Release built artifacts on partial failure + fix CUDA/SYCL bridge compile

[SnowCheetos]

Overview

Two independent changes, both aimed at the CUDA/SYCL source builds that keep breaking full releases. The v0.17.4 release (run 28914476140) built CPU/Vulkan/macOS in ~2.5 min each, then CUDA/SYCL ran the full ~15-min build and failed at the very end — and because the release job required all legs, the three good binaries were discarded.

1. ci(release) — publish what built, even if some targets fail

The release job used needs: build with no if:, so one failed matrix leg skipped the whole release. Now it runs with if: ${{ !cancelled() }}:

• Publishes whatever artifacts were uploaded (failed legs upload nothing → simply absent from the release).
• Errors out only if zero targets built, so we never cut an empty release.
• The overall run still goes red when a target fails — failures stay visible, they just no longer block the working binaries.

This is the direct fix for "the long build always gets ruined by the 2 failed backends."

2. fix(build) — the actual CUDA/SYCL failure (not OOM/timeout)

The failure looked like it might be a resource issue, but it was another API mismatch surfacing only after the full build:

autocommit_common_bridge.cpp:662: error: 'llama_params_fit' was not declared in this scope

The bridge guarded llama_params_fit behind #ifdef LLAMA_CPP_PREBUILT — no-oping it for the binary release but still calling it in the #else from-source branch. Since #42 pinned the submodule to b9837, the source build is also b9837, which doesn't expose that public symbol in either mode (the fitting logic moved into libcommon's internal common/fit.cpp). Fix: no-op unconditionally, matching the prebuilt backends that already ship (they've never run the fit either).

Verification

• release.yml parses as valid YAML.
• The bridge now compiles against a real b9837 checkout in source mode (no -DLLAMA_CPP_PREBUILT — exactly what CUDA/SYCL do), and still compiles in prebuilt mode. This was the precise step that failed at line 662.

Caveat: I can't run the CUDA/SYCL toolchains locally, so this fixes the confirmed compile error but I can't rule out a further issue at their link stage. That's exactly what change #1 now protects against — worst case, CPU/Vulkan/macOS still release.

After merge

Cut a v0.17.5 tag to exercise the full matrix. Expected: all five publish; if CUDA/SYCL still trip on something downstream, the other three release anyway.

[Copilot]

Pull request overview

This PR aims to make releases more resilient to partial build failures in the CI matrix and to fix a CUDA/SYCL source-build compilation break in the llama.cpp bridge code.

Changes:

• Update the Release workflow so the release job runs even when some build matrix legs fail, publishing whatever artifacts are available and failing only when none exist.
• Remove/disable usage of llama_params_fit in the CUDA/SYCL bridge to match the pinned llama.cpp release API surface and avoid compile errors.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
crates/llama-sys/src/autocommit_common_bridge.cpp	No-op autocommit_llama_params_fit unconditionally to avoid calling a missing symbol in source builds.
.github/workflows/release.yml	Make the release job tolerant of partial build failures and gate publishing on whether artifacts exist.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.