Pre-build Legion library (#1042)

* add optional flag for building legion only * added build path and legion-only flag * bug fix * pass new variable with config file * move nccl * bug fix * add cuda_arch list * export position move * cd into legion * quick fix * retrieve os version and cd directory * using ubuntu * directory fix * bug fix * add touch * create the release to flexflow-third-party * bug fix * bug fix * fix indentation * fix * bash launching * bug fix * bug fix * extract tar file` * bug fix * add parameter * bash fix * python version * bug fix * bug fix * bug fix * bug fix * bug fix * build bash * bug fix * bug fix * bug fix * bug fix * bug fix * auto running docker container * renew bash script * bug fix * bug fix * bug fix * non-running container * bug fix * make it easier to switch between inference and master branch * multiple fixes * bug fix * bug fix * add python version * bug fix * restore * enable building docker images for different hip versions * ignore shellcheck error code * support hip compilation in inference cmake files * fix * workflow and hardcode * bug fix * fix * cmake fix * python versions * cmake fixes * cmake fixes * move install * order * bug fix * nested if condition fix * update docker workflow and config scripts * update scripts * fix * fix * cleanup * rocm 5.6 by default in workflow * move outside * update workflow * incorp install.sh * bug fix * fix * fix * fix * bg fix * fix permissions * bug fix * bug fix * bug fix * bug fix * updated * bug fix * fix workflow * check * check * bug fix * fix * add python env * fix * cleanup * update workflow * newline * added runner * added endif * Code Cleanup * restore to self-hosted * bug fix * fix * fix * update workflow * fixes * fix cmake for hip rocm --------- Co-authored-by: Gabriele Oliaro <goliaro@cs.cmu.edu>
flexflow · Oct 23, 2023 · dd9f62d · dd9f62d
1 parent caf5d61
commit dd9f62d
Show file tree

Hide file tree

Showing 9 changed files with 452 additions and 242 deletions.
diff --git a/.github/workflows/helpers/prebuild_legion.sh b/.github/workflows/helpers/prebuild_legion.sh
@@ -0,0 +1,75 @@
+#! /usr/bin/env bash
+set -euo pipefail
+
+# Parse input params
+python_version=${python_version:-"empty"}
+gpu_backend=${gpu_backend:-"empty"}
+gpu_backend_version=${gpu_backend_version:-"empty"}
+
+if [[ "${gpu_backend}" != @(cuda|hip_cuda|hip_rocm|intel) ]]; then
+  echo "Error, value of gpu_backend (${gpu_backend}) is invalid. Pick between 'cuda', 'hip_cuda', 'hip_rocm' or 'intel'."
+  exit 1
+else
+  echo "Pre-building Legion with GPU backend: ${gpu_backend}"
+fi
+
+if [[ "${gpu_backend}" == "cuda" || "${FF_GPU_BACKEND}" == "hip_cuda" ]]; then
+    # Check that CUDA version is supported. Versions above 12.0 not supported because we don't publish docker images for it yet.
+    if [[ "$gpu_backend_version" != @(11.1|11.2|11.3|11.4|11.5|11.6|11.7|11.8|12.0) ]]; then
+        echo "cuda_version is not supported, please choose among {11.1|11.2|11.3|11.4|11.5|11.6|11.7|11.8|12.0}"
+        exit 1
+    fi
+    export cuda_version="$gpu_backend_version"
+elif [[ "${gpu_backend}" == "hip_rocm" ]]; then
+    # Check that HIP version is supported
+    if [[ "$gpu_backend_version" != @(5.3|5.4|5.5|5.6) ]]; then
+        echo "hip_version is not supported, please choose among {5.3, 5.4, 5.5, 5.6}"
+        exit 1
+    fi
+    export hip_version="$gpu_backend_version"
+else
+    echo "gpu backend: ${gpu_backend} and gpu_backend_version: ${gpu_backend_version} not yet supported."
+    exit 1
+fi
+
+# Cd into directory holding this script
+cd "${BASH_SOURCE[0]%/*}"
+
+export FF_GPU_BACKEND="${gpu_backend}"
+export FF_CUDA_ARCH=all
+export FF_HIP_ARCH=all
+export BUILD_LEGION_ONLY=ON
+export INSTALL_DIR="/usr/legion"
+export python_version="${python_version}"
+
+# Build Docker Flexflow Container
+echo "building docker"
+../../../docker/build.sh flexflow
+
+# Cleanup any existing container with the same name
+docker rm prelegion || true
+
+# Create container to be able to copy data from the image
+docker create --name prelegion flexflow-"${gpu_backend}"-"${gpu_backend_version}":latest
+
+# Copy legion libraries to host
+echo "extract legion library assets"
+mkdir -p ../../../prebuilt_legion_assets
+rm -rf ../../../prebuilt_legion_assets/tmp || true
+docker cp prelegion:$INSTALL_DIR ../../../prebuilt_legion_assets/tmp
+
+
+# Create the tarball file
+cd ../../../prebuilt_legion_assets/tmp
+export LEGION_TARBALL="legion_ubuntu-20.04_${gpu_backend}-${gpu_backend_version}_py${python_version}.tar.gz"
+
+echo "Creating archive $LEGION_TARBALL"
+tar -zcvf "../$LEGION_TARBALL" ./
+cd ..
+echo "Checking the size of the Legion tarball..."
+du -h "$LEGION_TARBALL"
+
+
+# Cleanup
+rm -rf tmp/*
+docker rm prelegion
diff --git a/.github/workflows/prebuild-legion.yml b/.github/workflows/prebuild-legion.yml
@@ -0,0 +1,84 @@
+name: "prebuild-legion"
+on:
+  push:
+    branches:
+      - "inference"
+    paths:
+      - "cmake/**"
+      - "config/**"
+      - "deps/legion/**"
+      - ".github/workflows/helpers/install_dependencies.sh"
+  workflow_dispatch:
+concurrency:
+  group: prebuild-legion-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  prebuild-legion:
+    name: Prebuild Legion with CMake
+    runs-on: ubuntu-20.04
+    defaults:
+      run:
+        shell: bash -l {0} # required to use an activated conda environment
+    strategy:
+      matrix:
+        gpu_backend: ["cuda", "hip_rocm"]
+        gpu_backend_version: ["11.8", "5.6"]
+        python_version: "3.11"
+        exclude:
+          - gpu_backend: "cuda"
+            gpu_backend_version: "5.6"
+          - gpu_backend: "hip_rocm"
+            gpu_backend_version: "11.8"
+      fail-fast: false
+    steps:
+      - name: Checkout Git Repository
+        uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Free additional space on runner
+        run: .github/workflows/helpers/free_space_on_runner.sh
+
+      - name: Build Legion
+        env:
+          FF_GPU_BACKEND: ${{ matrix.gpu_backend }}
+        run: .github/workflows/helpers/prebuild_legion.sh
+
+      - name: Archive compiled Legion library (CUDA)
+        env:
+          FF_GPU_BACKEND: ${{ matrix.gpu_backend }}
+        uses: actions/upload-artifact@v3
+        with:
+          name: legion_ubuntu-20.04_${{ matrix.gpu_backend }}-${{ matrix.gpu_backend_version }}_py${{ matrix.python_version }}
+          path: prebuilt_legion_assets/legion_ubuntu-20.04_${{ matrix.gpu_backend }}-${{ matrix.gpu_backend_version }}_py${{ matrix.python_version }}.tar.gz
+
+  create-release:
+    name: Create new release
+    runs-on: ubuntu-20.04
+    needs: prebuild-legion
+    steps:
+      - name: Checkout Git Repository
+        uses: actions/checkout@v3
+      - name: Free additional space on runner
+        run: .github/workflows/helpers/free_space_on_runner.sh
+      - name: Create folder for artifacts
+        run: mkdir artifacts unwrapped_artifacts
+      - name: Download artifacts
+        uses: actions/download-artifact@v3
+        with:
+          path: ./artifacts
+      - name: Display structure of downloaded files
+        working-directory: ./artifacts
+        run: ls -R
+      - name: Unwrap all artifacts
+        working-directory: ./artifacts
+        run: find . -maxdepth 2 -mindepth 2 -type f -name "*.tar.gz" -exec mv {} ../unwrapped_artifacts/  \;
+      - name: Get datetime
+        run: echo "RELEASE_DATETIME=$(date '+%Y-%m-%dT%H-%M-%S')" >> $GITHUB_ENV
+      - name: Release
+        env:
+          NAME: ${{ env.RELEASE_DATETIME }}
+          TAG_NAME: ${{ env.RELEASE_DATETIME }}
+          GITHUB_TOKEN: ${{ secrets.FLEXFLOW_TOKEN }}
+        run: gh release create $TAG_NAME ./unwrapped_artifacts/*.tar.gz --repo flexflow/flexflow-third-party