triton-inference-server / server

The Triton Inference Server provides an optimized cloud and edge inferencing solution.
https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/index.html
BSD 3-Clause "New" or "Revised" License
8.37k stars 1.49k forks source link

[feature request] C# / .NET bindings for in-proc C-API and in-proc wrapper's C++-API #7543

Open vadimkantorov opened 3 months ago

vadimkantorov commented 3 months ago

There exist relatively stable and supported tools for generating such bindings e.g. https://github.com/mono/CppSharp.

At least for the C-API it should run without problems similar to Javacpp generated bindings: https://github.com/bytedeco/javacpp-presets/blob/master/tritonserver/src/gen/java/org/bytedeco/tritonserver/global/tritonserver.java / https://github.com/bytedeco/javacpp-presets/tree/master/tritonserver/src/gen/java/org/bytedeco/tritonserver/tritonserver.

vadimkantorov commented 3 months ago

If anyone's searching, this seems to work to produce C-API bindings with CppSharp (https://github.com/vadimkantorov/tritonservercppsharp/blob/master/.github/workflows/tritonservercppsharp.yml and generated bindings at https://github.com/vadimkantorov/tritonservercppsharp/releases/tag/release)

So if anyone cares to try to actually test these generated bindings, maybe it would work.

name: tritonservercppsharp

on: workflow_dispatch

env:
  PLATFORM: x64
  FRAMEWORK: net80

jobs:
  tritonservercppsharp:
    runs-on: ubuntu-22.04
    steps:
      - name: Clone and build CppSharp
        run: |
          git clone --single-branch --depth 1 https://github.com/mono/CppSharp
          cd CppSharp
          bash build/build.sh generate -configuration Release -platform $PLATFORM -target-framework $FRAMEWORK
          bash build/build.sh download_llvm -platform $PLATFORM -target-framework $FRAMEWORK
          bash build/build.sh restore -platform $PLATFORM -target-framework $FRAMEWORK
          bash build/build.sh -platform $PLATFORM -build_only -target-framework $FRAMEWORK

      - name: Generating bindings with CppSharp
        run: |
          git clone --single-branch --depth 1 --branch r24.08 https://github.com/triton-inference-server/core
          ./CppSharp/bin/Release_x64/CppSharp.CLI -m tritonserver -g csharp -p linux -a x64 -o ./output/ -I=core/include core/include/triton/core/tritonserver.h core/include/triton/core/tritonbackend.h core/include/triton/core/tritoncache.h core/include/triton/core/tritonrepoagent.h

      - uses: actions/upload-artifact@v4
        with:
          path: output/

Generating C++ wrapper bindings seems to fail because CppSharp doesn't seem to support std::shared_ptr<T> yet: https://github.com/mono/CppSharp/blob/main/docs/UsersManual.md#standard-library-support. Maybe if shared_ptr is replaced by native pointer or value type, CppSharp would work here

vadimkantorov commented 2 months ago

In https://github.com/mono/CppSharp/issues/1860#issuecomment-2322838240 is claimed that using shared_ptr in C++ API are not ABI-portable. Not sure what it implies for tritonserver, but maybe if the C++ wrapper is considered for simplifying / binding, it would be better to model them without public usage of std::shared_ptr...

vadimkantorov commented 2 months ago

This seems to succeed:

name: tritonservercppsharp

on: workflow_dispatch

env:
  PLATFORM: x64
  FRAMEWORK: net80
  FRAMEWORKDOT: net8.0

jobs:
  tritonservercppsharp:
    runs-on: ubuntu-22.04
    steps:
      - name: Clone and build CppSharp
        run: |
          git clone --single-branch --depth 1 https://github.com/mono/CppSharp
          cd CppSharp
          bash build/build.sh generate -configuration Release -platform $PLATFORM -target-framework $FRAMEWORK
          bash build/build.sh download_llvm -platform $PLATFORM -target-framework $FRAMEWORK
          bash build/build.sh restore -platform $PLATFORM -target-framework $FRAMEWORK
          bash build/build.sh -platform $PLATFORM -build_only -target-framework $FRAMEWORK

      - name: Clone tritonserver
        run: |
          git clone --single-branch --depth 1 --branch r24.08 https://github.com/triton-inference-server/core
          git clone --single-branch --depth 1 --branch r24.08 https://github.com/triton-inference-server/developer_tools

      - name: Generate bindings oftritonserver
        run: |
          ./CppSharp/bin/Release_x64/CppSharp.CLI -m tritonserver -g csharp -p linux -a x64 -o ./variant1/ -I=core/include core/include/triton/core/tritonserver.h core/include/triton/core/tritonbackend.h core/include/triton/core/tritoncache.h core/include/triton/core/tritonrepoagent.h
          ./CppSharp/bin/Release_x64/CppSharp.CLI -m tritondevelopertoolsserver -g csharp -p linux -a x64 -o ./variant3/ -I=developer_tools/server/include -I=core/include -I=developer_tools/server/include/triton/developer_tools developer_tools/server/include/triton/developer_tools/server_wrapper.h developer_tools/server/include/triton/developer_tools/generic_server_wrapper.h