build: replace build.py with Conan 2 + CMakePresets build system

.gitignore

@@ -18,3 +18,4 @@ qa/L0_openai/openai
 tensorrtllm_models
 tensorrtllm_mistral_models/
 custom_tokenizer
+CMakeUserPresets.json

CMakePresets.json

@@ -0,0 +1,4 @@
+{
+  "version": 6,
+  "include": ["cmake/CMakePresets.json"]
+}

cmake/CMakePresets.json

@@ -0,0 +1,88 @@
+{
+  "version": 6,
+  "cmakeMinimumRequired": { "major": 3, "minor": 25, "patch": 0 },
+  "vendor": {
+    "note": "AUTHORITATIVE PRESET FILE. When building from server/ all downstream repos (core, common, backend) are included via add_subdirectory and inherit these variables. Downstream repos have their own CMakePresets.json for standalone builds only."
+  },
+  "configurePresets": [
+    {
+      "name": "conan-base",
+      "hidden": true,
+      "generator": "Ninja",
+      "binaryDir": "${sourceDir}/build/${presetName}",
+      "toolchainFile": "${sourceDir}/build/${presetName}/conan/generators/conan_toolchain.cmake",
+      "cacheVariables": { "CMAKE_EXPORT_COMPILE_COMMANDS": "ON" }
+    },
+    {
+      "name": "release",
+      "inherits": "conan-base",
+      "displayName": "Release (GPU + HTTP + gRPC + Metrics)",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE":      "Release",
+        "TRITON_ENABLE_GPU":     "ON",
+        "TRITON_ENABLE_HTTP":    "ON",
+        "TRITON_ENABLE_GRPC":    "ON",
+        "TRITON_ENABLE_METRICS": "ON"
+      }
+    },
+    {
+      "name": "debug",
+      "inherits": "conan-base",
+      "displayName": "Debug (GPU + HTTP + gRPC + Metrics)",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE":      "Debug",
+        "TRITON_ENABLE_GPU":     "ON",
+        "TRITON_ENABLE_HTTP":    "ON",
+        "TRITON_ENABLE_GRPC":    "ON",
+        "TRITON_ENABLE_METRICS": "ON"
+      }
+    },
+    {
+      "name": "cpu-only",
+      "inherits": "conan-base",
+      "displayName": "Release CPU-only",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE":          "Release",
+        "TRITON_ENABLE_GPU":         "OFF",
+        "TRITON_ENABLE_METRICS_GPU": "OFF",
+        "TRITON_ENABLE_HTTP":        "ON",
+        "TRITON_ENABLE_GRPC":        "ON",
+        "TRITON_ENABLE_METRICS":     "ON"
+      }
+    },
+    {
+      "name": "all-features",
+      "inherits": "conan-base",
+      "displayName": "Release - all features (no cloud storage)",
+      "cacheVariables": {
+        "CMAKE_BUILD_TYPE":                   "Release",
+        "TRITON_ENABLE_GPU":                  "ON",
+        "TRITON_ENABLE_HTTP":                 "ON",
+        "TRITON_ENABLE_GRPC":                 "ON",
+        "TRITON_ENABLE_METRICS":              "ON",
+        "TRITON_ENABLE_METRICS_GPU":          "ON",
+        "TRITON_ENABLE_METRICS_CPU":          "ON",
+        "TRITON_ENABLE_TRACING":              "ON",
+        "TRITON_ENABLE_ENSEMBLE":             "ON",
+        "TRITON_ENABLE_GCS":                  "OFF",
+        "TRITON_ENABLE_S3":                   "OFF",
+        "TRITON_ENABLE_AZURE_STORAGE":        "OFF",
+        "TRITON_ENABLE_IDENTITY_BACKEND":     "ON",
+        "TRITON_ENABLE_REPEAT_BACKEND":       "ON",
+        "TRITON_ENABLE_SQUARE_BACKEND":       "ON",
+        "TRITON_ENABLE_PYTHON_BACKEND":       "OFF",
+        "TRITON_ENABLE_ONNXRUNTIME_BACKEND":  "OFF",
+        "TRITON_ENABLE_PYTORCH_BACKEND":      "OFF",
+        "TRITON_ENABLE_TENSORRT_BACKEND":     "OFF",
+        "TRITON_ENABLE_VLLM_BACKEND":         "OFF",
+        "TRITON_ENABLE_TENSORRTLLM_BACKEND":  "OFF"
+      }
+    }
+  ],
+  "buildPresets": [
+    { "name": "release",      "configurePreset": "release" },
+    { "name": "debug",        "configurePreset": "debug" },
+    { "name": "cpu-only",     "configurePreset": "cpu-only" },
+    { "name": "all-features", "configurePreset": "all-features" }
+  ]
+}

cmake/TritonCheckCxxAbi.cmake

@@ -0,0 +1,110 @@
+# TritonCheckCxxAbi.cmake
+#
+# Detects the _GLIBCXX_USE_CXX11_ABI value that the current build will use and
+# validates that the Conan-provided binary packages were compiled with the same
+# ABI setting.  A mismatch causes silent link failures or runtime crashes when
+# mixing old ABI (=0) and new ABI (=1) translation units.
+#
+# Usage (call once, early in the root CMakeLists.txt after find_package calls):
+#   include(cmake/TritonCheckCxxAbi.cmake)
+#   triton_check_cxx_abi()
+#
+# The macro sets TRITON_GLIBCXX_USE_CXX11_ABI (cache variable) to 0 or 1.
+
+macro(triton_check_cxx_abi)
+    if(NOT CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+        return()
+    endif()
+
+    # ---------------------------------------------------------------
+    # 1. Probe the ABI the compiler will use for this build.
+    # ---------------------------------------------------------------
+    include(CheckCXXSourceCompiles)
+    set(_abi_probe_src "
+#include <string>
+#if _GLIBCXX_USE_CXX11_ABI
+int use_new_abi = 1;
+#else
+int use_new_abi = 0;
+#endif
+int main() { return use_new_abi; }
+")
+    # Ask the compiler directly via try_compile output rather than
+    # running the binary, so cross-compilation works too.
+    file(WRITE "${CMAKE_BINARY_DIR}/_abi_probe.cpp" "${_abi_probe_src}")
+    try_compile(
+        _abi_compile_ok
+        "${CMAKE_BINARY_DIR}/_abi_probe_build"
+        SOURCES "${CMAKE_BINARY_DIR}/_abi_probe.cpp"
+        CXX_STANDARD 17
+        OUTPUT_VARIABLE _abi_compile_output
+    )
+
+    if(NOT _abi_compile_ok)
+        message(WARNING "TritonCheckCxxAbi: could not compile ABI probe. Skipping ABI check.")
+        return()
+    endif()
+
+    # Extract the _GLIBCXX_USE_CXX11_ABI value by compiling a file that
+    # emits a preprocessor-expanded value we can grep.
+    file(WRITE "${CMAKE_BINARY_DIR}/_abi_value.cpp"
+        "#include <string>\nint abi_value = _GLIBCXX_USE_CXX11_ABI;\n")
+    execute_process(
+        COMMAND "${CMAKE_CXX_COMPILER}" -E -dM
+                "${CMAKE_BINARY_DIR}/_abi_value.cpp"
+        OUTPUT_VARIABLE _abi_macros
+        ERROR_QUIET
+    )
+
+    if(_abi_macros MATCHES "#define _GLIBCXX_USE_CXX11_ABI ([01])")
+        set(_detected_abi "${CMAKE_MATCH_1}")
+    else()
+        # Default: GCC >= 5 defaults to 1 (new ABI).
+        set(_detected_abi 1)
+    endif()
+
+    set(TRITON_GLIBCXX_USE_CXX11_ABI "${_detected_abi}"
+        CACHE STRING "Detected _GLIBCXX_USE_CXX11_ABI value (0=old, 1=new)" FORCE)
+    message(STATUS "TritonCheckCxxAbi: _GLIBCXX_USE_CXX11_ABI = ${TRITON_GLIBCXX_USE_CXX11_ABI}")
+
+    # ---------------------------------------------------------------
+    # 2. Validate Conan-provided binary packages that embed ABI info.
+    # ---------------------------------------------------------------
+    # Conan 2 records the compiler.libcxx setting in the package ID.
+    # If the Conan toolchain was generated we can cross-check via the
+    # conanbuildinfo vars; otherwise we skip silently.
+    if(DEFINED CONAN_COMPILER_LIBCXX)
+        if("${CONAN_COMPILER_LIBCXX}" STREQUAL "libstdc++" AND
+           "${TRITON_GLIBCXX_USE_CXX11_ABI}" STREQUAL "1")
+            message(FATAL_ERROR
+                "ABI MISMATCH: Build is using new C++11 ABI "
+                "(_GLIBCXX_USE_CXX11_ABI=1) but Conan packages were built "
+                "with libstdc++ (old ABI). Re-run 'conan install' with "
+                "compiler.libcxx=libstdc++11 in the host profile.")
+        endif()
+        if("${CONAN_COMPILER_LIBCXX}" STREQUAL "libstdc++11" AND
+           "${TRITON_GLIBCXX_USE_CXX11_ABI}" STREQUAL "0")
+            message(FATAL_ERROR
+                "ABI MISMATCH: Build is using old ABI "
+                "(_GLIBCXX_USE_CXX11_ABI=0) but Conan packages were built "
+                "with libstdc++11 (new ABI). Re-run 'conan install' with "
+                "compiler.libcxx=libstdc++ in the host profile, or remove "
+                "-D_GLIBCXX_USE_CXX11_ABI=0 from your compile flags.")
+        endif()
+    endif()
+
+    # ---------------------------------------------------------------
+    # 3. Warn if a user has manually set a conflicting compile flag.
+    # ---------------------------------------------------------------
+    if(DEFINED CMAKE_CXX_FLAGS AND
+       CMAKE_CXX_FLAGS MATCHES "_GLIBCXX_USE_CXX11_ABI=([01])")
+        set(_flag_abi "${CMAKE_MATCH_1}")
+        if(NOT "${_flag_abi}" STREQUAL "${_detected_abi}")
+            message(WARNING
+                "TritonCheckCxxAbi: CMAKE_CXX_FLAGS sets "
+                "_GLIBCXX_USE_CXX11_ABI=${_flag_abi} but the compiler "
+                "default is ${_detected_abi}. This will likely cause "
+                "linker errors with Conan-provided binary packages.")
+        endif()
+    endif()
+endmacro()

cmake/presets/CMakePresets.TritonClient.Release.CPU.ubuntu.aarch64.json

@@ -0,0 +1,64 @@
+{
+  "version": 6,
+  "vendor": {
+    "triton": {
+      "component": "TritonClient",
+      "platform": "Ubuntu aarch64, CPU-only (Graviton, Neoverse, Apple Silicon dev)",
+      "notes": [
+        "CPU-only client build for aarch64 — no CUDA required.",
+        "Targets AWS Graviton, Ampere Altra, and NVIDIA Grace (CPU die).",
+        "Run cmake from the client/ source directory."
+      ],
+      "conan_install": [
+        "conan install client/",
+        "  --profile:host=server/conan/profiles/linux-gcc13-release-aarch64",
+        "  --profile:build=server/conan/profiles/linux-gcc13-release",
+        "  -o '&:enable_gpu=False'",
+        "  -o '&:enable_grpc=True'",
+        "  -o '&:enable_http=True'",
+        "  --build=missing",
+        "  --output-folder=client/build/triton-client-release-cpu-ubuntu-aarch64/conan",
+        "",
+        "cmake --preset triton-client-release-cpu-ubuntu-aarch64",
+        "cmake --build --preset triton-client-release-cpu-ubuntu-aarch64",
+        "cmake --build --preset triton-client-release-cpu-ubuntu-aarch64 --target wheels"
+      ]
+    }
+  },
+  "configurePresets": [
+    {
+      "name": "triton-client-release-cpu-ubuntu-aarch64",
+      "displayName": "TritonClient — Release, CPU-only, Ubuntu, aarch64",
+      "description": "CPU-only client library build for Ubuntu Linux on aarch64.",
+      "generator": "Ninja",
+      "binaryDir": "${sourceDir}/build/${presetName}",
+      "toolchainFile": "${sourceDir}/build/${presetName}/conan/generators/conan_toolchain.cmake",
+      "cacheVariables": {
+        "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
+        "CMAKE_BUILD_TYPE":               "Release",
+        "TRITON_ENABLE_GPU":              "OFF",
+        "TRITON_ENABLE_CC_HTTP":          "ON",
+        "TRITON_ENABLE_CC_GRPC":          "ON",
+        "TRITON_ENABLE_PYTHON_HTTP":      "ON",
+        "TRITON_ENABLE_PYTHON_GRPC":      "ON",
+        "TRITON_ENABLE_PERF_ANALYZER":    "ON",
+        "TRITON_ENABLE_EXAMPLES":         "ON",
+        "TRITON_ENABLE_TESTS":            "ON"
+      }
+    }
+  ],
+  "buildPresets": [
+    {
+      "name": "triton-client-release-cpu-ubuntu-aarch64",
+      "configurePreset": "triton-client-release-cpu-ubuntu-aarch64",
+      "jobs": 8
+    }
+  ],
+  "testPresets": [
+    {
+      "name": "triton-client-release-cpu-ubuntu-aarch64",
+      "configurePreset": "triton-client-release-cpu-ubuntu-aarch64",
+      "filter": { "exclude": { "label": "requires-gpu" } }
+    }
+  ]
+}

cmake/presets/CMakePresets.TritonClient.Release.CPU.ubuntu.x86_64.json

@@ -0,0 +1,64 @@
+{
+  "version": 6,
+  "vendor": {
+    "triton": {
+      "component": "TritonClient",
+      "platform": "Ubuntu x86_64, CPU-only (no NVIDIA GPU required)",
+      "notes": [
+        "CPU-only client build — no CUDA toolkit required.",
+        "CUDA shared memory operations are disabled.",
+        "Run cmake from the client/ source directory."
+      ],
+      "conan_install": [
+        "conan install client/",
+        "  --profile:host=server/conan/profiles/linux-gcc13-release",
+        "  --profile:build=server/conan/profiles/linux-gcc13-release",
+        "  -o '&:enable_gpu=False'",
+        "  -o '&:enable_grpc=True'",
+        "  -o '&:enable_http=True'",
+        "  --build=missing",
+        "  --output-folder=client/build/triton-client-release-cpu-ubuntu-x86_64/conan",
+        "",
+        "cmake --preset triton-client-release-cpu-ubuntu-x86_64",
+        "cmake --build --preset triton-client-release-cpu-ubuntu-x86_64",
+        "cmake --build --preset triton-client-release-cpu-ubuntu-x86_64 --target wheels"
+      ]
+    }
+  },
+  "configurePresets": [
+    {
+      "name": "triton-client-release-cpu-ubuntu-x86_64",
+      "displayName": "TritonClient — Release, CPU-only, Ubuntu, x86_64",
+      "description": "CPU-only client library build for Ubuntu Linux on x86_64. No CUDA required.",
+      "generator": "Ninja",
+      "binaryDir": "${sourceDir}/build/${presetName}",
+      "toolchainFile": "${sourceDir}/build/${presetName}/conan/generators/conan_toolchain.cmake",
+      "cacheVariables": {
+        "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
+        "CMAKE_BUILD_TYPE":               "Release",
+        "TRITON_ENABLE_GPU":              "OFF",
+        "TRITON_ENABLE_CC_HTTP":          "ON",
+        "TRITON_ENABLE_CC_GRPC":          "ON",
+        "TRITON_ENABLE_PYTHON_HTTP":      "ON",
+        "TRITON_ENABLE_PYTHON_GRPC":      "ON",
+        "TRITON_ENABLE_PERF_ANALYZER":    "ON",
+        "TRITON_ENABLE_EXAMPLES":         "ON",
+        "TRITON_ENABLE_TESTS":            "ON"
+      }
+    }
+  ],
+  "buildPresets": [
+    {
+      "name": "triton-client-release-cpu-ubuntu-x86_64",
+      "configurePreset": "triton-client-release-cpu-ubuntu-x86_64",
+      "jobs": 8
+    }
+  ],
+  "testPresets": [
+    {
+      "name": "triton-client-release-cpu-ubuntu-x86_64",
+      "configurePreset": "triton-client-release-cpu-ubuntu-x86_64",
+      "filter": { "exclude": { "label": "requires-gpu" } }
+    }
+  ]
+}