Repository navigation
pr-set: pin Beam (ggml-org#30670), repin glm5-next compute buffer, drop pins merged upstream; fix preflight clone - #282
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a5143b3f6c
ℹ️ About Codex in GitHub
Your team has set up Codex to 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 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| "https://github.com/ggml-org/llama.cpp/pull/30520/commits/e64207bab6c2989a4ae5aa8b9e0cbe700e135ee3", | ||
| "https://github.com/ggml-org/llama.cpp/pull/30616/commits/9bbf448097779acb72cb5084c1a92cde5ff19323" | ||
| "https://github.com/ggml-org/llama.cpp/pull/30616/commits/9bbf448097779acb72cb5084c1a92cde5ff19323", | ||
| "https://github.com/unslothai/llama.cpp/pull/0/commits/2dfaee3be5fa8c01c417faeae1c6cc461f0ff32d" |
There was a problem hiding this comment.
This entry points to unslothai#0 at 2dfaee3..., although the reviewed pin described by this commit is unslothai#281 at cfb9f2cec.... When the manifest is consumed, the resolver in .github/workflows/unsloth-prebuilt.yml calls gh api repos/unslothai/llama.cpp/pulls/0 and exits if that nonexistent PR cannot be fetched, so every scheduled mixed prebuild aborts before merging or building; the lint workflow fails for the same reason. Update both this URL and the matching feature owner to the real PR.
Useful? React with 👍 / 👎.
| ] | ||
| }, | ||
| "muse-spark": { "owner": "ggml-org#30616", "checks": [ { "kind": "arch", "arch": "muse-spark" }, { "kind": "mtmd", "projector": "muse-spark-audio" } ] }, | ||
| "beam": { "owner": "unslothai#0", "checks": [ { "kind": "arch", "arch": "beam" } ] }, |
There was a problem hiding this comment.
Even after correcting the pin identity, this sole architecture probe does not exercise the PR's new Jinja list-on-objects behavior. The preflight builds only test-llama-archs, test-backend-ops, and test-mtmd-impl before running this matrix, and no workflow runs the existing test-jinja target, so a merge regression in that independently shipped behavior can leave the feature gate green. Add a Jinja probe that builds and executes test-jinja (or otherwise record and test this part of the pin).
Useful? React with 👍 / 👎.
…s merged upstream ggml-org#30520 (Nemotron 3.5) and ggml-org#30616 (Muse Spark) are merged upstream and in every aged base the nightly can pick, so their pins are dropped and their feature checks keep running with no owner. ggml-org#30616 applied a second time conflicted on master. #243 is repinned to its merge of b11680, which resolves the conflict with ggml-org#30613. #281 is repinned to its current head.
danielhanchen
changed the title
ggml-org#30670 is the upstream PR for the same change (one squashed commit, identical changed lines).
With --filter=blob:none, git 2.55 on the runners fetches missing blobs during the merge, GitHub refuses them, the merge exits 128, and the silenced merge --abort then ends the step with no message.
danielhanchen
changed the title
Pin set update for the prebuilt (aged upstream
b####plus the pins inpr-set.json, merged in order with diff3 andadditive_merge.pyfor pure add/add conflicts).Changes
00d962b64feature-checks.jsongets abeamentry owned by ggml-org#30670 (test-llama-archs -a beam)f427abb1cto43521db75src/models/glm5-next.cppon every tag from b11680, after upstream ggml-org#30613 (share the k-pool input views between layers). Fixed in the pin branch with a merge of b11680: upstream's text kept, only the chunked re-pooling left in the pin (+34/-12). Old vs new pin: identical perplexity and greedy output on the generated glm5-next test modelThe dropped pins' feature checks stay, with
owner: null, since the features now arrive through the base tag.Evidence
llama_pinset.py sim --tags window(b11680 aged through b11696 newest, 11 tags): all 18 pins merge on every tag,pin_contract18/18 intact,merge_checksclean; nothing is refused. Rerun with ggml-org#30670 in place of #281 on b11681 and b11696: same result (ggml-org#30670 additive).llama_pinset.py audit: 18/18 pins clean.llama_pinset.py verify --cuda(sm_100), final pin set with ggml-org#30670:Real weights: the Beam UD-IQ1_S GGUF loads and generates on CUDA with the same change (#281 build).
Preflight fix
unsloth-pin-preflight.ymlcloned with--filter=blob:none. With git 2.55 on the runners a merge then fetches missing blobs, GitHub refuses them, the merge exits 128, and the silencedgit merge --abortended the step with no message (master's own preflight failed this way). It now does a full clone and tolerates a failed abort.Not covered here: ROCm, Vulkan, macOS, Windows and ARM legs (nightly only).