Fix GPU CI startup and move CPU jobs to GitHub Actions

[jpsamaroo]

The Buildkite run for #769 aborts CUDA, ROCm, and oneAPI tests during setup: installing a GPU backend downgrades LLVM after dependencies are already loaded. Workers also launch in the repository project instead of the temporary Pkg.test environment, so their dependency versions can differ from the driver's.

Resolve optional packages together before imports and worker startup, and launch workers with Base.active_project(). Install PythonCall and CondaPkg only when the Python suite is selected, so Metal tests no longer abort during unrelated Python initialization. Default CPU runs still select the Python suite.

Run Finch tests in a dedicated GitHub Actions job on ubuntu-latest, preserving Julia 1.11, the 90-minute timeout, test selection, skip condition, and coverage reporting. Remove the duplicate CPU tests and benchmarks from the OpenCL Buildkite pipeline; both Buildkite pipelines now contain GPU tests, benchmarks, and their reports. CPU coverage remains in GitHub Actions.

GPU runtime fixes are covered by #758; the reduction fix remains in #769.

The Python failures also occur in unrelated PR #771. Both fail in CondaPkg 0.2.37 when it opens a nonexistent global Julia Project.toml; the upstream fix is merged. PythonCall also has a merged fix for conda-forge Python 3.15 shared-library discovery. These fixes were unreleased at validation time. Testing both fixed commits together passes the Python suite; no dependency workaround is added here.

Validation on Julia 1.11.9:

• Forced LLVM 10 → 9.13.2 with CUDA 5.8.5; CUDA loads on the driver and three workers with matching versions and the same active test project.
• Focused CPU mapreduce suite: 421/421 tests pass.
• GPU bootstrap excludes Python. Selecting extlang/python installs both dependencies and enters the suite.
• Python suite: 6/6 pass with one and four driver threads plus three workers, using CondaPkg commit 0187461 and PythonCall commit 4a7e029 in an isolated test project. Released versions still encounter the upstream initialization failures described above.
• CI_TEST_FINCH=1 test-runner simulation selects only array/linalg/matmul_finch.
• Workflow and pipeline YAML parse successfully; all remaining Buildkite step dependencies resolve and every job is a GPU test, benchmark, or report.
• All six Buildkite report and matrix-expansion tests pass.
• Julia syntax and git diff --check pass. Full Finch execution, GPU execution, and hosted CI reruns remain for CI.

[github-actions]

Dagger benchmarks: dirty vs master

Counts are benchmark metrics (time, allocations, bytes). Timing changes require non-overlapping median ± IQR bands; both revisions need at least five timed samples, and timing regressions need an independent confirmation run.

Backend	Regressions	Improvements	Within noise	Inconclusive time	Suites
Threads	0	1	10	1	4/4
Distributed	4	3	1	1	4/4
Distributed+Threads	0	0	0	1	4/4
MPI	0	0	0	2	4/4
MPI+Threads	1	2	18	1	4/4
Total	5	6	29	6	20/20

Regressions across all backends

Backend	Benchmark	Metric	Change
Distributed	linalg/dagger/N=1024 (block 128)/matmul (A*A)	allocs	+113.6%
Distributed	linalg/dagger/N=1024 (block 128)/matmul (A*A)	memory	+112.3%
Distributed	linalg/dagger/N=1024 (block 128)/syrk (A'*A)	allocs	+61.8%
MPI+Threads	stencil/dagger/N=1024 (block 128)/assign (const)	time	+45.1%
Distributed	linalg/dagger/N=1024 (block 128)/syrk (A'*A)	memory	+40.5%

Improvements across all backends

Backend	Benchmark	Metric	Change
MPI+Threads	linalg/dagger/N=256 (block 256)/matvec (A*x)	time	-46.3%
Threads	array/dagger/N=1024 (block 512)/norm	time	-42.7%
MPI+Threads	array/dagger/N=1024 (block 512)/norm	time	-38.9%
Distributed	array/dagger/N=1024 (block 128)/add (X + X)	memory	-38.2%
Distributed	sparse/dagger/N=1024 (block 64)/cg solve (laplacian)	memory	-34.5%
Distributed	sparse/dagger/N=1024 (block 64)/cg solve (laplacian)	allocs	-31.2%

Threads (1 process × 4 threads) — 0 regressions, 1 improvements

Suite	Regressions	Improvements	Within noise	Inconclusive time
array	0	1	4	0
linalg	0	0	3	0
sparse	0	0	0	0
stencil	0	0	3	1

Improvements

Benchmark	Metric	Change
array/dagger/N=1024 (block 512)/norm	time	-42.7%

Within noise

Benchmark	Metric	Change
stencil/dagger/N=256 (block 128)/alloc (neighbors Wrap)	time	+174.6%
array/dagger/N=256 (block 256)/broadcast (X .+ 1)	time	+71.0%
stencil/dagger/N=256 (block 128)/multi-expr	time	+38.0%
array/dagger/N=256 (block 128)/map (sin.(X))	time	+37.8%
stencil/dagger/N=1024 (block 512)/assign (const)	time	+37.8%
linalg/dagger/N=256 (block 128)/syrk (A'*A)	time	+31.3%
array/dagger/N=256 (block 256)/norm	time	+28.8%
linalg/dagger/N=1024 (block 512)/matvec (A*x)	time	-34.5%
array/dagger/N=256 (block 128)/norm	time	-36.6%
linalg/dagger/N=256 (block 128)/matmul (A*A)	time	-36.7%

Inconclusive timing changes

Benchmark	Metric	Change
stencil/dagger/N=1024 (block 512)/neighbors (Pad)	time	+50.0%

Distributed (4 processes × 1 thread) — 4 regressions, 3 improvements

Suite	Regressions	Improvements	Within noise	Inconclusive time
array	0	1	0	0
linalg	4	0	1	1
sparse	0	2	0	0
stencil	0	0	0	0

Regressions

Benchmark	Metric	Change
linalg/dagger/N=1024 (block 128)/matmul (A*A)	allocs	+113.6%
linalg/dagger/N=1024 (block 128)/matmul (A*A)	memory	+112.3%
linalg/dagger/N=1024 (block 128)/syrk (A'*A)	allocs	+61.8%
linalg/dagger/N=1024 (block 128)/syrk (A'*A)	memory	+40.5%

Improvements

Benchmark	Metric	Change
array/dagger/N=1024 (block 128)/add (X + X)	memory	-38.2%
sparse/dagger/N=1024 (block 64)/cg solve (laplacian)	memory	-34.5%
sparse/dagger/N=1024 (block 64)/cg solve (laplacian)	allocs	-31.2%

Within noise

Benchmark	Metric	Change
linalg/dagger/N=1024 (block 128)/matmul (A*A)	time	+58.6%

Inconclusive timing changes

Benchmark	Metric	Change
linalg/dagger/N=1024 (block 128)/solve (A\b via lu)	time	+165.9%

Distributed+Threads (2 processes × 2 threads) — 0 regressions, 0 improvements

Suite	Regressions	Improvements	Within noise	Inconclusive time
array	0	0	0	1
linalg	0	0	0	0
sparse	0	0	0	0
stencil	0	0	0	0

Inconclusive timing changes

Benchmark	Metric	Change
array/dagger/N=1024 (block 512)/alloc (rand)	time	+142.7%

MPI (4 ranks × 1 thread) — 0 regressions, 0 improvements

Suite	Regressions	Improvements	Within noise	Inconclusive time
array	0	0	0	2
linalg	0	0	0	0
sparse	0	0	0	0
stencil	0	0	0	0

Inconclusive timing changes

Benchmark	Metric	Change
array/dagger/N=1024 (block 512)/add (X + X)	time	+46.7%
array/dagger/N=1024 (block 512)/norm	time	+45.0%

MPI+Threads (2 ranks × 2 threads) — 1 regressions, 2 improvements

Suite	Regressions	Improvements	Within noise	Inconclusive time
array	0	1	9	0
linalg	0	1	3	0
sparse	0	0	0	0
stencil	1	0	6	1

Regressions

Benchmark	Metric	Change
stencil/dagger/N=1024 (block 128)/assign (const)	time	+45.1%

Improvements

Benchmark	Metric	Change
linalg/dagger/N=256 (block 256)/matvec (A*x)	time	-46.3%
array/dagger/N=1024 (block 512)/norm	time	-38.9%

Within noise

Benchmark	Metric	Change
array/dagger/N=1024 (block 512)/broadcast (X .+ 1)	time	+96.1%
array/dagger/N=256 (block 128)/norm	time	+95.3%
stencil/dagger/N=256 (block 256)/update (+)	time	+94.9%
stencil/dagger/N=256 (block 256)/neighbors (Pad)	time	+56.1%
stencil/dagger/N=256 (block 256)/assign (const)	time	+54.0%
array/dagger/N=256 (block 256)/broadcast (X .+ 1)	time	+51.5%
stencil/dagger/N=1024 (block 512)/update (+)	time	+48.9%
array/dagger/N=1024 (block 128)/reduce (sum)	time	+46.0%
linalg/dagger/N=256 (block 256)/qr	time	+45.2%
stencil/dagger/N=256 (block 256)/multi-expr	time	+43.6%
array/dagger/N=1024 (block 512)/map (sin.(X))	time	+40.2%
linalg/dagger/N=256 (block 128)/cholesky	time	+39.7%
array/dagger/N=256 (block 256)/add (X + X)	time	+39.6%
array/dagger/N=256 (block 128)/alloc (rand)	time	+37.6%
stencil/dagger/N=256 (block 256)/neighbors (Wrap)	time	+36.1%
array/dagger/N=1024 (block 512)/alloc (rand)	time	-35.7%
linalg/dagger/N=1024 (block 512)/matmul (A*A)	time	-47.6%
array/dagger/N=256 (block 128)/transpose (permutedims)	time	-56.8%

Inconclusive timing changes

Benchmark	Metric	Change
stencil/dagger/N=1024 (block 512)/multi-expr	time	+93.2%

Full results and plots (download the benchmark-results-* artifacts).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 61.88%. Comparing base (0238175) to head (c52f47e).
⚠️ Report is 10 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #772      +/-   ##
==========================================
+ Coverage   60.68%   61.88%   +1.20%     
==========================================
  Files         125      125              
  Lines       18159    18155       -4     
==========================================
+ Hits        11019    11235     +216     
+ Misses       7140     6920     -220     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.