Create pkgci.yml and pkgci_build_packages.yml.

[ScottTodd]

Progress on #584.

Summary

• Added .github/workflows/pkgci_build_packages.yml that builds sharktank, shortfin, and shark-ai dev packages and upload them to GitHub artifacts.
  • With ccache configured and tracing disabled, this job takes around 2 minutes to run.
• Added .github/workflows/pkgci.yml that just runs pkgci_build_packages.yml for now. Other jobs can be migrated to depend on that job and use/test the packages.

Other details

• Experimented a bit with ccache. I'm seeing a ~30% cache hit rate but that still gets the shortfin build from 2m30s (test logs here) down to 1m30s on standard GitHub-hosted ubuntu-24.04 runners (test logs here).
• Plumbed the SHORTFIN_ENABLE_TRACING setting through scripts / Docker. For dev packages we can keep tracing disabled (unless there is a clear reason to add it). If the cache hit rate improves then we might be able to enable tracing for low cost.
• Dropped cache: "pip" from build_packages.yml since it is counterproductive for a job that only installs packaging. Multiple workflows seem to be writing to the same cache and I see no way to customize the cache key. That, or the cache is unnecessarily large and we just need to prune it manually.

[marbre]

Nice! Just two minor things.

[marbre]

Do we need aarch64 support? We probably can just install ccache for x86_64.

[ScottTodd]

We might at some point. I've written this code a few times before, most recently at https://github.com/iree-org/base-docker-images/blob/main/build_tools/install_ccache.sh, and supporting both architectures isn't much extra code. Better to write cross-platform/architecture code whenever possible instead of artificially limiting ourselves.

I plan on updating https://github.com/nod-ai/base-docker-images/blob/main/dockerfiles/manylinux_x86_64.Dockerfile to more closely match https://github.com/iree-org/base-docker-images/blob/main/dockerfiles/manylinux_x86_64.Dockerfile, as part of upgrading from manylinux2014 to manylinux_2_28:

shark-ai/shortfin/build_tools/build_linux_package.sh

Lines 39 to 43 in 779adc3

	# TODO(#130): Update to manylinux_2_28, upstream or a fork
	# * upstream uses a version of gcc that has build warnings/errors
	# * https://github.com/nod-ai/base-docker-images is a bit out of date but can include a recent clang
	# MANYLINUX_DOCKER_IMAGE="${MANYLINUX_DOCKER_IMAGE:-quay.io/pypa/manylinux_2_28_${ARCH}:latest}"
	MANYLINUX_DOCKER_IMAGE="${MANYLINUX_DOCKER_IMAGE:-quay.io/pypa/manylinux2014_${ARCH}:latest}"

[ScottTodd]

Next steps with this:

• Refactor at least one workflow to use the package artifacts
• Put some more thought into how workflows can get different versions of the IREE / iree-turbine / etc. dependencies. Should they keep doing this?

  shark-ai/.github/workflows/ci-shark-ai.yml

  Lines 63 to 72 in e37b934

  	# Install latest iree-tubrine.
  	pip install --no-compile -f https://iree.dev/pip-release-links.html --src deps \
  	-e "git+https://github.com/iree-org/iree-turbine.git#egg=iree-turbine"
  	# Try with the latest IREE nightly releases, not what iree-turbine pins.
  	# We could also pin to a known working or stable version.
  	# This should eventually stabilize. Do the best we can for now.
  	pip install -f https://iree.dev/pip-release-links.html --upgrade --pre \
  	iree-base-compiler \
  	iree-base-runtime

[ScottTodd]

Picking this back up now.

[renxida]

Extremely excited for every bit of CI time reduction.

[renxida]

YES!