Skip to content

fix(release): pin Windows CUDA to sccache-compatible version - #606

Merged
michaelneale merged 1 commit into
mainfrom
micn/release-pin-windows-cuda
May 20, 2026
Merged

fix(release): pin Windows CUDA to sccache-compatible version#606
michaelneale merged 1 commit into
mainfrom
micn/release-pin-windows-cuda

Conversation

@michaelneale

Copy link
Copy Markdown
Collaborator

What

The Windows CUDA release lane now pins CUDA via Jimver/cuda-toolkit@v0.2.35 to vars.CUDA_VERSION || '12.6.3' (currently 12.8.0 in repo variables), matching the CI lane. The previous choco install cuda path in scripts/install-windows-sdk.ps1 pulled the latest CUDA Toolkit (13.2) which is incompatible with sccache 0.15.0.

The rocm and vulkan branches of install-windows-sdk.ps1 are unchanged — only the cuda matrix entry is rewired.

Why

Both v0.66.0-rc5 and v0.66.0-rc6 release runs failed on Build Windows CUDA with:

sccache: error: failed to execute compile
sccache: caused by: error reading compile response from server
sccache: caused by: An existing connection was forcibly closed by the remote host. (os error 10054)
ninja: build stopped: subcommand failed.

This is the documented incompatibility in mozilla/sccache#2470 (CUDA + MSVC + Ninja Multi-Config). sccache 0.15.0 cannot parse the shell line nvcc 13.x emits on Windows. The crash is on the same file (argsort.cu) every run.

PR #604 made the in-script retry actually run by tolerating a dead sccache server during cleanup. That fix is correct on its own — without it, the retry path crashed before the second cmake invocation. But with #604 merged, the retry still hits the same deterministic crash because the underlying nvcc/sccache incompatibility is unfixable in the script.

The CI Windows CUDA lane (.github/workflows/ci.yml:864) has been working throughout because it already uses Jimver/cuda-toolkit@v0.2.35 with WINDOWS_CUDA_VERSION: ${{ vars.CUDA_VERSION || '12.6.3' }}. PR #604's PR CI Windows CUDA succeeded against the same code that failed in the release lane — only difference was the CUDA version. This PR mirrors the CI setup in release.

Risk

  • Adds Jimver/cuda-toolkit@v0.2.35 as a new dep in release.yml. Same action+SHA CI already uses, so no new surface area.
  • The Cache Windows SDK installers step now skips on matrix.backend == 'cuda'. Cache there was tied to install-windows-sdk.ps1 hash, and that script is no longer called for cuda — caching is now handled by Jimver's use-github-cache: true + use-local-cache: true instead.
  • ROCm and Vulkan paths unchanged.

Validation

The PR's own CI will exercise the Windows CUDA lane on this change. If that goes green, retry v0.66.0-rc6 (or cut rc7) — should produce the missing Windows CUDA bundle and unblock the release publish gate.

The Windows CUDA release lane has been failing deterministically because
`scripts/install-windows-sdk.ps1` does `choco install cuda` which pulls
the latest CUDA Toolkit (currently 13.2). sccache 0.15.0 cannot parse
nvcc 13.2 output and crashes mid-build with:

  sccache: error: failed to execute compile
  sccache: caused by: An existing connection was forcibly closed by the
    remote host. (os error 10054)

This is the documented incompatibility in mozilla/sccache#2470 (CUDA +
MSVC + Ninja). PR #604 made the in-script retry tolerate a dead sccache
server, which let the retry actually run — but the retry hits the same
deterministic crash on the same file (argsort.cu), so the workflow
still failed on rc6.

CI's Windows CUDA lane has been working because it pins CUDA via
`Jimver/cuda-toolkit@v0.2.35` to `vars.CUDA_VERSION || '12.6.3'`
(currently 12.8.0 in repo vars). Mirror that setup here for the
release lane: install CUDA via the same pinned action for the cuda
matrix entry, and keep `install-windows-sdk.ps1` for rocm/vulkan
where it works correctly.

Validated by:
- PR #604 PR CI Windows CUDA succeeded with the same Jimver+12.8.0 path
- Release rc5/rc6 Windows CUDA failed with choco+13.2 path

Other Windows lanes (CPU, ROCm, Vulkan) keep using
`install-windows-sdk.ps1` unchanged.
Copilot AI review requested due to automatic review settings May 20, 2026 09:35
@michaelneale
michaelneale merged commit c251f1d into main May 20, 2026
13 checks passed
@michaelneale
michaelneale deleted the micn/release-pin-windows-cuda branch May 20, 2026 09:37

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR aligns the Windows CUDA release workflow with the existing CI approach by pinning CUDA to a sccache-compatible version (via Jimver/cuda-toolkit@v0.2.35) instead of installing the latest CUDA via Chocolatey, preventing deterministic sccache crashes with newer nvcc on Windows.

Changes:

  • Add a WINDOWS_CUDA_VERSION env var in the Windows GPU release job, defaulting to ${{ vars.CUDA_VERSION || '12.6.3' }}.
  • Replace the CUDA branch of Windows SDK installation in release with Jimver/cuda-toolkit@v0.2.35 and a dedicated CUDA verification step.
  • Skip the Windows SDK installer cache + install-windows-sdk.ps1 step for the CUDA matrix entry (ROCm/Vulkan unchanged).

run: |
if (-not $env:CUDA_PATH -or -not (Test-Path $env:CUDA_PATH)) {
throw "CUDA_PATH was not configured by Jimver/cuda-toolkit."
}
michaelneale added a commit that referenced this pull request May 21, 2026
Brings in #579 advisory capacity, #583 hardened materialization cache,
#562 version bump, #606/#604 Windows CUDA build fixes, #608 lint rule,
and #560 UI mockup.

* origin/main:
  fix(ci): small update for lint rule (#608)
  mockup: Reserves high-fidelity UI mockup (#560)
  Add advisory capacity evaluation for model targets (#579)
  Harden Skippy layer package materialization cache (#583)
  fix(release): pin Windows CUDA to sccache-compatible version (#606)
  fix(build-windows): tolerate dead sccache server in CUDA retry path (#604)
  chore(version): synchronize version bump everywhere (#562)
michaelneale added a commit that referenced this pull request May 21, 2026
* main:
  docs(AGENTS): add confidence-testing recipe for routing/MoA/gossip changes (#613)
  ci(sdk-smoke): install lld in macOS swift smoke job (#610)
  MoA: mesh mode and many inference critical fixes, and quic keep alive (#566)
  fix(ci): small update for lint rule (#608)
  mockup: Reserves high-fidelity UI mockup (#560)
  Add advisory capacity evaluation for model targets (#579)
  Harden Skippy layer package materialization cache (#583)
  fix(release): pin Windows CUDA to sccache-compatible version (#606)
  fix(build-windows): tolerate dead sccache server in CUDA retry path (#604)
  chore(version): synchronize version bump everywhere (#562)
  fix(mesh): skip filtered peers in gossip dial loop to unwedge `--auto` (#602)
  docs(agents): clarify just build vs release-build for serious testing (#599)
  build: ozempic — slim binary -42 MB / -47 MB (#592)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants