Fix pip install test workflows (#3542)

Summary:
X-link: facebookresearch/FBGEMM#630

- Fix pip install test workflows

Pull Request resolved: #3542

Reviewed By: sryap

Differential Revision: D67817349

Pulled By: q10

fbshipit-source-id: 27ffce0b43666df34c8d6c1bf86717abeaf07366

.github/scripts/fbgemm_gpu_install.bash

@@ -180,8 +180,8 @@ install_fbgemm_gpu_pip () {
     echo "Usage: ${FUNCNAME[0]} ENV_NAME FBGEMM_GPU_CHANNEL[/VERSION] FBGEMM_GPU_VARIANT_TYPE[/VARIANT_VERSION]"
     echo "Example(s):"
     echo "    ${FUNCNAME[0]} build_env 0.8.0 cpu                  # Install the CPU variant, specific version from release channel"
-    echo "    ${FUNCNAME[0]} build_env release cuda/12.4.1        # Install the CUDA variant, latest version from release channel"
-    echo "    ${FUNCNAME[0]} build_env test/0.8.0 cuda/12.4.1     # Install the CUDA 12.4 variant, specific version from test channel"
+    echo "    ${FUNCNAME[0]} build_env release cuda/12.6.3        # Install the CUDA 12.3 variant, latest version from release channel"
+    echo "    ${FUNCNAME[0]} build_env test/0.8.0 cuda/12.6.3     # Install the CUDA 12.3 variant, specific version from test channel"
     echo "    ${FUNCNAME[0]} build_env nightly rocm/6.2           # Install the ROCM 6.2 variant, latest version from nightly channel"
     return 1
   else

.github/scripts/fbgemm_gpu_test.bash

@@ -336,10 +336,10 @@ test_setup_conda_environment () {
   local pytorch_channel_version="$5"
   local pytorch_variant_type="$6"
   local pytorch_variant_version="$7"
-  if [ "$pytorch_variant_type" == "" ]; then
-    echo "Usage: ${FUNCNAME[0]} ENV_NAME COMPILER PYTHON_VERSION PYTORCH_INSTALLER PYTORCH_CHANNEL[/VERSION] PYTORCH_VARIANT_TYPE [PYTORCH_VARIANT_VERSION]"
+  if [ "$pytorch_variant_version" == "" ]; then
+    echo "Usage: ${FUNCNAME[0]} ENV_NAME COMPILER PYTHON_VERSION PYTORCH_INSTALLER PYTORCH_CHANNEL[/VERSION] PYTORCH_VARIANT_TYPE PYTORCH_VARIANT_VERSION"
     echo "Example(s):"
-    echo "    ${FUNCNAME[0]} build_env clang 3.13 pip test/1.0.0 cuda 12.4.1       # Setup environment with pytorch-test 1.0.0 for Clang + Python 3.13 + CUDA 12.4.1"
+    echo "    ${FUNCNAME[0]} build_env clang 3.13 pip test/1.0.0 cuda 12.6.3       # Setup environment with pytorch-test 1.0.0 for Clang + Python 3.13 + CUDA 12.6.3"
     return 1
   else
     echo "################################################################################"
@@ -350,16 +350,13 @@ test_setup_conda_environment () {
     echo ""
   fi
 
-  if [ "$env_name" == "" ]; then
-    local env_name="test_py${python_version}_${pytorch_installer}_pytorch_${pytorch_channel_version}_${pytorch_variant_type}"
-    if [ "$pytorch_variant_version" != "" ]; then
-      local env_name="${env_name}_${pytorch_variant_version}"
-    fi
-  fi
-
   echo "Creating the Build Environment: ${env_name} ..."
   create_conda_environment  "${env_name}" "${python_version}"           || return 1
 
+  if [ "$pytorch_variant_type" == "cuda" ] || [ "$pytorch_variant_type" == "genai" ]; then
+    print_exec conda env config vars set -n "${env_name}" BUILD_CUDA_VERSION="${pytorch_variant_version}"
+  fi
+
   # Install C++ compiler and build tools (all FBGEMM_GPU variants)
   if [ "$compiler" == "gcc" ] || [ "$compiler" == "clang" ]; then
     install_cxx_compiler  "${env_name}" "${compiler}"   || return 1
@@ -503,6 +500,7 @@ test_fbgemm_gpu_setup_and_pip_install () {
     local variant_versions=(
       11.8.0
       12.4.1
+      12.6.3
     )
   elif [ "$variant_type" == "rocm" ]; then
     local variant_versions=(

.github/scripts/test_torchrec.bash

@@ -84,7 +84,7 @@ setup_miniconda "$miniconda_prefix"
 echo "## 2. Create Conda environment"
 ################################################################################
 
-if [ "${cuda_version}" == "" ]; then
+if [ "${cuda_version}" != "" ]; then
   pytorch_variant="cuda ${cuda_version}"
 else
   pytorch_variant="cpu"

.github/scripts/utils_build.bash

@@ -357,8 +357,8 @@ install_build_tools () {
   local conda_prefix=$(conda run ${env_prefix} printenv CONDA_PREFIX)
   (print_exec ln -s "${conda_prefix}/lib/librhash.so" "${conda_prefix}/lib/librhash.so.0") || return 1
 
-  # For some reason, the build package for Python 3.12 is missing from Conda, so
-  # we have to install through PyPI instead.
+  # For some reason, the build package for Python 3.12+ is missing from conda,
+  # so we have to install through pip instead.
   #
   # LibMambaUnsatisfiableError: Encountered problems while solving:
   #   - package build-0.10.0-py310h06a4308_0 requires python >=3.10,<3.11.0a0, but none of the providers can be installed

.github/workflows/fbgemm_gpu_docs.yml

@@ -42,7 +42,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: [ "3.12" ]
+        python-version: [ "3.13" ]
 
     steps:
     - name: Setup Build Container

.github/workflows/fbgemm_gpu_pip.yml

@@ -26,12 +26,12 @@ on:
   workflow_dispatch:
     inputs:
       pytorch_version:
-        description: PyTorch Version (e.g. '2.1.0', 'nightly', 'test')
+        description: PyTorch Version (e.g. '2.6.0', 'nightly', 'test')
         type: string
         required: true
         default: "nightly"
       fbgemm_gpu_channel_version:
-        description: FBGEMM-GPU Channel + Version (e.g. '0.5.0', 'nightly', 'test/0.8.0r0')
+        description: FBGEMM-GPU Channel + Version (e.g. '1.1.0', 'nightly', 'test/1.0.0r1')
         type: string
         required: true
         default: "nightly"
@@ -85,6 +85,9 @@ jobs:
     - name: Create Conda Environment
       run: . $PRELUDE; create_conda_environment $BUILD_ENV ${{ matrix.python-version }}
 
+    - name: Install C/C++ Compilers for Updated LIBGCC
+      run: . $PRELUDE; install_cxx_compiler $BUILD_ENV ${{ matrix.compiler }}
+
     - name: Install Build Tools
       run: . $PRELUDE; install_build_tools $BUILD_ENV
 
@@ -113,6 +116,7 @@ jobs:
       PRELUDE: .github/scripts/setup_env.bash
       BUILD_ENV: test_install
       BUILD_VARIANT: cuda
+      BUILD_CUDA_VERSION: ${{ matrix.cuda-version }}
       ENFORCE_CUDA_DEVICE: 1
     strategy:
       fail-fast: false
@@ -143,6 +147,13 @@ jobs:
     - name: Create Conda Environment
       run: . $PRELUDE; create_conda_environment $BUILD_ENV ${{ matrix.python-version }}
 
+    - name: Install C/C++ Compilers for Updated LIBGCC
+      # NOTE: gcc is required for torch dynamo to work properly, as some of
+      # the compilation flags used by torch dynamo are gcc-specific:
+      #
+      #   clang-16: error: unknown argument: '-fno-tree-loop-vectorize'
+      run: . $PRELUDE; install_cxx_compiler $BUILD_ENV gcc
+
     - name: Install Build Tools
       run: . $PRELUDE; install_build_tools $BUILD_ENV
 
@@ -213,6 +224,9 @@ jobs:
     - name: Create Conda Environment
       run: . $PRELUDE; create_conda_environment $BUILD_ENV ${{ matrix.python-version }}
 
+    - name: Install ROCm AMD-SMI
+      run: . $PRELUDE; install_rocm_amdsmi_ubuntu $BUILD_ENV
+
     - name: Install Build Tools
       run: . $PRELUDE; install_build_tools $BUILD_ENV
 

cmake/modules/CxxCompilerSetup.cmake

@@ -43,14 +43,6 @@ endif()
 set(CMAKE_C_EXTENSIONS OFF)
 set(CMAKE_C_STANDARD_REQUIRED ON)
 
-if(DEFINED GLIBCXX_USE_CXX11_ABI)
-  if(${GLIBCXX_USE_CXX11_ABI} EQUAL 1)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D_GLIBCXX_USE_CXX11_ABI=1")
-  else()
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D_GLIBCXX_USE_CXX11_ABI=0")
-  endif()
-endif()
-
 BLOCK_PRINT(
   "Default C compiler flags"
   "(values may be overridden by CMAKE_CXX_STANDARD and CXX_STANDARD):"

fbgemm_gpu/docs/src/fbgemm_gpu-development/InstallationInstructions.rst

@@ -25,8 +25,8 @@ older than the one that the FBGEMM release corresponds to.
 | FBGEMM Release  | Corresponding    | Supported        | Supported      | (Experimental) Supported  |
 |                 | PyTorch Release  | Python Versions  | CUDA Versions  | ROCm Versions             |
 +=================+==================+==================+================+===========================+
-| 1.1.0           | 2.6.x            | 3.9, 3.10, 3.11, | 11.8, 12.4     | 6.0, 6.1                  |
-|                 |                  | 3.12, 3.13       |                |                           |
+| 1.1.0           | 2.6.x            | 3.9, 3.10, 3.11, | 11.8, 12.4,    | 6.0, 6.1                  |
+|                 |                  | 3.12, 3.13       | 12.6           |                           |
 +-----------------+------------------+------------------+----------------+---------------------------+
 | 1.0.0           | 2.5.x            | 3.9, 3.10, 3.11, | 11.8, 12.1,    | 6.0, 6.1                  |
 |                 |                  | 3.12             | 12.4           |                           |

fbgemm_gpu/setup.py

@@ -237,7 +237,9 @@ def _get_cxx11_abi():
                 value = int(torch._C._GLIBCXX_USE_CXX11_ABI)
             except ImportError:
                 value = 0
-            return "-DGLIBCXX_USE_CXX11_ABI=" + str(value)
+            # NOTE: The correct spelling for the flag is
+            # `_GLIBCXX_USE_CXX11_ABI`, not `GLIBCXX_USE_CXX11_ABI`
+            return f"-D_GLIBCXX_USE_CXX11_ABI={value}"
 
         torch_root = os.path.dirname(torch.__file__)
         os.environ["CMAKE_BUILD_PARALLEL_LEVEL"] = str(os.cpu_count() // 2)

netlify.toml

@@ -18,7 +18,7 @@
 
     # Set up Conda environment
     setup_miniconda             $HOME/miniconda
-    create_conda_environment    $BUILD_ENV 3.12
+    create_conda_environment    $BUILD_ENV 3.13
 
     # Install tools
     install_cxx_compiler        $BUILD_ENV