
+    auto indices_meta = to_meta(indices);
+    auto out_meta = at::meta::index(self_meta, indices_meta);
+
+    std::vector<torch::lazy::Shape> shapes{torch::lazy::Shape(out_meta.scalar_type(), out_meta.sizes().vec())};
+    TORCH_INTERNAL_ASSERT(shapes.size() == 1);
+    if(torch::lazy::symbolicShapeEnabled()) {
+      std::vector<torch::jit::IValue> inputs = { self, indices };
+      const char* schema_str = "aten::index.Tensor(Tensor self, Tensor?[] indices) -> Tensor";
+      applySymbolicShapesOnLT(schema_str, inputs, shapes);
+    }
+
+    node = torch::lazy::MakeNode<IndexTensor>(lazy_self->GetIrValue(), list, std::move(shapes));
+    CacheNode(node);
+  }
+
+  auto result = torch::lazy::CreateAtenFromLtcTensor(
+          torch::lazy::LazyTensor::Create(std::move(node), *common_device));
+
+  return result;
+}
+
+at::Tensor LazyNativeFunctions::index_put(const at::Tensor & self, const c10::List<c10::optional<at::Tensor>> & indices, const at::Tensor & values, bool accumulate) {
+  TORCH_LAZY_FN_COUNTER("lazy::");
+  auto common_device = torch::lazy::GetBackendDevice(self);
+  TORCH_INTERNAL_ASSERT(common_device);
+  LazyTensorPtr lazy_self = torch::lazy::GetLtcTensorOrCreateForWrappedNumber(self, *common_device);
+  LazyTensorPtr lazy_valeus = torch::lazy::GetLtcTensorOrCreateForWrappedNumber(values, *common_device);
+
+  std::vector<torch::lazy::Value> indices_vector;
+  for (const auto & it : indices) {
+    c10::optional<at::Tensor> tensor = it;
+    LazyTensorPtr lazy_tensor = torch::lazy::TryGetLtcTensor(tensor.value_or(at::Tensor()));
+    indices_vector.push_back(lazy_tensor ? lazy_tensor->GetIrValue() : torch::lazy::Value(MakeNode<IValueConstant>(c10::IValue()), 0));
+  }
+
+  auto indices_list = MakeNode<TorchMlirOptionalTensorList>(indices_vector);
+
+  torch::lazy::NodePtr node = torch::lazy::ReuseNode<IndexPut>(lazy_self->GetIrValue(), indices_list, lazy_valeus->GetIrValue(), accumulate);
+
+  if (!node) {
+    auto self_meta = to_meta(self);
+    auto indices_meta = to_meta(indices);
+    auto values_meta = to_meta(values);
+
+    auto out_meta = at::compositeexplicitautograd::index_put(self_meta, indices_meta, values_meta, accumulate);
+
+    std::vector<torch::lazy::Shape> shapes{torch::lazy::Shape(out_meta.scalar_type(), out_meta.sizes().vec())};
+    TORCH_INTERNAL_ASSERT(shapes.size() == 1);
+    if(torch::lazy::symbolicShapeEnabled()) {
+      std::vector<torch::jit::IValue> inputs = { self, indices, values };
+      const char* schema_str = "aten::index_put(Tensor self, Tensor?[] indices, Tensor values, bool accumulate=False) -> Tensor";
+      applySymbolicShapesOnLT(schema_str, inputs, shapes);
+    }
+
+    node = torch::lazy::MakeNode<IndexPut>(lazy_self->GetIrValue(), indices_list, lazy_valeus->GetIrValue(), accumulate, std::move(shapes));
+    CacheNode(node);
+  }
+
+  auto result = torch::lazy::CreateAtenFromLtcTensor(
+          torch::lazy::LazyTensor::Create(std::move(node), *common_device));
+
+  return result;
+}
+
 // This is needed by the torch.tensor constructor.
 // LazyTensor always opts into functionalization.
 // "lifting" a tensor for functionalization means wrapping it in a FunctionalTensorWrapper object.

python/torch_mlir/csrc/base_lazy_backend/mlir_node.cpp

@@ -120,5 +120,38 @@ torch::lazy::TorchMlirOpVector TorchMlirTensorList::Lower(
   return {listnode->output()};
 }
 
+///////////////////////////////////////////////////////////////////////////////
+// TorchMlirOptionalTensorList
+///////////////////////////////////////////////////////////////////////////////
+
+OpKind TorchMlirOptionalTensorList::ClassOpKind() {
+  // Note: this OpKind is separate from ltc_ops.h since it would be a circular
+  // import otherwise
+  static const OpKind tensor_list_opkind =
+      OpKind::Get("lazy_tensors::optional_tensor_list");
+  return tensor_list_opkind;
+}
+
+TorchMlirOptionalTensorList::TorchMlirOptionalTensorList(OpList values)
+    : TorchMlirNode(
+          /*op=*/TorchMlirOptionalTensorList::ClassOpKind(),
+          /*operands=*/values,
+          /*shapes=*/std::vector<Shape>(),
+          /*num_outputs=*/1,
+          /*hash_seed=*/kHashSeed) {}
+
+torch::lazy::TorchMlirOpVector TorchMlirOptionalTensorList::Lower(
+    TorchMlirFunction function, TorchMlirLoweringContext* loctx) const {
+  std::vector<torch::jit::Value*> tensor_list;
+  CHECK(!operands().empty());
+  for (const torch::lazy::Output& operand : operands()) {
+    tensor_list.emplace_back(loctx->GetOutputOp(operand));
+  }
+  auto graph = function->graph();
+  auto listnode =
+      graph->insertNode(graph->createList(c10::OptionalType::create(c10::TensorType::get()), tensor_list));
+  return {listnode->output()};
+}
+
 } // namespace lazy
 } // namespace torch

python/torch_mlir/csrc/base_lazy_backend/mlir_node.h

@@ -91,5 +91,18 @@ struct TORCH_API TorchMlirTensorList : public TorchMlirNode {
       TorchMlirLoweringContext* loctx) const override;
 };
 
+// TorchMlirOptionalTensorList is similar to TorchMlirTensorList but it can also represent
+// optional tensors, so the output type for this op is !torch.list<optional<vtensor>>.
+struct TORCH_API TorchMlirOptionalTensorList : public TorchMlirNode {
+  static OpKind ClassOpKind();
+
+  TorchMlirOptionalTensorList() = delete;
+  TorchMlirOptionalTensorList(OpList values);
+
+  torch::lazy::TorchMlirOpVector Lower(
+      TorchMlirFunction function,
+      TorchMlirLoweringContext* loctx) const override;
+};
+
 } // namespace lazy
 } // namespace torch

python/torch_mlir/csrc/base_lazy_backend/mlir_node_lowering.cpp

@@ -43,7 +43,12 @@ TorchMlirOpVector LowerTorchMlirBuiltin(
   for (auto arg : arguments) {
     torch::jit::Value* value = arg.value(dummy_graph);
     if (value->type()->kind() == c10::TypeKind::ListType) {
-      value->setType(c10::ListType::create(c10::TensorType::get()));
+      auto list_element_type = value->type()->cast<c10::ListType>()->getElementType();
+      if (list_element_type->cast<c10::OptionalType>()) {
+        value->setType(c10::ListType::create(c10::OptionalType::create(c10::TensorType::get())));
+      } else {
+        value->setType(c10::ListType::create(c10::TensorType::get()));
+      }
     }
   }
 

python/torch_mlir/csrc/base_lazy_backend/ops/index.cpp

@@ -0,0 +1,99 @@
+//===- index.cpp ----------------------------------------------------------===//
+//
+// 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
+// Also available under a BSD-style license. See LICENSE.
+//
+//===----------------------------------------------------------------------===//
+
+#include "index.h"
+
+namespace torch {
+namespace lazy {
+
+IndexTensor::IndexTensor(const torch::lazy::Value& self,
+                         const torch::lazy::Value& indices,
+                         std::vector<torch::lazy::Shape>&& shapes)
+    : torch::lazy::TorchMlirNode(IndexTensor::ClassOpKind(),
+                                 OpList{self, indices}, std::move(shapes),
+                                 /* num_outputs */ 1, torch::lazy::MHash()) {}
+
+std::string IndexTensor::ToString() const {
+  std::stringstream ss;
+  ss << torch::lazy::TorchMlirNode::ToString();
+  return ss.str();
+}
+
+bool IndexTensor::CanBeReused(const torch::lazy::Value& self,
+                              const torch::lazy::Value& indices) const {
+  return false;
+}
+
+TorchMlirOpVector IndexTensor::Lower(TorchMlirFunction function,
+                                     TorchMlirLoweringContext* loctx) const {
+  PRINT_FUNCTION();
+  std::vector<torch::jit::NamedValue> arguments;
+  std::vector<torch::jit::NamedValue> kwarguments;
+  arguments.reserve(2);
+  kwarguments.reserve(0);
+
+  size_t i = 0;
+  arguments.emplace_back(loctx->GetOutputOp(operand(i++)));
+  arguments.emplace_back(loctx->GetOutputOp(operand(i++)));
+
+  torch::lazy::TorchMlirOpVector index_out = torch::lazy::LowerTorchMlirBuiltin(
+      function, op().op, shapes(), arguments, kwarguments);
+  TORCH_CHECK_EQ(index_out.size(), 1);
+
+  return index_out;
+}
+
+IndexPut::IndexPut(const torch::lazy::Value& self,
+                   const torch::lazy::Value& indices,
+                   const torch::lazy::Value& values, bool accumulate,
+                   std::vector<torch::lazy::Shape>&& shapes)
+    : torch::lazy::TorchMlirNode(
+          IndexPut::ClassOpKind(), OpList{self, indices, values},
+          std::move(shapes),
+          /* num_outputs */ 1, torch::lazy::MHash(accumulate)),
+      accumulate(accumulate) {}
+
+std::string IndexPut::ToString() const {
+  std::stringstream ss;
+  ss << torch::lazy::TorchMlirNode::ToString();
+  ss << ", accumulate=" << accumulate;
+  return ss.str();
+}
+
+bool IndexPut::CanBeReused(const torch::lazy::Value& self,
+                           const torch::lazy::Value& indices,
+                           const torch::lazy::Value& values,
+                           bool accumulate) const {
+  return false;
+}
+
+TorchMlirOpVector IndexPut::Lower(TorchMlirFunction function,
+                                  TorchMlirLoweringContext* loctx) const {
+  PRINT_FUNCTION();
+  std::vector<torch::jit::NamedValue> arguments;
+  std::vector<torch::jit::NamedValue> kwarguments;
+  arguments.reserve(4);
+  kwarguments.reserve(0);
+
+  size_t i = 0;
+  arguments.emplace_back(loctx->GetOutputOp(operand(i++)));
+  arguments.emplace_back(loctx->GetOutputOp(operand(i++)));
+  arguments.emplace_back(loctx->GetOutputOp(operand(i++)));
+  arguments.emplace_back("accumulate", accumulate);
+
+  torch::lazy::TorchMlirOpVector index_out = torch::lazy::LowerTorchMlirBuiltin(
+      function, op().op, shapes(), arguments, kwarguments);
+
+  TORCH_CHECK_EQ(index_out.size(), 1);
+
+  return index_out;
+}
+
+}  // namespace lazy
+}  // namespace torch

python/torch_mlir/csrc/base_lazy_backend/ops/index.h

@@ -0,0 +1,58 @@
+//===- index.h ------------------------------------------------------------===//
+//
+// 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
+// Also available under a BSD-style license. See LICENSE.
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#include "../mlir_node.h"
+
+namespace torch {
+namespace lazy {
+
+class IndexTensor : public torch::lazy::TorchMlirNode {
+ public:
+  static torch::lazy::OpKind ClassOpKind() {
+    return torch::lazy::OpKind(at::aten::index);
+  }
+
+  IndexTensor(const torch::lazy::Value& self, const torch::lazy::Value& indices,
+              std::vector<torch::lazy::Shape>&& shapes);
+
+  std::string ToString() const override;
+
+  bool CanBeReused(const torch::lazy::Value& self,
+                   const torch::lazy::Value& indices) const;
+
+  TorchMlirOpVector Lower(TorchMlirFunction function,
+                          TorchMlirLoweringContext* loctx) const override;
+};
+
+class IndexPut : public torch::lazy::TorchMlirNode {
+ public:
+  static torch::lazy::OpKind ClassOpKind() {
+    return torch::lazy::OpKind(at::aten::index_put);
+  }
+
+  IndexPut(const torch::lazy::Value& self, const torch::lazy::Value& indices,
+           const torch::lazy::Value& values, bool accumulate,
+           std::vector<torch::lazy::Shape>&& shapes);
+
+  std::string ToString() const override;
+
+  bool CanBeReused(const torch::lazy::Value& self,
+                   const torch::lazy::Value& indices,
+                   const torch::lazy::Value& values, bool accumulate) const;
+
+  TorchMlirOpVector Lower(TorchMlirFunction function,
+                          TorchMlirLoweringContext* loctx) const override;
+
+  bool accumulate;
+};
+
+}  // namespace lazy
+}  // namespace torch

python/torch_mlir/csrc/base_lazy_backend/ops/ivalue.cpp

@@ -0,0 +1,36 @@
+//===- ivalue.cpp
+//----------------------------------------------------------===//
+//
+// 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
+// Also available under a BSD-style license. See LICENSE.
+//
+//===----------------------------------------------------------------------===//
+
+#include "ivalue.h"
+
+#include <ATen/core/ivalue.h>
+
+namespace torch {
+namespace lazy {
+
+IValueConstant::IValueConstant(const c10::IValue& value)
+    : torch::lazy::TorchMlirNode(IValueConstant::ClassOpKind(), OpList{},
+                                 std::vector<Shape>{},
+                                 /* num_outputs */ 1, torch::lazy::MHash()),
+      value(value) {}
+
+std::string IValueConstant::ToString() const {
+  std::stringstream ss;
+  ss << torch::lazy::TorchMlirNode::ToString();
+  return ss.str();
+}
+
+TorchMlirOpVector IValueConstant::Lower(TorchMlirFunction function,
+                                        TorchMlirLoweringContext* loctx) const {
+  return {loctx->graph()->insertConstant(value)};
+}
+
+}  // namespace lazy
+}  // namespace torch

python/torch_mlir/csrc/base_lazy_backend/ops/ivalue.h

@@ -0,0 +1,37 @@
+//===- index.h ------------------------------------------------------------===//
+//
+// 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
+// Also available under a BSD-style license. See LICENSE.
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#include "../mlir_node.h"
+
+namespace torch {
+namespace lazy {
+
+// IValueConstant IR Node represents a `prim::Constant` constructed with IValue
+// parameter which is helpfull in different usecases when we need custom
+// native ops lowering to torch-mlir IR nodes.
+class IValueConstant : public torch::lazy::TorchMlirNode {
+ public:
+  static torch::lazy::OpKind ClassOpKind() {
+    return torch::lazy::OpKind(at::prim::Constant);
+  }
+
+  IValueConstant(const c10::IValue& value);
+
+  std::string ToString() const override;
+
+  TorchMlirOpVector Lower(TorchMlirFunction function,
+                          TorchMlirLoweringContext* loctx) const override;
+
+  c10::IValue value;
+};
+
+}  // namespace lazy
+}  // namespace torch