Merge branch 'rocm-main' into ci-upstream-sync-151_1

Author: charleshofer

.bazelrc

@@ -130,6 +130,8 @@ build:clang --copt=-Wno-gnu-offsetof-extensions
 build:clang --copt=-Qunused-arguments
 # Error on struct/class mismatches, since this causes link failures on Windows.
 build:clang --copt=-Werror=mismatched-tags
+# Don't error out on C++23 extensions. Needed for building the clang-19.
+build:clang --copt=-Wno-error=c23-extensions
 
 # Configs for CUDA
 build:cuda --repo_env TF_NEED_CUDA=1

.github/CODEOWNERS

@@ -0,0 +1,3 @@
+# Require approvals from someone on the JAX team before PRs are merged
+* @ROCm/jax-devs
+

.github/workflows/ci-build.yaml

@@ -1,4 +1,4 @@
-name: CI
+name: ROCm CPU CI
 
 # We test all supported Python versions as follows:
 # - 3.10 : Documentation build
@@ -11,10 +11,10 @@ on:
   # but only for the main branch
   push:
     branches:
-      - main
+      - rocm-main
   pull_request:
     branches:
-      - main
+      - rocm-main
 
 permissions:
   contents: read  # to fetch code
@@ -42,12 +42,8 @@ jobs:
       - run: pre-commit run --show-diff-on-failure --color=always --all-files
 
   build:
-    # Don't execute in fork due to runner type
-    if: github.repository == 'jax-ml/jax'
     name: "build ${{ matrix.name-prefix }} (py ${{ matrix.python-version }} on ubuntu-20.04, x64=${{ matrix.enable-x64}})"
-    runs-on: linux-x86-n2-32
-    container:
-      image: index.docker.io/library/ubuntu@sha256:6d8d9799fe6ab3221965efac00b4c34a2bcc102c086a58dff9e19a08b913c7ef # ratchet:ubuntu:20.04
+    runs-on: ROCM-Ubuntu
     timeout-minutes: 60
     strategy:
       matrix:
@@ -65,10 +61,6 @@ jobs:
             num_generated_cases: 1
     steps:
     - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v4.2.2
-    - name: Image Setup
-      run: |
-        apt update
-        apt install -y libssl-dev
     - name: Set up Python ${{ matrix.python-version }}
       uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38  # v5.4.0
       with:
@@ -95,12 +87,12 @@ jobs:
         echo "JAX_THREEFRY_PARTITIONABLE=$JAX_THREEFRY_PARTITIONABLE"
         echo "JAX_ENABLE_CHECKS=$JAX_ENABLE_CHECKS"
         echo "JAX_SKIP_SLOW_TESTS=$JAX_SKIP_SLOW_TESTS"
-        pytest -n auto --tb=short --maxfail=20 tests examples
+        pytest -n 4 --tb=short --maxfail=20 tests examples
 
 
   documentation:
     name: Documentation - test code snippets
-    runs-on: ubuntu-latest
+    runs-on: ROCM-Ubuntu
     timeout-minutes: 10
     strategy:
       matrix:
@@ -128,19 +120,13 @@ jobs:
 
   documentation_render:
     name: Documentation - render documentation
-    runs-on: linux-x86-n2-16
-    container:
-      image: index.docker.io/library/ubuntu@sha256:6d8d9799fe6ab3221965efac00b4c34a2bcc102c086a58dff9e19a08b913c7ef # ratchet:ubuntu:20.04
-    timeout-minutes: 10
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
     strategy:
       matrix:
         python-version: ['3.10']
     steps:
     - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v4.2.2
-    - name: Image Setup
-      run: |
-        apt update
-        apt install -y libssl-dev libsqlite3-dev
     - name: Set up Python ${{ matrix.python-version }}
       uses: actions/setup-python@42375524e23c412d93fb67b49958b491fce71c38  # v5.4.0
       with:
@@ -194,9 +180,7 @@ jobs:
 
   ffi:
     name: FFI example
-    runs-on: linux-x86-g2-16-l4-1gpu
-    container:
-      image: index.docker.io/tensorflow/build:latest-python3.12@sha256:48e99608fe9434ada5b14e19fdfd8e64f4cfc83aacd328b9c2101b210e984295 # ratchet:index.docker.io/tensorflow/build:latest-python3.12
+    runs-on: ROCM-Ubuntu
     timeout-minutes: 30
     steps:
     - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v4.2.2
@@ -207,7 +191,8 @@ jobs:
     - name: Install JAX
       run: |
         pip install uv~=0.5.30
-        uv pip install --system .[cuda12]
+        pip install uv
+        uv pip install --system .
     - name: Build and install example project
       run: uv pip install --system ./examples/ffi[test]
       env:
@@ -216,10 +201,11 @@ jobs:
         # a different toolchain. GCC is the default compiler on the
         # 'ubuntu-latest' runner, but we still set this explicitly just to be
         # clear.
-        CMAKE_ARGS: -DCMAKE_CXX_COMPILER=g++ -DJAX_FFI_EXAMPLE_ENABLE_CUDA=ON
+        CMAKE_ARGS: -DCMAKE_CXX_COMPILER=g++ #-DJAX_FFI_EXAMPLE_ENABLE_CUDA=ON
     - name: Run CPU tests
       run: python -m pytest examples/ffi/tests
       env:
         JAX_PLATFORM_NAME: cpu
     - name: Run GPU tests
       run: python -m pytest examples/ffi/tests
+

.github/workflows/rocm-ci.yml

@@ -1,23 +1,27 @@
-name: ROCm GPU Post-Merge Check
+name: ROCm GPU CI
 
 on:
-  # Trigger the workflow after a push into the main branch
+  # Trigger the workflow on push or pull request,
+  # but only for the rocm-main branch
   push:
     branches:
-      - main
-
-permissions:
-  contents: read
+      - rocm-main
+      - 'rocm-jaxlib-v*'
+  pull_request:
+    branches:
+      - rocm-main
+      - 'rocm-jaxlib-v*'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref || github.ref }}
+  cancel-in-progress: true
 
 jobs:
-  build-jax-in-docker:
-    runs-on: linux-x86_64-cirrascale-64-8gpu-amd-mi250
+  build-jax-in-docker: # strategy and matrix come here
+    runs-on:  mi-250
     env:
       BASE_IMAGE: "ubuntu:22.04"
-      TEST_IMAGE: ubuntu-jax-upstream-${{ github.run_id }}_${{ github.run_number }}_${{ github.run_attempt }}
+      TEST_IMAGE: ubuntu-jax-${{ github.run_id }}_${{ github.run_number }}_${{ github.run_attempt }}
       PYTHON_VERSION: "3.10"
       ROCM_VERSION: "6.2.4"
       WORKSPACE_DIR: workdir_${{ github.run_id }}_${{ github.run_number }}_${{ github.run_attempt }}
@@ -32,6 +36,9 @@ jobs:
           ls
       - name: Print system info
         run: |
+          whoami
+          printenv
+          df -h
           rocm-smi
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v4.2.2
         with:
@@ -50,9 +57,12 @@ jobs:
         uses: actions/upload-artifact@v4
         with:
           name: rocm_jax_r${{ env.ROCM_VERSION }}_py${{ env.PYTHON_VERSION  }}_id${{ github.run_id }}
-          path: ${{ env.WORKSPACE_DIR }}/dist/*.whl
-          retention-days: 2
+          path: ./dist/*.whl
       - name: Run tests
+        env:
+          ROCM_TEST_INCLUDE_SKIPS: "1"
+          GPU_COUNT: "8"
+          GFX: "gfx90a"
         run: |
           cd $WORKSPACE_DIR
           python3 build/rocm/ci_build test $TEST_IMAGE --test-cmd "pytest tests/core_test.py"

.github/workflows/rocm-nightly-upstream-sync.yml

@@ -0,0 +1,66 @@
+# Pulls the latest changes from upstream into main and opens a PR to merge
+# them into rocm-main branch.
+
+name: ROCm Nightly Upstream Sync
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: '0 6 * * 1-5'
+permissions:
+  contents: write
+  pull-requests: write
+env:
+  SYNC_BRANCH_NAME: ci-upstream-sync-${{ github.run_number }}_${{ github.run_attempt }}
+jobs:
+  sync-main:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Generate an app token
+        id: generate-token
+        uses: actions/create-github-app-token@v1
+        with:
+          app-id: ${{ vars.ROCM_REPO_MANAGEMENT_API_2_ID }}
+          private-key: ${{ secrets.ROCM_REPO_MANAGEMENT_API_2_PRIV_KEY }}
+      - name: Sync our main with upstream main
+        run: |
+          gh auth status
+          gh repo sync rocm/jax -b main
+        env:
+          GH_TOKEN: ${{ steps.generate-token.outputs.token }}
+  create-sync-branch:
+    needs: sync-main
+    runs-on: ubuntu-latest
+    env:
+      GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v4.2.2
+      - name: Create branch
+        run: |
+          git fetch
+          git checkout origin/main 
+          git checkout -b $SYNC_BRANCH_NAME
+          # Try and merge rocm-main into this new branch so that we don't run upstream's CI code
+          git config --global user.email "[email protected]"
+          git config --global user.name "GitHub Actions"
+          git merge origin/rocm-main || true
+          # If the merge creates conflicts, we want to abort and push to origin anyways so that a dev can resolve the conflicts
+          git merge --abort || true
+          git push origin HEAD
+  open-sync-pr:
+    needs: create-sync-branch
+    runs-on: ubuntu-latest
+    steps:
+      - name: Generate an app token
+        id: generate-token
+        uses: actions/create-github-app-token@v1
+        with:
+          app-id: ${{ vars.ROCM_REPO_MANAGEMENT_API_2_ID }}
+          private-key: ${{ secrets.ROCM_REPO_MANAGEMENT_API_2_PRIV_KEY }}
+      - name: Open a PR to rocm-main
+        run: |
+          gh pr create --repo $GITHUB_REPOSITORY --head $SYNC_BRANCH_NAME --base rocm-main --title "CI: $(date +%x) upstream sync" --body "Daily sync with upstream"
+          gh pr merge --repo $GITHUB_REPOSITORY --merge --auto $SYNC_BRANCH_NAME
+        env:
+          GH_TOKEN: ${{ steps.generate-token.outputs.token }}
+

.github/workflows/rocm-open-upstream-pr.yml

@@ -0,0 +1,41 @@
+name: ROCm Open Upstream PR
+on:
+  pull_request:
+    types: [ labeled ]
+    branches: [ rocm-main ]
+jobs:
+  open-upstream:
+    if: ${{ github.event.label.name == 'open-upstream' }}
+    permissions:
+      contents: write
+      pull-requests: write
+    runs-on: ubuntu-latest
+    env:
+      NEW_BRANCH_NAME: "${{ github.head_ref }}-upstream"
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v4.2.2
+      - name: Rebase code to main
+        run: |
+          git config --global user.email "[email protected]"
+          git config --global user.name "Github Actions"
+          git fetch
+          git checkout -b $NEW_BRANCH_NAME origin/${{ github.head_ref }}
+          git rebase --onto origin/main origin/rocm-main
+          # Force push here so that we don't run into conflicts with the origin branch
+          git push origin HEAD --force
+      - name: Leave link to create PR
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          # Bash is not friendly with newline characters, so make our own
+          NL=$'\n'
+          # Encode the PR title and body for passing as URL get parameters
+          TITLE_ENC=$(jq -rn --arg x "[ROCm] ${{ github.event.pull_request.title }}" '$x|@uri')
+          BODY_ENC=$(jq -rn --arg x $"${{ github.event.pull_request.body }}${NL}${NL}Created from: rocm/jax#${{ github.event.pull_request.number }}" '$x|@uri')
+          # Create a link to the that will open up a new PR form to upstream and autofill the fields
+          CREATE_PR_LINK="https://github.com/jax-ml/jax/compare/main...ROCm:jax:$NEW_BRANCH_NAME?expand=1&title=$TITLE_ENC&body=$BODY_ENC"
+          # Add a comment with the link to the PR
+          COMMENT_BODY="Feature branch from main is ready. [Create a new PR][1] destined for upstream?${NL}${NL}[1]: $CREATE_PR_LINK"
+          gh pr comment ${{ github.event.pull_request.number }} --repo rocm/jax --body "$COMMENT_BODY"
+

.github/workflows/upstream-nightly.yml

@@ -22,7 +22,7 @@ on:
 
 jobs:
   upstream-dev:
-    runs-on: ubuntu-latest
+    runs-on: ROCM-Ubuntu
     permissions:
       contents: read
       issues: write  # for failed-build-issue

build/build.py

@@ -262,7 +262,7 @@ def add_artifact_subcommand_arguments(parser: argparse.ArgumentParser):
   rocm_group.add_argument(
       "--rocm_amdgpu_targets",
       type=str,
-      default="gfx900,gfx906,gfx908,gfx90a,gfx940,gfx941,gfx942,gfx1030,gfx1100,gfx1200,gfx1201",
+      default="gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1030,gfx1100,gfx1101,gfx1200,gfx1201",
       help="A comma-separated list of ROCm amdgpu targets to support.",
   )
 

build/rocm/Dockerfile.ms

@@ -13,7 +13,7 @@ RUN apt-get update && apt-get install -y \
     && rm -rf /var/lib/apt/lists/*
 
 # Add target file to help determine which device(s) to build for
-ARG GPU_DEVICE_TARGETS="gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx1100 gfx1200 gfx1201"
+ARG GPU_DEVICE_TARGETS="gfx900 gfx906 gfx908 gfx90a gfx942 gfx1030 gfx1100 gfx1101 gfx1200 gfx1201"
 ENV GPU_DEVICE_TARGETS=${GPU_DEVICE_TARGETS}
 
 # Install ROCm
@@ -62,7 +62,6 @@ RUN --mount=type=cache,mode=0755,target=/root/.cache/pip \
         pytest-reportlog \
         pytest-rerunfailures \
         pytest-json-report \
-        pytest-csv \
         cloudpickle \
         portpicker \
         matplotlib \

build/rocm/README.md

@@ -207,3 +207,4 @@ This will generate three wheels in the `dist/` directory:
 ### Simplified Build Script
 
 For a streamlined process, consider using the `jax/build/rocm/dev_build_rocm.py` script.
+