Skip to content

SolidCraft: an open-source parametric 3D CAD application built in pure Rust. Desktop, web, and automation through MCP.

CI: GPU path tests in four shards on 16 vCPU (#7096)

bherbruck84b1aeecaebb
diff --git a/.config/nextest.toml b/.config/nextest.toml
index 853777e723..2e00c234cd 100644
--- a/.config/nextest.toml
+++ b/.config/nextest.toml
@@ -1,6 +1,6 @@
# cargo-nextest settings for CI (`cargo xtask test` does not use nextest).
# CI builds the tests once and runs them in one job on a 32 vCPU runner (.github/workflows/ci.yml),
-# except the GPU path tracer's tests (the gpu-path group below), which run in three shards of their own.
+# except the GPU path tracer's tests (the gpu-path group below), which run in four shards of their own.
[profile.ci]
# Heavy tests (one to twelve minutes each on a CI runner) that pull requests skip. They run on
@@ -73,7 +73,7 @@ priority = 100
# Each nextest test has its own process/device. Concurrent cold lavapipe path
# compiles contend for CPU and can exhaust the 600 s pipeline setup budget.
# Serialize this module per CI shard; other tests retain their normal parallelism.
-# CI runs these in their own job (GPU path tests in ci.yml), split into hash partitions.
+# CI runs these in their own job (GPU path tests in ci.yml), split into count partitions.
[test-groups.gpu-path]
max-threads = 1
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 3f4bcc583d..0b9577a502 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -10,7 +10,7 @@
# no native job runs for them, on pull requests or on main.
# Tests are built and run in one job on a big runner, slowest first, starting beside the repository
# checks (no archive to upload and download). The GPU path tracer's tests run one at a time
-# (.config/nextest.toml), so they have a job of their own in three shards (GPU path tests). A few
+# (.config/nextest.toml), so they have a job of their own in four shards (GPU path tests). A few
# heavy tests (see .config/nextest.toml) are skipped on pull requests and run on pushes to main and
# nightly. The reference tests and the doc tests run in the merge queue, on main and nightly, not
# on pull requests (the fast lane, in Select checks); a break there is fixed forward on main. The
@@ -32,7 +32,7 @@
# reads `ci` from the newest run of a commit, so a run without it would block a PR that passed.
# Cross-compilation is opt-in through Run workflow; release builds remain in release.yml.
# Every Linux job runs on Blacksmith's Ubuntu 24.04 runners (blacksmith-<N>vcpu-ubuntu-2404), sized
-# per job: 32 vCPU for the tests, 16 for Clippy, the archive build (nightly) and the merge queue's, 8 for compiles, the GPU path tests, doc and reference tests, 4 for the light checks, 2 for Select checks and ci.
+# per job: 32 vCPU for the tests, 16 for Clippy, the GPU path tests, the archive build (nightly) and the merge queue's, 8 for compiles, doc and reference tests, 4 for the light checks, 2 for Select checks and ci.
# The Windows shader job runs on Blacksmith's Windows Server 2025 runners (blacksmith-8vcpu-windows-2025);
# the Windows app startup smoke stays on windows-latest, which has a desktop session that presents frames.
# Pull requests, forks included, the merge queue, pushes to main and the nightly run all use them:
@@ -545,20 +545,22 @@ jobs:
retention-days: 7
gpu-path-tests:
- name: GPU path tests (${{ matrix.shard }}/3)
+ name: GPU path tests (${{ matrix.shard }}/4)
# The GPU path tracer's tests (about 108) compile their pipelines on lavapipe, one test at a
# time (the gpu-path test group in .config/nextest.toml keeps them within the 600 s pipeline
# setup budget). Serial, they took about 36 minutes, the longest chain in Native Linux tests;
- # here three shards run beside it. Their filter is the one Native Linux tests leaves out, so
+ # here four shards run beside it. Their filter is the one Native Linux tests leaves out, so
# every test runs exactly once.
+ # 16 vCPU, not 8: lavapipe compiles on every core, and on 8 vCPU one test went past the 600 s
+ # budget (665 s, against 204 s for the same test in another run) and a shard took 35 minutes.
needs: changes
if: ${{ !cancelled() && needs.changes.result == 'success' && needs.changes.outputs.gpu == 'true' }}
- runs-on: blacksmith-8vcpu-ubuntu-2404
+ runs-on: blacksmith-16vcpu-ubuntu-2404
timeout-minutes: 45
strategy:
fail-fast: false
matrix:
- shard: [1, 2, 3]
+ shard: [1, 2, 3, 4]
env:
# As in Native Linux tests: lavapipe draws, and a test without an adapter fails.
SOLIDCRAFT_GPU: software
@@ -604,9 +606,10 @@ jobs:
SHARD: ${{ matrix.shard }}
run: |
# None of these is in the default filter (the heavy tests), so pull requests, the merge
- # queue, main and nightly all run every one of them, as before.
+ # queue, main and nightly all run every one of them, as before. `count` deals the tests
+ # out in turn (27 each); `hash` split them 41/40/27 across three shards.
cargo nextest run -p solidcraft-gpu --cargo-profile ci --locked --profile ci --no-tests=pass \
- --ignore-default-filter -E 'test(/^path::/)' --partition "hash:$SHARD/3"
+ --ignore-default-filter -E 'test(/^path::/)' --partition "count:$SHARD/4"
- name: Compiler cache statistics
if: always()
run: sccache --show-stats