chore: ⬆️ Update mudler/vllm.cpp to a42b8187caff02c570c28e19e4dc2b1d7f55ed14

[localai-bot]

Changes: https://github.com/mudler/vllm.cpp/compare/9e1c9025ae61167a3335454d7cc0de6093c21845..a42b8187caff02c570c28e19e4dc2b1d7f55ed14

[mudler]

@localai-org-maint-bot update the PR and bump the vllm.cpp pin, fix CI issues

[localai-org-maint-bot]

Updated and pushed at 4d6b45807. The PR is rebased onto latest master and now pins the current vllm.cpp HEAD, 74d7bc803c51f2b7b8ea91a7e57f8b2745b3d9e9, which includes the intervening compiler-portability fixes. Local verification passed: make -C backend/go/vllm-cpp test (21 passed, 12 real-model specs skipped as designed), pin reachability, and git diff --check. Fresh CI has been triggered by the push.

[localai-org-maint-bot]

Resumed after the dependency bot advanced the branch to 502b84f44. The new pin ca5c7adc already contains merged vllm.cpp#24, but the fresh Darwin build still fails with the same two AppleClang -Wgnu-folding-constant diagnostics under -Werror (job 91543540465). The source-level pragma is therefore not controlling the compilation that emits these warnings. Since multiple flag-order/source-suppression variants have now failed, I am stopping rather than layering on another speculative override; the warning boundary needs to be redesigned at the MLX/external-header build boundary.

[localai-org-maint-bot]

Resumed and fixed the actual Darwin failure. The two go-m1cpu folding diagnostics were nonfatal; the fatal error was AppleClang promoting the unguarded LagunaMarlinMoeEnabled helper to -Werror,-Wunused-function in a Metal/non-Marlin build. Opened mudler/vllm.cpp#25 with the matching feature guard and regression test, rebased this PR onto current master, pinned 065373dcc, and force-pushed with lease at 615bbe89e.

Local verification: vllm.cpp device-leakage suite 25/25, DSR baseline unchanged at 32, checkpoint/agent-record checks green; LocalAI vllm-cpp suite 21 passed / 12 real-model specs skipped, packaging shell syntax and Metal/CUDA Make dry-runs pass. Fresh Darwin CI is now the binding AppleClang verification.

[localai-org-maint-bot]

Resumed after the completed Darwin log exposed the remaining root cause. The two go-m1cpu folding diagnostics were nonfatal; the actual failure was AppleClang rejecting Voxtral's GCC-only -Wstringop-overflow pragma under -Werror,-Wunknown-warning-option. Updated mudler/vllm.cpp#25 with a compiler-boundary regression and a GCC-only pragma guard, then advanced this PR to 6199acab and pushed LocalAI head 5c3478ce9.

Fresh verification: the upstream regression passed red/green, the Voxtral translation unit passes GCC -Wall -Wextra -Werror, vllm.cpp checkpoint/agent-record/device-leakage gates pass, and LocalAI pin reachability, packaging shell syntax, Metal+MLX Make dry-run, committed diff, and worktree cleanliness pass. LocalAI's Go suite could not be rerun in this worktree because proto generation requires unavailable host unzip; it passed on the preceding head and this commit only changes the vllm.cpp pin. Fresh CI is queued, with Darwin AppleClang as the binding platform verification.