fix: Upgrade Torch version, enable options

- Upgrade Torch version across the stack
- Update Dynamo sample with advanced usage to indicate usage of new
`options` argument in `torch.compile`
- Enable options argument in `torch.compile` including improved input
handling in the default torch_tensorrt backend
- ResNet example now features `torch_tensorrt.dynamo.compile`, while
transformers example features `torch_tensorrt.compile(...,
ir="dynamo_compile", ...)`
- Fix bugs in core runtime and `TRTInterpreter` to address issues
arising with latest PyTorch distribution
- Add feature in `TRTInterpreter` to specify output data types
- Add `pass_through_build_failures` argument to
`torch_tensorrt.dynamo.torch_compile` frontend

Author: gs-olive

.bazelrc

@@ -22,7 +22,7 @@
 # +------------------------------------------------------------+
 # Enable colorful output of GCC
 build --cxxopt="-fdiagnostics-color=always"
-build --cxxopt='-std=c++14'
+build --cxxopt='-std=c++17'
 #build --linkopt="-Wl,--no-as-needed"
 
 

.circleci/config.yml

@@ -269,10 +269,10 @@ commands:
     parameters:
       torch-build:
         type: string
-        default: "2.1.0.dev20230419+cu118"
+        default: "2.1.0.dev20230601+cu118"
       torchvision-build:
         type: string
-        default: "0.16.0.dev20230419+cu118"
+        default: "0.16.0.dev20230601+cu118"
       torch-build-index:
         type: string
         default: "https://download.pytorch.org/whl/nightly/cu118"
@@ -1352,10 +1352,10 @@ parameters:
   # Nightly platform config
   torch-build:
     type: string
-    default: "2.1.0.dev20230419+cu118"
+    default: "2.1.0.dev20230601+cu118"
   torchvision-build:
     type: string
-    default: "0.16.0.dev20230419+cu118"
+    default: "0.16.0.dev20230601+cu118"
   torch-build-index:
     type: string
     default: "https://download.pytorch.org/whl/nightly/cu118"

CMakeLists.txt

@@ -2,8 +2,8 @@
 cmake_minimum_required(VERSION 3.17)
 project(Torch-TensorRT LANGUAGES CXX)
 
-# use c++14 like PyTorch
-set(CMAKE_CXX_STANDARD 14)
+# use c++17 like PyTorch
+set(CMAKE_CXX_STANDARD 17)
 
 # Build the libraries with -fPIC
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)

README.md

@@ -116,7 +116,7 @@ torch.jit.save(trt_ts_module, "trt_torchscript_module.ts") # save the TRT embedd
 These are the following dependencies used to verify the testcases. Torch-TensorRT can work with other versions, but the tests are not guaranteed to pass.
 
 - Bazel 5.2.0
-- Libtorch 2.1.0.dev20230419 (built with CUDA 11.8)
+- Libtorch 2.1.0.dev20230601 (built with CUDA 11.8)
 - CUDA 11.8
 - cuDNN 8.8.0
 - TensorRT 8.6.1

WORKSPACE

@@ -51,17 +51,17 @@ new_local_repository(
 http_archive(
     name = "libtorch",
     build_file = "@//third_party/libtorch:BUILD",
-    sha256 = "1a526a9cd19c1015674d26921dbb94bcd2d632a6f9c431a21c43f4e24768d834",
+    sha256 = "c8407ae3462c344ae3814e82023e22ece759ebe75023f35bdf62e9c0a7e79035",
     strip_prefix = "libtorch",
-    urls = ["https://download.pytorch.org/libtorch/nightly/cu118/libtorch-cxx11-abi-shared-with-deps-2.1.0.dev20230419%2Bcu118.zip"],
+    urls = ["https://download.pytorch.org/libtorch/nightly/cu118/libtorch-cxx11-abi-shared-with-deps-2.1.0.dev20230601%2Bcu118.zip"],
 )
 
 http_archive(
     name = "libtorch_pre_cxx11_abi",
     build_file = "@//third_party/libtorch:BUILD",
-    sha256 = "60c5912a5085a6a7073b3804b10d41d6cc054693bbeb7a45e0247050c2837bac",
+    sha256 = "76f983bd6d784cc0a95c679034d297abe36911c16b2188498b13a9028177e28e",
     strip_prefix = "libtorch",
-    urls = ["https://download.pytorch.org/libtorch/nightly/cu118/libtorch-shared-with-deps-2.1.0.dev20230419%2Bcu118.zip"],
+    urls = ["https://download.pytorch.org/libtorch/nightly/cu118/libtorch-shared-with-deps-2.1.0.dev20230601%2Bcu118.zip"],
 )
 
 # Download these tarballs manually from the NVIDIA website

core/runtime/TRTEngine.cpp

@@ -111,7 +111,8 @@ TRTEngine::TRTEngine(
     for (size_t pyt_idx = 0; pyt_idx < inputs_size; pyt_idx++) {
       auto binding_name = _in_binding_names[pyt_idx];
       auto trt_idx = cuda_engine->getBindingIndex(binding_name.c_str());
-      std::string engine_binded_name = cuda_engine->getIOTensorName(pyt_idx);
+      std::string engine_binded_name = cuda_engine->getIOTensorName(trt_idx);
+
       TORCHTRT_CHECK(
           (binding_name == engine_binded_name),
           "Could not find a TensorRT engine binding for input named " << binding_name);

examples/dynamo/dynamo_compile_advanced_usage.py

@@ -9,7 +9,6 @@
 # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 import torch
-from torch_tensorrt.dynamo.backend import create_backend
 from torch_tensorrt.fx.lower_setting import LowerPrecision
 
 # %%
@@ -39,15 +38,19 @@ def forward(self, x: torch.Tensor, y: torch.Tensor):
 
 # Next, we compile the model using torch.compile
 # For the default settings, we can simply call torch.compile
-# with the backend "tensorrt", and run the model on an
+# with the backend "torch_tensorrt", and run the model on an
 # input to cause compilation, as so:
-optimized_model = torch.compile(model, backend="tensorrt")
+optimized_model = torch.compile(model, backend="torch_tensorrt")
 optimized_model(*sample_inputs)
 
 # %%
 # Compilation with `torch.compile` Using Custom Settings
 # ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
+# First, we use Torch utilities to clean up the workspace
+# after the previous compile invocation
+torch._dynamo.reset()
+
 # Define sample half inputs and initialize model
 sample_inputs_half = [
     torch.rand((5, 7)).half().cuda(),
@@ -58,20 +61,25 @@ def forward(self, x: torch.Tensor, y: torch.Tensor):
 # %%
 
 # If we want to customize certain options in the backend,
-# but still use the torch.compile call directly, we can call the
-# convenience/helper function create_backend to create a custom backend
-# which has been pre-populated with certain keys
-custom_backend = create_backend(
-    lower_precision=LowerPrecision.FP16,
-    debug=True,
-    min_block_size=2,
-    torch_executed_ops={},
-    optimization_level=4,
-    use_experimental_rt=True,
-)
+# but still use the torch.compile call directly, we can provide
+# custom options to the backend via the "options" keyword
+# which takes in a dictionary mapping options to values.
+#
+# For accepted backend options, see the CompilationSettings dataclass:
+# py/torch_tensorrt/dynamo/backend/_settings.py
+backend_kwargs = {
+    "lower_precision": LowerPrecision.FP16,
+    "debug": True,
+    "min_block_size": 2,
+    "torch_executed_ops": {"torch.ops.aten.sub.Tensor"},
+    "optimization_level": 4,
+    "use_experimental_rt": True,
+}
 
 # Run the model on an input to cause compilation, as so:
-optimized_model_custom = torch.compile(model_half, backend=custom_backend)
+optimized_model_custom = torch.compile(
+    model_half, backend="torch_tensorrt", options=backend_kwargs
+)
 optimized_model_custom(*sample_inputs_half)
 
 # %%

py/requirements.txt

@@ -2,7 +2,7 @@ numpy
 packaging
 pybind11==2.6.2
 --extra-index-url https://download.pytorch.org/whl/nightly/cu118
-torch==2.1.0.dev20230419+cu118
-torchvision==0.16.0.dev20230419+cu118
+torch==2.1.0.dev20230601+cu118
+torchvision==0.16.0.dev20230601+cu118
 --extra-index-url https://pypi.ngc.nvidia.com
 tensorrt==8.6.1

py/torch_tensorrt/dynamo/backend/__init__.py

@@ -49,6 +49,7 @@ def compile(
     min_block_size=MIN_BLOCK_SIZE,
     torch_executed_ops=[],
     torch_executed_modules=[],
+    pass_through_build_failures=PASS_THROUGH_BUILD_FAILURES,
     max_aux_streams=MAX_AUX_STREAMS,
     version_compatible=VERSION_COMPATIBLE,
     optimization_level=OPTIMIZATION_LEVEL,
@@ -94,6 +95,7 @@ def compile(
         workspace_size=workspace_size,
         min_block_size=min_block_size,
         torch_executed_ops=torch_executed_ops,
+        pass_through_build_failures=pass_through_build_failures,
         max_aux_streams=max_aux_streams,
         version_compatible=version_compatible,
         optimization_level=optimization_level,

py/torch_tensorrt/dynamo/backend/backends.py

@@ -2,6 +2,7 @@
 from typing import Sequence
 import torch
 from functools import partial
+from dataclasses import replace, fields
 import torch._dynamo as td
 
 from torch_tensorrt.dynamo.backend._settings import CompilationSettings
@@ -28,7 +29,24 @@ def torch_tensorrt_backend(
     gm: torch.fx.GraphModule,
     sample_inputs: Sequence[torch.Tensor],
     settings: CompilationSettings = CompilationSettings(),
+    **kwargs
 ):
+    # If the user specifies keyword args, overwrite those fields in settings
+    # Validate all specified kwargs to ensure they are true fields of the dataclass
+    #
+    # Note: kwargs provided by torch.compile are wrapped in the "options" key
+    if kwargs:
+        if "options" in kwargs and len(kwargs) == 1:
+            kwargs = kwargs["options"]
+
+        valid_attrs = {attr.name for attr in fields(settings)}
+        valid_kwargs = {k: v for k, v in kwargs.items() if k in valid_attrs}
+        settings = replace(settings, **valid_kwargs)
+
+    # Enable debug/verbose mode if requested
+    if settings.debug:
+        logger.setLevel(logging.DEBUG)
+
     DEFAULT_BACKEND = aot_torch_tensorrt_aten_backend
 
     return DEFAULT_BACKEND(gm, sample_inputs, settings=settings)