ci: migrate from pip to uv for faster installs

[hudson-ai]

Motivation

Dependency installation is a significant bottleneck in our CI pipelines. Looking at recent runs of the pull_request workflow:

• CPU integration tests spend ~200s installing deps out of ~300s total (65% of wall time)
• Windows unit tests spend 50–110s on installs (38–63% of wall time)
• Even Linux/macOS unit tests spend ~20s (17–29%)

uv is a drop-in replacement for pip that's 10–100× faster at dependency resolution and installation. Conservatively, this should cut install times by 5–10×, saving minutes per CI run — especially on the heavier workflows (cpu_tests, gpu_tests, notebooks).

Changes

Replace pip install with uv pip install --system across all 8 CI workflows that install Python packages:

• Add astral-sh/setup-uv@v5 step to all jobs
• Replace all pip install / python -m pip install → uv pip install --system
• Remove now-unnecessary pip install --upgrade pip steps
• Use uv build instead of python -m pip install build && python -m build in pypi_upload

No changes to pyproject.toml, the build backend (still setuptools), or test configuration. The --system flag tells uv to install into the system Python rather than creating a venv, matching the existing CI behavior.

Affected workflows

Workflow	Install time (before)	% of total
call_cpu_tests	~200s	65%
call_gpu_tests	similar	similar
pull_request (unit_tests, Windows)	50–110s	38–63%
pull_request (unit_tests, Linux/macOS)	11–22s	17–29%
code_quality	light	—
ci_docs	moderate	—
ci_credentials	moderate	—
notebook_tests	moderate	—
pypi_upload	light	—

What this doesn't do (yet)

This PR only swaps the installer. A natural follow-up would be adding uv lock to generate a uv.lock file for fully reproducible CI environments with pinned transitive dependencies.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 60.34%. Comparing base (376121b) to head (3c1e837).
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1431       +/-   ##
===========================================
+ Coverage   42.56%   60.34%   +17.77%     
===========================================
  Files          62       62               
  Lines        5248     5248               
===========================================
+ Hits         2234     3167      +933     
+ Misses       3014     2081      -933     

☔ View full report in Codecov by Sentry.
📢 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.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[hudson-ai]

Measured speedups from PR gate run

Comparing the latest PR gate run (uv) against the most recent main run (pip):

CPU integration tests

Job	pip	uv	Speedup
cpu_tests (3.10)	204s	56s	3.6×
cpu_tests (3.14)	174s	75s	2.3×

Unit tests — Windows (biggest absolute wins)

Job	pip	uv	Speedup
Win 3.14	110s	21s	5.2×
Win 3.10	92s	25s	3.7×
Win 3.12	59s	19s	3.1×
Win 3.11	54s	22s	2.5×
Win 3.13	53s	17s	3.1×

Unit tests — Linux

Job	pip	uv	Speedup
Linux 3.10	20s	3s	6.7×
Linux 3.11	19s	3s	6.3×
Linux 3.12	21s	3s	7×
Linux 3.13	20s	5s	4×
Linux 3.14	22s	3s	7.3×

Unit tests — macOS

Job	pip	uv	Speedup
macOS 3.10	11s	4s	2.8×
macOS 3.11	17s	7s	2.4×
macOS 3.12	13s	3s	4.3×
macOS 3.13	13s	3s	4.3×
macOS 3.14	15s	3s	5×

All 17 jobs passed ✅

[hudson-ai]

ci_linux results

Compared the ci/uv-install run against the latest main run. 37/38 jobs passed (one pre-existing ONNX segfault on Python 3.10, unrelated).

Per-job install speedups

CPU tests: install time dropped from ~175–228s to 73–91s (2.0–2.7×)
GPU tests: install time dropped from ~790–880s to 660–700s (~1.2×, dominated by CUDA setup)

Wall-time impact

Since these jobs are parallelized, what matters is the critical path:

Critical path job	pip (total)	uv (total)	Wall time saved
Longest GPU job	~1040s	~885s	~2.5 min
Longest CPU job	~600s	~500s	~1.5 min

The biggest impact is on the PR gate workflow, where install time is a larger fraction of total job time — especially on Windows (up to 5× faster installs, see previous comment).

[riedgar-ms]

If this is shared amongst all builds, perhaps make it a variable?

[riedgar-ms]

Fine with me