[T145005253] Check for presence of NVIDIA drivers in OSS builds

- Check for the actual presence of NVIDIA drivers in OSS builds, and error out with friendly message instead of cryptic `RuntimeError: No such operator fbgemm::jagged_2d_to_dense` errors when `fbgemm_gpu` is installed and loaded on a system with a GPU but without GPU drivers installed - Allow for testing of multiple CUDA versions of the build, but restrict artifact publishing to just one CUDA version
pytorch · Mar 7, 2023 · e0d0be8 · e0d0be8
1 parent f4e0b70
commit e0d0be8
Show file tree

Hide file tree

Showing 7 changed files with 86 additions and 37 deletions.
diff --git a/.github/scripts/setup_env.bash b/.github/scripts/setup_env.bash
@@ -14,6 +14,7 @@ print_exec () {
   echo "+ $*"
   echo ""
   "$@"
+  echo ""
 }
 
 exec_with_retries () {
@@ -238,6 +239,30 @@ free_disk_space () {
 # Info Functions
 ################################################################################
 
+print_gpu_info () {
+  echo "################################################################################"
+  echo "[INFO] Check GPU info ..."
+  install_system_packages lshw
+  print_exec sudo lshw -C display
+
+  echo "################################################################################"
+  echo "[INFO] Check NVIDIA GPU info ..."
+
+  if [[ "${ENFORCE_NVIDIA_GPU}" ]]; then
+    # Ensure that nvidia-smi is available and returns GPU entries
+    if ! nvidia-smi; then
+      echo "[CHECK] NVIDIA driver is required, but does not appear to have been installed.  This will cause FBGEMM_GPU installation to fail!"
+      return 1
+    fi
+
+  else
+    if which nvidia-smi; then
+      # If nvidia-smi is installed on a machine without GPUs, this will return error
+      (print_exec nvidia-smi) || true
+    fi
+  fi
+}
+
 print_system_info () {
   echo "################################################################################"
   echo "# Print System Info"
@@ -264,18 +289,6 @@ print_system_info () {
   print_exec uname -a
   print_exec cat /proc/version
   print_exec cat /etc/os-release
-
-  echo "################################################################################"
-  echo "[INFO] Check GPU info ..."
-  install_system_packages lshw
-  print_exec sudo lshw -C display
-
-  if which nvidia-smi; then
-    echo "################################################################################"
-    echo "[INFO] Check NVIDIA GPU info ..."
-    # If nvidia-smi is installed on a machine without GPUs, this will return error
-    (print_exec nvidia-smi) || true
-  fi
 }
 
 print_ec2_info () {
@@ -336,7 +349,7 @@ setup_miniconda () {
   print_exec . ~/.bashrc
 
   echo "[SETUP] Updating Miniconda base packages ..."
-  print_exec conda update -n base -c defaults -y conda
+  (exec_with_retries conda update -n base -c defaults -y conda) || return 1
 
   # Print Conda info
   print_exec conda info

diff --git a/.github/workflows/fbgemm_gpu_ci.yml b/.github/workflows/fbgemm_gpu_ci.yml
@@ -38,6 +38,9 @@ jobs:
     - name: Display System Info
       run: . $PRELUDE; print_system_info
 
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
+
     - name: Free Disk Space
       run: . $PRELUDE; free_disk_space
 
@@ -150,6 +153,9 @@ jobs:
     - name: Display System Info
       run: . $PRELUDE; print_system_info
 
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
+
     - name: Setup Miniconda
       run: |
         . $PRELUDE; setup_miniconda $HOME/miniconda

diff --git a/.github/workflows/fbgemm_nightly_build.yml b/.github/workflows/fbgemm_nightly_build.yml
@@ -57,6 +57,9 @@ jobs:
     - name: Display System Info
       run: . $PRELUDE; print_system_info
 
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
+
     - name: Setup Miniconda
       run: |
         . $PRELUDE; setup_miniconda $HOME/miniconda
@@ -103,12 +106,15 @@ jobs:
     env:
       PRELUDE: .github/scripts/setup_env.bash
       BUILD_ENV: build_binary
+      ENFORCE_NVIDIA_GPU: 1
     strategy:
       fail-fast: false
       matrix:
         os: [ linux.g5.4xlarge.nvidia.gpu ]
         python-version: [ "3.8", "3.9", "3.10" ]
         cuda-version: [ "11.7.1", "11.8.0" ]
+        # Specify exactly ONE CUDA version for artifact publish
+        cuda-version-publish: [ "11.7.1" ]
     needs: build_artifact
 
     steps:
@@ -118,10 +124,10 @@ jobs:
         submodules: true
 
     - name: Display System Info
-      run: . $PRELUDE; print_system_info
+      run: . $PRELUDE; print_system_info; print_ec2_info
 
-    - name: Display EC2 Info
-      run: . $PRELUDE; print_ec2_info
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
 
     - name: Setup Miniconda
       run: |
@@ -157,7 +163,12 @@ jobs:
       run: . $PRELUDE; cd fbgemm_gpu/test; run_fbgemm_gpu_tests $BUILD_ENV
 
     - name: Push FBGEMM_GPU Nightly Binary to PYPI
-      if: ${{ github.event_name != 'pull_request' && github.event_name != 'push' }}
+      if: |
+        ${{
+          github.event_name != 'pull_request' &&
+          github.event_name != 'push' &&
+          matrix.cuda-version == matrix.cuda-version-publish
+        }}
       env:
         PYPI_TOKEN: ${{ secrets.PYPI_TOKEN }}
       run: . $PRELUDE; publish_to_pypi $BUILD_ENV fbgemm_gpu_nightly-*.whl "$PYPI_TOKEN"
diff --git a/.github/workflows/fbgemm_nightly_build_cpu.yml b/.github/workflows/fbgemm_nightly_build_cpu.yml
@@ -56,6 +56,9 @@ jobs:
     - name: Display System Info
       run: . $PRELUDE; print_system_info
 
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
+
     - name: Setup Miniconda
       run: |
         . $PRELUDE; setup_miniconda $HOME/miniconda
@@ -110,10 +113,10 @@ jobs:
         submodules: true
 
     - name: Display System Info
-      run: . $PRELUDE; print_system_info
+      run: . $PRELUDE; print_system_info; print_ec2_info
 
-    - name: Display EC2 Info
-      run: . $PRELUDE; print_ec2_info
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
 
     - name: Setup Miniconda
       run: |

diff --git a/.github/workflows/fbgemm_release_build.yml b/.github/workflows/fbgemm_release_build.yml
@@ -38,7 +38,7 @@ jobs:
       matrix:
         os: [ linux.12xlarge ]
         python-version: [ "3.8", "3.9", "3.10" ]
-        cuda-version: [ "11.7.1" ]
+        cuda-version: [ "11.7.1", "11.8.0" ]
 
     steps:
     - name: Checkout the Repository
@@ -49,6 +49,9 @@ jobs:
     - name: Display System Info
       run: . $PRELUDE; print_system_info
 
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
+
     - name: Setup Miniconda
       run: |
         . $PRELUDE; setup_miniconda $HOME/miniconda
@@ -95,12 +98,15 @@ jobs:
     env:
       PRELUDE: .github/scripts/setup_env.bash
       BUILD_ENV: build_binary
+      ENFORCE_NVIDIA_GPU: 1
     strategy:
       fail-fast: false
       matrix:
         os: [ linux.g5.4xlarge.nvidia.gpu ]
         python-version: [ "3.8", "3.9", "3.10" ]
-        cuda-version: [ "11.7.1" ]
+        cuda-version: [ "11.7.1", "11.8.0" ]
+        # Specify exactly ONE CUDA version for artifact publish
+        cuda-version-publish: [ "11.7.1" ]
     needs: build_artifact
     steps:
     - name: Checkout the Repository
@@ -109,10 +115,10 @@ jobs:
         submodules: true
 
     - name: Display System Info
-      run: . $PRELUDE; print_system_info
+      run: . $PRELUDE; print_system_info; print_ec2_info
 
-    - name: Display EC2 Info
-      run: . $PRELUDE; print_ec2_info
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
 
     - name: Setup Miniconda
       run: |
@@ -148,7 +154,12 @@ jobs:
       run: . $PRELUDE; cd fbgemm_gpu/test; run_fbgemm_gpu_tests $BUILD_ENV
 
     - name: Push FBGEMM_GPU Binary to PYPI
-      if: ${{ github.event_name != 'pull_request' && github.event_name != 'push' }}
+      if: |
+        ${{
+          github.event_name != 'pull_request' &&
+          github.event_name != 'push' &&
+          matrix.cuda-version == matrix.cuda-version-publish
+        }}
       env:
         PYPI_TOKEN: ${{ secrets.PYPI_TOKEN }}
       run: . $PRELUDE; publish_to_pypi $BUILD_ENV fbgemm_gpu-*.whl "$PYPI_TOKEN"
diff --git a/.github/workflows/fbgemm_release_build_cpu.yml b/.github/workflows/fbgemm_release_build_cpu.yml
@@ -48,6 +48,9 @@ jobs:
     - name: Display System Info
       run: . $PRELUDE; print_system_info
 
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
+
     - name: Setup Miniconda
       run: |
         . $PRELUDE; setup_miniconda $HOME/miniconda
@@ -102,10 +105,10 @@ jobs:
         submodules: true
 
     - name: Display System Info
-      run: . $PRELUDE; print_system_info
+      run: . $PRELUDE; print_system_info; print_ec2_info
 
-    - name: Display EC2 Info
-      run: . $PRELUDE; print_ec2_info
+    - name: Display GPU Info
+      run: . $PRELUDE; print_gpu_info
 
     - name: Setup Miniconda
       run: |

diff --git a/fbgemm_gpu/setup.py b/fbgemm_gpu/setup.py
@@ -7,34 +7,36 @@
 import argparse
 import os
 import random
-import re
 import subprocess
 import sys
 import setuptools_git_versioning as gitversion
 import torch
+
 from datetime import date
 from skbuild import setup
 from typing import List, Optional
 
 
 def generate_package_version(package_name: str):
-    print(f"[SETUP.PY] Generating the package version ...")
+    print("[SETUP.PY] Generating the package version ...")
 
-    if 'nightly' in package_name:
+    if "nightly" in package_name:
         # Use date stamp for nightly versions
-        print(f"[SETUP.PY] Package is for NIGHTLY; using timestamp for the versioning")
+        print("[SETUP.PY] Package is for NIGHTLY; using timestamp for the versioning")
         today = date.today()
         version = f"{today.year}.{today.month}.{today.day}"
 
-    elif 'test' in package_name:
+    elif "test" in package_name:
         # Use date stamp for nightly versions
-        print(f"[SETUP.PY] Package is for TEST: using random number for the versioning")
+        print("[SETUP.PY] Package is for TEST: using random number for the versioning")
         version = (f"0.0.{random.randint(0, 1000)}",)
 
     else:
         # Use git tag / branch / commit info to generate a PEP-440-compliant version string
-        print(f"[SETUP.PY] Package is for RELEASE: using git info for the versioning")
-        print(f"[SETUP.PY] TAG: {gitversion.get_tag()}, BRANCH: {gitversion.get_branch()}, SHA: {gitversion.get_sha()}")
+        print("[SETUP.PY] Package is for RELEASE: using git info for the versioning")
+        print(
+            f"[SETUP.PY] TAG: {gitversion.get_tag()}, BRANCH: {gitversion.get_branch()}, SHA: {gitversion.get_sha()}"
+        )
         version = gitversion.version_from_git()
 
     print(f"[SETUP.PY] Setting the package version: {version}")