CI: build relocatable packages via TheRock SDK by thananon · Pull Request #262 · ROCm/TransferBench

thananon · 2026-04-23T23:40:46Z

Summary

Adds GitHub Actions workflow and build script for producing relocatable packages (DEB/RPM/TGZ) via the TheRock SDK. This workflow is the primary path for continuous integration and nightly/release builds of TransferBench packages, enabling self-contained deployments without requiring a pre-installed ROCm stack on the target machine.

Prerequisites: Merged #261 (manylinux portability fixes) to ensure builds succeed on manylinux_2_28 (AlmaLinux 8) with gcc 8-era libstdc++.

Implementation

.github/workflows/build-relocatable-packages.yml: Two-job workflow (Ubuntu 22.04 → DEB/TGZ, manylinux_2_28 → RPM/TGZ) that fetches the latest TheRock SDK tarball, builds with -DBUILD_RELOCATABLE_PACKAGE=ON, and uploads artifacts to GitHub Actions + optional S3 (via OIDC).
build_packages_local.sh: Standalone script that downloads TheRock SDK tarballs from S3 by ROCm version and GPU family (gfx94X-dcgpu, gfx950-dcgpu, gfx110X-all, gfx120X-all, gfx1151), extracts them, drives CMake with relocatable RPATH, and produces amdrocm<MAJOR>-transferbench packages.
CMake additions: New BUILD_RELOCATABLE_PACKAGE option that bypasses rocm_install/rocm_create_package and drives CPack directly with version-namespaced package naming (amdrocm7-transferbench) and relocatable RPATH.

S3 and OIDC

Workflow uploads packages to S3 using GitHub OIDC for authentication (no long-lived credentials). S3 paths:

Nightly/develop: s3://.../nightly/transferbench/{deb,rpm,tar} with apt/yum repo metadata.
Release branches: s3://.../release/transferbench/{deb,rpm,tar} with apt/yum metadata.
Pull requests: s3://.../transferbench/<branch>/<run-number>/{ubuntu-22.04,manylinux_2_28} (no metadata).

S3 upload only executes in the ROCm/TransferBench repository and requires AWS_S3_BUCKET (organization variable) and AWS_ROLE_ARN (organization secret).

Supported GPU Families

TheRock SDK tarballs are fetched per GPU family:

gfx94X-dcgpu: gfx940, gfx941, gfx942, gfx943 (MI300 series)
gfx950-dcgpu: gfx950 (upcoming MI400 architecture)
gfx110X-all: gfx1100, gfx1101, gfx1102 (RDNA3 client)
gfx120X-all: gfx1200, gfx1201 (RDNA4)
gfx1151: gfx1151 (Phoenix APU)

The workflow defaults to gfx94X-dcgpu (data-center workhorse); other families can be selected via workflow_dispatch input.

Trigger Conditions

Push to develop, mainline, release/**, users/thananon/therock-packages-workflow-v2 (the last one is temporary for pre-merge validation of this PR; will be removed after merge)
Pull request to develop or mainline
Daily scheduled run at 13:00 UTC (5:00 AM PST)
Manual via workflow_dispatch with optional ROCm version/GPU family override

Testing

Local invocation (no GitHub Actions):

export GPU_FAMILY=gfx94X-dcgpu BUILD_TYPE=Release
./build_packages_local.sh
ls build/*.{deb,rpm,tar.gz}

The workflow triggers on push to this PR's branch (users/thananon/therock-packages-workflow-v2) so the first CI run will validate the full build→package→upload flow before the PR is merged.

Adds a CI workflow that builds DEB/RPM/TGZ packages of TransferBench against the TheRock nightly ROCm SDK, modeled on the ROCmValidationSuite packaging workflow. Packages install to /opt/rocm/extras-<MAJOR> with $ORIGIN-relative RPATH so they are relocatable. - build_packages_local.sh: single source of truth for both local and CI builds. Detects Ubuntu vs AlmaLinux/manylinux, installs deps, fetches TheRock SDK tarball, configures CMake with relocatable RPATH and the new BUILD_RELOCATABLE_PACKAGE option, builds, and invokes CPack for DEB/RPM/TGZ. - .github/workflows/build-relocatable-packages.yml: parallel Ubuntu 22.04 + manylinux_2_28 jobs triggered on push, PR, daily cron, and workflow_dispatch. OIDC-based S3 upload gated on AWS_S3_BUCKET being set; apt/yum repo metadata generated for non-PR builds. Build report artifact summarizes S3 paths. - .github/workflows/README_BUILD_PACKAGES.md: workflow docs covering triggers, local usage, S3 layout, IAM trust policy, and apt/yum install snippets. - CMakeLists.txt: new BUILD_RELOCATABLE_PACKAGE option that bypasses rocm_install/rocm_create_package, names the package amdrocm<MAJOR>-transferbench, and honors caller-set install prefix and CPACK_*_PACKAGE_RELEASE env vars. Default cmake .. behavior is unchanged. Co-Authored-By: Claude Opus 4 <[email protected]>

The tarballs are published as: therock-dist-linux-<family>-<version>.tar.gz not <family>-<version>.tar.gz as I had it. Also TheRock does not publish a per-family LATEST.txt, so the auto-fetch path now lists the bucket via S3 ListObjectsV2 and picks the highest version key with `sort -V`. Updates the pinned fallback to a version that actually exists on the bucket today (7.13.0a20260423). Co-Authored-By: Claude Opus 4 <[email protected]>

The bucket includes non-release ad-hoc builds (e.g. therock-dist-linux-gfx94X-dcgpu-ADHOCBUILD-7.0.0rc20250625.tar.gz) which `sort -V | tail -1` was selecting because 'A' lexically sorts after digits. The downstream `printf '%02d'` then crashed trying to parse `ADHOCBUILD-7` as the ROCm major. Restrict the auto-fetch grep to keys matching <prefix><MAJOR>.<MINOR>.<patch+suffix>.tar.gz so only properly versioned releases are considered. Co-Authored-By: Claude Opus 4 <[email protected]>

- Gate S3 upload + repo metadata steps on github.repository == 'ROCm/TransferBench' so forks that set AWS_S3_BUCKET don't try unauthenticated uploads. - Add EUID root check to build_packages_local.sh up front (clearer error than failing later in apt/dnf). - Rename TAROBALL_BASE -> TARBALL_BASE (typo). - Sanitize PKG_RELEASE: collapse non-alphanumerics into dots so feature-branch names stay valid in DEB/RPM release fields. - Drop ${ROCM_PATH}/lib from RPATH_LIST so the ephemeral SDK download path is not embedded into the packaged binary. - Make CPACK_RPM_EXCLUDE_FROM_AUTO_FILELIST_ADDITION track the actual CPACK_PACKAGING_INSTALL_PREFIX instead of hard-coded /opt/rocm/extras-N paths. Co-Authored-By: Claude Opus 4 <[email protected]>

Copilot

Pull request overview

Adds a GitHub Actions-based CI path (and a local build script) to produce relocatable TransferBench packages using TheRock ROCm SDK, with a new CMake packaging mode that drives CPack directly.

Changes:

Added build_packages_local.sh to download TheRock SDK, configure relocatable RPATH, build, and run CPack to emit DEB/RPM/TGZ.
Added .github/workflows/build-relocatable-packages.yml to build packages on Ubuntu 22.04 and manylinux_2_28, upload artifacts, and optionally publish to S3 via OIDC with repo metadata generation.
Updated CMakeLists.txt with a BUILD_RELOCATABLE_PACKAGE option to bypass rocm_create_package and set CPack metadata/package naming.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.

File	Description
`build_packages_local.sh`	Local/CI build driver: installs deps, downloads TheRock SDK, configures relocatable install/RPATH, builds + runs CPack.
`CMakeLists.txt`	Adds relocatable packaging mode and new package naming/CPack settings.
`.github/workflows/build-relocatable-packages.yml`	CI workflow to build/package on Ubuntu + manylinux, upload artifacts, and optionally push to S3 with OIDC + repo metadata.
`.github/workflows/README_BUILD_PACKAGES.md`	Documentation for the new workflow/script, triggers, S3 layout, and usage instructions.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Packages declare runtime deps on hsa-rocr/numactl and only ship the TransferBench binary; the install tree is movable via $ORIGIN RPATH but target systems still need the ROCm/HSA runtime. Update README so the claim matches the actual packaging behavior.

nileshnegi · 2026-04-24T04:46:29Z

+  -DCMAKE_VERBOSE_MAKEFILE=ON
+  -DBUILD_RELOCATABLE_PACKAGE=ON
+  -DBUILD_LOCAL_GPU_TARGET_ONLY=OFF
+  -DENABLE_NIC_EXEC=ON


i don't think we want this ON by default
@gilbertlee-amd thoughts?

nileshnegi · 2026-04-24T04:49:18Z

+    build-essential cmake git curl tar xz-utils ca-certificates pkg-config \
+    python3 python3-pip \
+    libnuma-dev libibverbs-dev rdma-core ibverbs-providers \
+    libopenmpi-dev openmpi-bin \


is this aligned with how TheRock sets/expects MPI dependencies for other packages?

thananon and others added 5 commits April 23, 2026 18:31

CI: trigger workflow on PR-B branch for pre-merge validation

56b63c6

thananon requested review from a team as code owners April 23, 2026 23:40

nileshnegi requested a review from Copilot April 23, 2026 23:48

Copilot started reviewing on behalf of nileshnegi April 23, 2026 23:49 View session

Copilot AI reviewed Apr 23, 2026

View reviewed changes

Comment thread CMakeLists.txt

Comment thread CMakeLists.txt

Comment thread .github/workflows/build-relocatable-packages.yml

Comment thread .github/workflows/build-relocatable-packages.yml

Comment thread .github/workflows/README_BUILD_PACKAGES.md Outdated

thananon added 2 commits April 23, 2026 20:06

ci: trigger relocatable packages workflow on candidate branch

a679c34

nileshnegi reviewed Apr 24, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

CI: build relocatable packages via TheRock SDK#262

CI: build relocatable packages via TheRock SDK#262
thananon wants to merge 7 commits intodevelopfrom
users/thananon/therock-packages-workflow-v2

thananon commented Apr 23, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

nileshnegi Apr 24, 2026

Uh oh!

nileshnegi Apr 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

thananon commented Apr 23, 2026

Summary

Implementation

S3 and OIDC

Supported GPU Families

Trigger Conditions

Testing

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

nileshnegi Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

nileshnegi Apr 24, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants