Prebuilds: accept unslothai/llama.cpp PRs in pr-set.json#30
Conversation
|
@codex review |
|
/gemini review |
|
Codex Review: Didn't find any major issues. Breezy! Reviewed commit: ℹ️ About Codex in GitHubCodex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback". |
|
/gemini review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5eed57e071
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
Expands
scripts/unsloth/pr-set.jsonto accept pins from unslothai/llama.cpp PRs, not just ggml-org/llama.cpp. This lets fork-only changes ship in the nightly prebuilds without waiting for an upstream PR.How it works
https://github.com/{ggml-org,unslothai}/llama.cpp/pull/<n>/commits/<sha>.b####tag.repofield, and the-mix-<hash>tag now hashesrepo#number:sha, so the same PR number in the two repos can't collide. Side effect: the first nightly after this lands rebuilds the current pin set once under a new tag, same content.unslothai/llama.cpp#16) and keep the plain#24423form for upstream ones.Hardenings and docs
gh: Not Found. The lint reports it and keeps checking the remaining entries.Verification
Simulated the full resolve flow locally on base b9902. The mixed set (the 3 current ggml-org pins plus #16) merges cleanly end to end, with correctly labeled merge commits and intact build-info stamping; #11 and #15 conflict with the current base and abort with the reorder-or-drop message as intended. Exercised the URL validation against real PRs from both repos plus malformed and wrong-repo URLs. Also exercised the 404 paths in both workflows against the live API. Nothing downstream consumes the manifest's
merged_prs, so the addedrepofield is additive.