feat: Cudagraphs integration for Torch-TRT + Non-default Stream Utilization #2881

gs-olive · 2024-06-03T16:40:54Z

Description

Add option to enable Cudagraphs in the runtime for additional acceleration of TRT engines
Add C++ and Python toggles and well as full integration for C++ and Python runtimes
Add support for dynamic shape cases via shape keys with cache invalidation
Add test cases for cudagraphs support
Add support for non-default CUDA Stream usage by default

Addresses #2736

Type of change

New feature (non-breaking change which adds functionality)

Checklist:

[ x ] My code follows the style guidelines of this project (You can use the linters)
[ x ] I have performed a self-review of my own code
[ x ] I have commented my code, particularly in hard-to-understand areas and hacks
[ x ] I have made corresponding changes to the documentation
[ x ] I have added tests to verify my fix or my feature
[ x ] New and existing unit tests pass locally with my changes
[ x ] I have added the relevant labels to my PR in so that relevant reviewers are notified

narendasan · 2024-06-04T16:11:00Z

core/runtime/execute_engine.cpp

@@ -168,27 +210,66 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
      auto dims = core::util::toVec(out_shape);
      auto type = util::TRTDataTypeToScalarType(compiled_engine->exec_ctx->getEngine().getTensorDataType(name.c_str()));
      outputs[pyt_idx] = std::move(at::empty(dims, {at::kCUDA}).to(type).contiguous());
+
+      // In cudagraphs mode, the allocated output buffers are stored for reuse


Is there a downside to always using the buffers?

Also what is the required lifetime of these buffers, can they be deallocated at any point or do they need to be reserved as long as the cuda graph is active?

If not in cudagraphs mode, the main downside is the requirement that buffers need to change size/be reallocated in Dynamic shape usecases.

My understanding is that these I/O buffers need to be reserved for the duration of the cudagraph being active

core/runtime/execute_engine.cpp

py/torch_tensorrt/dynamo/runtime/_PythonTorchTensorRTModule.py

py/torch_tensorrt/runtime/cudagraphs.py

tests/py/dynamo/runtime/test_cudagraphs.py

core/runtime/execute_engine.cpp

peri044

LGTM

peri044

LGTM, pending rebase

zewenli98

LGTM

- Add option to enable Cudagraphs in the runtime for additional acceleration of TRT engines - Add C++ and Python toggles and well as full integration for C++ and Python runtimes - Add support for dynamic shape cases via shape keys with cache invalidation - Add test cases for cudagraphs support

- Add support for non-default streams

…zation (#2881)

gs-olive requested a review from narendasan June 3, 2024 16:40

gs-olive self-assigned this Jun 3, 2024

facebook-github-bot added the cla signed label Jun 3, 2024

github-actions bot requested a review from peri044 June 3, 2024 16:41