Wrapper module around TRT + pytorch subgraphs (#3270)

Author: keehyuna

core/runtime/execute_engine.cpp

@@ -94,6 +94,7 @@ bool _validate_shapes(std::vector<at::Tensor> inputs, c10::intrusive_ptr<TRTEngi
 void setup_input_tensors(
     std::vector<at::Tensor> inputs,
     c10::intrusive_ptr<TRTEngine> compiled_engine,
+    bool cudagraphs_enabled,
     bool need_cudagraphs_record) {
   // this is a buffer to store shape tensor input addresses throughout the runtime scope
   std::list<std::vector<int64_t>> inputShapeTensorValues;
@@ -127,7 +128,7 @@ void setup_input_tensors(
           compiled_engine->exec_ctx->setTensorAddress(name.c_str(), inputShapeTensorValues.back().data()),
           "Error while setting the tensor address for shape inputs");
 
-      if (CUDAGRAPHS_MODE) {
+      if (cudagraphs_enabled) {
         // @peri044 I dont know if this makes sense since they are supposed to be GPU buffers
         compiled_engine->input_buffers[i] = input_cpu;
       }
@@ -147,7 +148,7 @@ void setup_input_tensors(
       TORCHTRT_CHECK(
           compiled_engine->exec_ctx->setInputShape(name.c_str(), dims), "Error while setting the input shape");
 
-      if (CUDAGRAPHS_MODE) {
+      if (cudagraphs_enabled) {
         // If using CUDAGraphs copy formatted input to the corresponding persistent input buffer
         compiled_engine->input_buffers[i].copy_(formatted_inputs.back(), true);
         TORCHTRT_CHECK(
@@ -201,17 +202,17 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
     LOG_INFO("" << log_info);
     compiled_engine->cudagraph.enable_debug_mode();
   }
-
+  bool cudagraphs_enabled = (CUDAGRAPHS_MODE == SUBGRAPH_CUDAGRAPHS);
   bool shape_changed = _validate_shapes(inputs, compiled_engine);
 
   // Whether cudagraphs needs to record the graph on this pass
   auto result = compiled_engine->runtime_states.set_runtime_states(
-      CUDAGRAPHS_MODE, compiled_engine->use_pre_allocated_outputs, shape_changed);
+      cudagraphs_enabled, compiled_engine->use_pre_allocated_outputs, shape_changed);
 
   bool need_cudagraphs_record = std::get<0>(result);
   bool can_use_pre_allocated_outputs = std::get<1>(result);
 
-  if (!CUDAGRAPHS_MODE || shape_changed) {
+  if (!cudagraphs_enabled || shape_changed) {
     compiled_engine->cudagraph.reset();
   }
 
@@ -273,8 +274,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
           std::make_unique<torch::autograd::profiler::RecordProfile>(compiled_engine->input_profile_path);
     }
 
-    setup_input_tensors(inputs, compiled_engine, need_cudagraphs_record);
-
+    setup_input_tensors(inputs, compiled_engine, cudagraphs_enabled, need_cudagraphs_record);
     // Check if input shapes can be inferred.
     int32_t const io_size{compiled_engine->cuda_engine->getNbIOTensors()};
     std::vector<char const*> names(io_size);
@@ -306,7 +306,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
         compiled_engine->output_buffers[pyt_idx] = std::move(outputs[pyt_idx].clone());
       }
 
-      if (CUDAGRAPHS_MODE) {
+      if (cudagraphs_enabled) {
         TORCHTRT_CHECK(
             compiled_engine->exec_ctx->setTensorAddress(
                 name.c_str(), compiled_engine->output_buffers[pyt_idx].data_ptr()),
@@ -346,7 +346,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
     caller_exec_complete.record(compiled_engine->caller_stream);
     caller_exec_complete.block(compiled_engine->engine_stream);
 
-    if (!CUDAGRAPHS_MODE) {
+    if (!cudagraphs_enabled) {
       // Direct execution uses the caller buffers directly
       compiled_engine->exec_ctx->enqueueV3(compiled_engine->engine_stream);
     } else {
@@ -377,7 +377,7 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
   trt_exec_complete.record(compiled_engine->engine_stream);
   trt_exec_complete.block(compiled_engine->caller_stream);
 
-  if (CUDAGRAPHS_MODE) {
+  if (cudagraphs_enabled) {
     // If in CUDAGraph mode, results need to be copied to the result buffers (on caller stream)
     for (size_t o = 0; o < compiled_engine->output_buffers.size(); o++) {
       outputs[o].copy_(compiled_engine->output_buffers[o], false);

core/runtime/register_jit_hooks.cpp

@@ -112,8 +112,10 @@ TORCH_LIBRARY(tensorrt, m) {
   m.def("set_multi_device_safe_mode", [](bool multi_device_safe_mode) -> void {
     MULTI_DEVICE_SAFE_MODE = multi_device_safe_mode;
   });
-  m.def("get_cudagraphs_mode", []() -> bool { return CUDAGRAPHS_MODE; });
-  m.def("set_cudagraphs_mode", [](bool cudagraphs_mode) -> void { CUDAGRAPHS_MODE = cudagraphs_mode; });
+  m.def("get_cudagraphs_mode", []() -> int64_t { return CUDAGRAPHS_MODE; });
+  m.def("set_cudagraphs_mode", [](int64_t cudagraphs_mode) -> void {
+    CUDAGRAPHS_MODE = CudaGraphsMode(cudagraphs_mode);
+  });
   m.def("set_logging_level", [](int64_t level) -> void {
     util::logging::get_logger().set_reportable_log_level(util::logging::LogLevel(level));
   });

core/runtime/runtime.cpp

@@ -8,7 +8,7 @@ namespace core {
 namespace runtime {
 
 bool MULTI_DEVICE_SAFE_MODE = false;
-bool CUDAGRAPHS_MODE = false;
+CudaGraphsMode CUDAGRAPHS_MODE = STANDARD;
 
 c10::optional<RTDevice> get_most_compatible_device(
     const RTDevice& target_device,
@@ -130,11 +130,11 @@ void set_multi_device_safe_mode(bool multi_device_safe_mode) {
   MULTI_DEVICE_SAFE_MODE = multi_device_safe_mode;
 }
 
-bool get_cudagraphs_mode() {
+CudaGraphsMode get_cudagraphs_mode() {
   return CUDAGRAPHS_MODE;
 }
 
-void set_cudagraphs_mode(bool cudagraphs_mode) {
+void set_cudagraphs_mode(CudaGraphsMode cudagraphs_mode) {
   CUDAGRAPHS_MODE = cudagraphs_mode;
 }
 

core/runtime/runtime.h

@@ -18,7 +18,14 @@ namespace runtime {
 using EngineID = int64_t;
 const std::string ABI_VERSION = "6";
 extern bool MULTI_DEVICE_SAFE_MODE;
-extern bool CUDAGRAPHS_MODE;
+
+typedef enum {
+  STANDARD = 0,
+  SUBGRAPH_CUDAGRAPHS,
+  WHOLE_GRAPH_CUDAGRAPHS,
+} CudaGraphsMode;
+
+extern CudaGraphsMode CUDAGRAPHS_MODE;
 
 typedef enum {
   ABI_TARGET_IDX = 0,
@@ -51,9 +58,9 @@ bool get_multi_device_safe_mode();
 
 void set_multi_device_safe_mode(bool multi_device_safe_mode);
 
-bool get_cudagraphs_mode();
+CudaGraphsMode get_cudagraphs_mode();
 
-void set_cudagraphs_mode(bool cudagraphs_mode);
+void set_cudagraphs_mode(CudaGraphsMode cudagraphs_mode);
 
 class DeviceList {
   using DeviceMap = std::unordered_map<int, RTDevice>;

docsrc/user_guide/runtime.rst

@@ -86,7 +86,7 @@ Cudagraphs can accelerate certain models by reducing kernel overheads, as docume
     torch_tensorrt.runtime.set_cudagraphs_mode(False)
 
     # Enables Cudagraphs Mode, then resets the mode to its prior setting
-    with torch_tensorrt.runtime.enable_cudagraphs():
+    with torch_tensorrt.runtime.enable_cudagraphs(trt_module):
         ...
 
 In the current implementation, use of a new input shape (for instance in dynamic shape 

examples/dynamo/torch_export_cudagraphs.py

@@ -11,9 +11,8 @@
 # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 import torch
-import torchvision.models as models
-
 import torch_tensorrt
+import torchvision.models as models
 
 # %%
 # Compilation with `torch_tensorrt.compile` Using Default Settings
@@ -47,8 +46,8 @@
 # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 # We can enable the cudagraphs API with a context manager
-with torch_tensorrt.runtime.enable_cudagraphs():
-    out_trt = opt(inputs)
+with torch_tensorrt.runtime.enable_cudagraphs(opt) as cudagraphs_module:
+    out_trt = cudagraphs_module(inputs)
 
 # Alternatively, we can set the cudagraphs mode for the session
 torch_tensorrt.runtime.set_cudagraphs_mode(True)
@@ -64,6 +63,49 @@
 inputs_2 = torch.randn((8, 3, 224, 224)).cuda()
 inputs_3 = torch.randn((4, 3, 224, 224)).cuda()
 
-with torch_tensorrt.runtime.enable_cudagraphs():
-    out_trt_2 = opt(inputs_2)
-    out_trt_3 = opt(inputs_3)
+with torch_tensorrt.runtime.enable_cudagraphs(opt) as cudagraphs_module:
+    out_trt_2 = cudagraphs_module(inputs_2)
+    out_trt_3 = cudagraphs_module(inputs_3)
+
+# %%
+# Cuda graphs with module that contains graph breaks
+# ----------------------------------
+#
+# When CUDA Graphs are applied to a TensorRT model that contains graph breaks, each break introduces additional
+# overhead. This occurs because graph breaks prevent the entire model from being executed as a single, continuous
+# optimized unit. As a result, some of the performance benefits typically provided by CUDA Graphs, such as reduced
+# kernel launch overhead and improved execution efficiency, may be diminished.
+# Using a wrapped runtime module with CUDA Graphs allows you to encapsulate sequences of operations into graphs
+# that can be executed efficiently, even in the presence of graph breaks.
+# If TensorRT module has graph breaks, CUDA Graph context manager returns a wrapped_module. This module captures entire
+# execution graph, enabling efficient replay during subsequent inferences by reducing kernel launch overheads
+# and improving performance. Note that initializing with the wrapper module involves a warm-up phase where the
+# module is executed several times. This warm-up ensures that memory allocations and initializations are not
+# recorded in CUDA Graphs, which helps maintain consistent execution paths and optimize performance.
+
+
+class SampleModel(torch.nn.Module):
+    def forward(self, x):
+        return torch.relu((x + 2) * 0.5)
+
+
+model = SampleModel().eval().cuda()
+input = torch.randn((1, 3, 224, 224)).to("cuda")
+
+# The 'torch_executed_ops' compiler option is used in this example to intentionally introduce graph breaks within the module.
+# Note: The Dynamo backend is required for the CUDA Graph context manager to handle modules in an Ahead-Of-Time (AOT) manner.
+opt_with_graph_break = torch_tensorrt.compile(
+    model,
+    ir="dynamo",
+    inputs=[input],
+    min_block_size=1,
+    pass_through_build_failures=True,
+    torch_executed_ops={"torch.ops.aten.mul.Tensor"},
+)
+
+# %%
+# If module has graph breaks, whole submodules are recorded and replayed by cuda graphs
+with torch_tensorrt.runtime.enable_cudagraphs(
+    opt_with_graph_break
+) as cudagraphs_module:
+    cudagraphs_module(input)

py/torch_tensorrt/_compile.py

@@ -12,6 +12,9 @@
 from torch_tensorrt._features import ENABLED_FEATURES
 from torch_tensorrt._Input import Input
 from torch_tensorrt.dynamo import _defaults
+from torch_tensorrt.dynamo.runtime._CudaGraphsTorchTensorRTModule import (
+    CudaGraphsTorchTensorRTModule,
+)
 from torch_tensorrt.fx import InputTensorSpec
 from torch_tensorrt.fx.lower import compile as fx_compile
 from torch_tensorrt.fx.utils import LowerPrecision
@@ -586,14 +589,16 @@ def save(
     Save the model to disk in the specified output format.
 
     Arguments:
-        module (Optional(torch.jit.ScriptModule | torch.export.ExportedProgram | torch.fx.GraphModule)): Compiled Torch-TensorRT module
+        module (Optional(torch.jit.ScriptModule | torch.export.ExportedProgram | torch.fx.GraphModule | CudaGraphsTorchTensorRTModule)): Compiled Torch-TensorRT module
         inputs (torch.Tensor): Torch input tensors
         arg_inputs (Tuple[Any, ...]): Same as inputs. Alias for better understanding with kwarg_inputs.
         kwarg_inputs (dict[Any, ...]): Optional, kwarg inputs to the module forward function.
         output_format (str): Format to save the model. Options include exported_program | torchscript.
         retrace (bool): When the module type is a fx.GraphModule, this option re-exports the graph using torch.export.export(strict=False) to save it.
                 This flag is experimental for now.
     """
+    if isinstance(module, CudaGraphsTorchTensorRTModule):
+        module = module.compiled_module
     module_type = _parse_module_type(module)
     accepted_formats = {"exported_program", "torchscript"}
     if arg_inputs is not None and not all(