fix: Add lowering pass to remove output repacking

- Automatically remove output repacking for
`convert_method_to_trt_engine` calls, to improve parity between models
which can be converted directly to TRT engines, and models which can be
fully compiled
- Add new internal `CompileSpec` argument for lowering which indicates
whether the lowering passes originate from a
`convert_method_to_trt_engine` call or a regular `compile` call, which
affects whether the lowering pass is applied
- Regular TorchScript graphs cannot have this pass applied, as it can
otherwise break the output graph. Newer versions of Torch disallow graph
outputs with 0 or 2+ arguments which are not packed in a struct
- Current lowering pass detects outputs which are flat Lists or Tuples
of Tensors and returns the outputs as-is (direct from the TRT Engine),
so the entire model can be converted to a single TRT engine

Author: gs-olive

core/lowering/lowering.cpp

@@ -144,6 +144,9 @@ void LowerGraph(std::shared_ptr<torch::jit::Graph>& g, std::vector<torch::jit::I
   passes::RemoveSingleUse0DTensors(g);
   passes::RemoveUnnecessaryCasts(g);
   passes::ReplaceAtenInt(g);
+  if (lower_info.converting_to_trt_engine) {
+    passes::RemoveCollectionCast(g);
+  }
   passes::UnpackAndCastMaskedFill(g, lower_info.getGPUDeviceString());
   passes::UnpackAndCastNumToTensor(g, lower_info.getGPUDeviceString());
   passes::UnpackAndCastFull(g, lower_info.getGPUDeviceString());

core/lowering/lowering.h

@@ -16,6 +16,10 @@ struct LowerInfo {
   // Since these QDQ nodes will be identical as they share same input, one of them is eliminated due to CSE lowering
   // pass. Disable this in order to not disturb TensorRT's QAT optimizations.
   bool disable_cse = false;
+
+  // Whether the originating caller is `convert_method_to_trt_engine` (true) or `compile` (false)
+  bool converting_to_trt_engine = false;
+
   ir::Device target_device;
   std::vector<std::string> forced_fallback_modules;
   friend std::ostream& operator<<(std::ostream& os, const LowerInfo& l);

core/lowering/passes/passes.h

@@ -33,6 +33,7 @@ void RemoveNOPs(std::shared_ptr<torch::jit::Graph> graph);
 void RemoveSingleUse0DTensors(std::shared_ptr<torch::jit::Graph>& g);
 void RemoveUnnecessaryCasts(std::shared_ptr<torch::jit::Graph>& graph);
 void ReplaceAtenInt(std::shared_ptr<torch::jit::Graph>& g);
+void RemoveCollectionCast(std::shared_ptr<torch::jit::Graph>& g);
 void UnpackAddMM(std::shared_ptr<torch::jit::Graph>& graph);
 void UnpackBatchNorm(std::shared_ptr<torch::jit::Graph>& graph);
 void UnpackLogSoftmax(std::shared_ptr<torch::jit::Graph>& graph);

core/lowering/passes/remove_unnecessary_casts.cpp

@@ -356,6 +356,45 @@ void ReplaceAtenInt(std::shared_ptr<torch::jit::Graph>& g) {
   LOG_GRAPH("Post removing aten.Int.Tensor operations: " << *g);
 }
 
+void RemoveCollectionCast(std::shared_ptr<torch::jit::Graph>& g) {
+  // Removes unnecessary collection-casting of graph outputs
+  // Only to be used if the overall output is intended to be a TRT Engine
+  // Will cause errors if used directly as a TorchScript graph
+
+  // Validate the output is a single value with type Tuple or List
+  if (!(g->outputs().size() == 1 &&
+        (g->outputs()[0]->node()->kind() == torch::jit::prim::TupleConstruct ||
+         g->outputs()[0]->node()->kind() == torch::jit::prim::ListConstruct))) {
+    return;
+  }
+
+  // Ensure all inputs to the Tuple/List Construct operator are regular Tensors
+  // (nested structures cannot be preserved in TensorRT)
+  auto all_tensors = true;
+  auto collection_inputs = g->outputs()[0]->node()->inputs();
+
+  for (size_t i = 0; i < collection_inputs.size(); ++i) {
+    all_tensors &= collection_inputs[i]->type()->isSubtypeOf(c10::TensorType::get());
+  }
+
+  if (!all_tensors) {
+    return;
+  }
+
+  // For each input to the collection packing operator, add its value directly
+  // as an output of the graph
+  for (size_t i = 0; i < collection_inputs.size(); ++i) {
+    g->registerOutput(collection_inputs[i]);
+  }
+
+  // Remove the original output value of the graph (the collection object)
+  g->eraseOutput(0);
+
+  // Clean up remnant collection node in graph
+  torch::jit::EliminateDeadCode(g);
+  LOG_GRAPH("Post removing collection casting operations: " << *g);
+}
+
 } // namespace passes
 } // namespace lowering
 } // namespace core

cpp/src/compile_spec.cpp

@@ -78,9 +78,11 @@ torchtrt::core::CompileSpec init_compile_spec(CompileSpec& external) {
   }
 }
 
-torchtrt::core::CompileSpec to_internal_compile_spec(CompileSpec external) {
+torchtrt::core::CompileSpec to_internal_compile_spec(CompileSpec external, bool converting_to_trt_engine) {
   torchtrt::core::CompileSpec internal = init_compile_spec(external);
 
+  internal.lower_info.converting_to_trt_engine = converting_to_trt_engine;
+
   for (auto p : external.enabled_precisions) {
     internal.convert_info.engine_settings.enabled_precisions.insert(toTRTDataType(p));
   }

cpp/src/torch_tensorrt.cpp

@@ -10,7 +10,7 @@ namespace torch_tensorrt {
 torch_tensorrt::core::runtime::RTDevice to_internal_rt_device(Device device);
 namespace torchscript {
 // Defined in compile_spec.cpp
-torch_tensorrt::core::CompileSpec to_internal_compile_spec(CompileSpec external);
+torch_tensorrt::core::CompileSpec to_internal_compile_spec(CompileSpec external, bool converting_to_trt_engine = false);
 
 bool check_method_operator_support(const torch::jit::script::Module& module, std::string method_name) {
   return torch_tensorrt::core::CheckMethodOperatorSupport(module, method_name);
@@ -23,7 +23,8 @@ std::string convert_method_to_trt_engine(
   LOG_DEBUG(get_build_info());
   // Want to export a much simpler (non TRT header dependent) API so doing the
   // type conversion here
-  return torch_tensorrt::core::ConvertGraphToTRTEngine(module, method_name, to_internal_compile_spec(info));
+  return torch_tensorrt::core::ConvertGraphToTRTEngine(
+      module, method_name, to_internal_compile_spec(info, /*bool converting_to_trt_engine=*/true));
 }
 
 torch::jit::script::Module compile(const torch::jit::script::Module& module, CompileSpec info) {

py/torch_tensorrt/csrc/tensorrt_classes.cpp

@@ -326,9 +326,11 @@ core::CompileSpec init_compile_spec(CompileSpec external) {
   }
 }
 
-core::CompileSpec CompileSpec::toInternalCompileSpec() {
+core::CompileSpec CompileSpec::toInternalCompileSpec(bool converting_to_trt_engine) {
   core::CompileSpec info = init_compile_spec(*this);
 
+  info.lower_info.converting_to_trt_engine = converting_to_trt_engine;
+
   for (auto p : enabled_precisions) {
     info.convert_info.engine_settings.enabled_precisions.insert(toTRTDataType(p));
   }

py/torch_tensorrt/csrc/tensorrt_classes.h

@@ -123,7 +123,7 @@ std::string to_str(EngineCapability value);
 nvinfer1::EngineCapability toTRTEngineCapability(EngineCapability value);
 
 struct CompileSpec : torch::CustomClassHolder {
-  core::CompileSpec toInternalCompileSpec();
+  core::CompileSpec toInternalCompileSpec(bool converting_to_trt_engine = false);
   std::string stringify();
   void appendInput(const c10::intrusive_ptr<Input>& ir) {
     inputs.push_back(*ir);

py/torch_tensorrt/csrc/torch_tensorrt_py.cpp

@@ -158,7 +158,8 @@ torch::jit::Module CompileGraph(const torch::jit::Module& mod, CompileSpec& info
 
 py::bytes ConvertGraphToTRTEngine(const torch::jit::Module& mod, const std::string& method_name, CompileSpec& info) {
   py::gil_scoped_acquire gil;
-  auto trt_engine = core::ConvertGraphToTRTEngine(mod, method_name, info.toInternalCompileSpec());
+  auto trt_engine = core::ConvertGraphToTRTEngine(
+      mod, method_name, info.toInternalCompileSpec(/*bool converting_to_trt_engine=*/true));
   return py::bytes(trt_engine);
 }
 

tests/core/lowering/test_remove_unnecessary_casts.cpp

@@ -589,3 +589,89 @@ TEST(LoweringPasses, RemoveAtenIntConstTensorValuesAgree) {
   // Validate identical graphs after pooling constants and canonicalizing
   ASSERT_TRUE((tg->toString() == sg->toString()));
 }
+
+TEST(LoweringPasses, RemoveCollectionCastTuple) {
+  // Ensure the lowering pass transforms the first graph into the second
+  std::string source_graph = R"IR(
+    graph(%x.1 : Tensor):
+      %3 : int = prim::Constant[value=1]()
+      %2 : int = prim::Constant[value=2]()
+      %a.1 : Tensor = aten::mul(%x.1, %2)
+      %b.1 : Tensor = aten::add(%a.1, %2, %3)
+      %c.1 : Tensor = aten::relu(%b.1)
+      %d.1 : Tensor = aten::sqrt(%c.1)
+      %8 : (Tensor, Tensor, Tensor) = prim::TupleConstruct(%c.1, %d.1, %b.1)
+      return (%8))IR";
+
+  std::string target_graph = R"IR(
+    graph(%x.1 : Tensor):
+      %3 : int = prim::Constant[value=1]()
+      %2 : int = prim::Constant[value=2]()
+      %a.1 : Tensor = aten::mul(%x.1, %2)
+      %b.1 : Tensor = aten::add(%a.1, %2, %3)
+      %c.1 : Tensor = aten::relu(%b.1)
+      %d.1 : Tensor = aten::sqrt(%c.1)
+      return (%c.1, %d.1, %b.1))IR";
+
+  // Ensure the lowering pass transforms the first graph into the second
+  torch_tensorrt::core::util::logging::get_logger().set_reportable_log_level(
+      torch_tensorrt::core::util::logging::LogLevel::kGRAPH);
+  auto sg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(source_graph, sg.get());
+
+  torch_tensorrt::core::lowering::passes::RemoveCollectionCast(sg);
+  torch::jit::ConstantPooling(sg);
+  sg = torch::jit::Canonicalize(sg, false);
+
+  auto tg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(target_graph, tg.get());
+
+  torch::jit::ConstantPooling(tg);
+  tg = torch::jit::Canonicalize(tg, false);
+
+  // Validate identical graphs after pooling constants and canonicalizing
+  ASSERT_TRUE((tg->toString() == sg->toString()));
+}
+
+TEST(LoweringPasses, RemoveCollectionCastList) {
+  // Ensure the lowering pass transforms the first graph into the second
+  std::string source_graph = R"IR(
+    graph(%x.1 : Tensor):
+      %3 : int = prim::Constant[value=1]()
+      %2 : int = prim::Constant[value=2]()
+      %a.1 : Tensor = aten::mul(%x.1, %2)
+      %b.1 : Tensor = aten::add(%a.1, %2, %3)
+      %c.1 : Tensor = aten::relu(%b.1)
+      %d.1 : Tensor = aten::sqrt(%c.1)
+      %8 : (Tensor, Tensor, Tensor) = prim::ListConstruct(%b.1, %c.1, %d.1)
+      return (%8))IR";
+
+  std::string target_graph = R"IR(
+    graph(%x.1 : Tensor):
+      %3 : int = prim::Constant[value=1]()
+      %2 : int = prim::Constant[value=2]()
+      %a.1 : Tensor = aten::mul(%x.1, %2)
+      %b.1 : Tensor = aten::add(%a.1, %2, %3)
+      %c.1 : Tensor = aten::relu(%b.1)
+      %d.1 : Tensor = aten::sqrt(%c.1)
+      return (%b.1, %c.1, %d.1))IR";
+
+  // Ensure the lowering pass transforms the first graph into the second
+  torch_tensorrt::core::util::logging::get_logger().set_reportable_log_level(
+      torch_tensorrt::core::util::logging::LogLevel::kGRAPH);
+  auto sg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(source_graph, sg.get());
+
+  torch_tensorrt::core::lowering::passes::RemoveCollectionCast(sg);
+  torch::jit::ConstantPooling(sg);
+  sg = torch::jit::Canonicalize(sg, false);
+
+  auto tg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(target_graph, tg.get());
+
+  torch::jit::ConstantPooling(tg);
+  tg = torch::jit::Canonicalize(tg, false);
+
+  // Validate identical graphs after pooling constants and canonicalizing
+  ASSERT_TRUE((tg->toString() == sg->toString()));
+}