Repository navigation
ci(bump-hrx): merge AMD's pin into our line instead of rebasing 159 commits (fixes the 5-day silent sync failure) - #328
Merged
Conversation
added 2 commits
October 6, 2026 04:35
…ommits
bump-hrx failed on every scheduled run from 2026-10-01 to 2026-10-05 at the
"Sync the fork and rebase our commits" step, with "Open the bump PR" skipped.
The token step passes, so it was never auth.
Measured cause: our 1bit/hrx-vulkan-patched carries 159 commits that are
genuinely ours (`git cherry` against AMD's pin reports 0 already upstream), and
AMD's pin had moved 70 commits over the same files. Rebasing replays all 159 and
died at commit 1 of 159:
Rebase (1/159) Auto-merging .../dispatch-mul-mat.cpp
CONFLICT (content): ... could not apply d7b4d78f7 (IQ3_XXS matvec kernel)
Merging the same two lines conflicts in 15 files, all under ggml/src/ggml-hrx/,
and resolves once. So the premise ("rebase our commits onto AMD's pin") is what
was wrong, not the plumbing.
Also: a failing scheduled job that tells no one is the real defect here — the
HRX pin silently froze for five days. A failing run now files or updates an
issue naming the conflicting files.
The rebase path's ancestry guard, the tag-before-move safety, and the
force-with-lease push are unchanged, so the commit the engine pins stays
reachable exactly as before.
AMD's pins come as a PAIR, but this workflow moved only llama.cpp, and .gitmodules records hrx-system as "rebased by hand". So every AMD move left hrx-system silently behind — the same class of silent drift that let the llama.cpp pin freeze for five days. Adds a step that merges AMD's hrx-system pin into 1bit-MONSTER/hrx-system 1bit/main with the same merge-not-rebase treatment, tags the previous tip before moving the branch, and force-with-lease pushes. On conflict it names the conflicting files and fails. Also reads the branch into an explicit ref rather than FETCH_HEAD: after two fetches a bare FETCH_HEAD silently compares the wrong pair — that bug cost a debugging cycle on the llama.cpp half. The bump PR now moves BOTH gitlinks instead of naming hrx-system as a manual follow-up.
This was referenced Oct 6, 2026
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What was wrong
bump-hrxfailed on every scheduled run from 2026-10-01 to 2026-10-05, always at the "Sync the fork and rebase our commits" step, with "Open the bump PR" skipped. Step 2 "Require the token" passes, so this was never auth.Measured cause. Our
1bit/hrx-vulkan-patchedcarries 159 commits that are genuinely ours —git cherryagainst AMD's pin reports 0 already upstream — and AMD's pin had moved 70 commits over the same files. Rebasing replays all 159 and died at commit 1 of 159:The same two lines merge with only 15 conflicting files, all under
ggml/src/ggml-hrx/— so the premise ("rebase our commits onto AMD's pin") is what's wrong, not the plumbing.What this changes
git merge --no-edit --no-ffrather thanrebase --onto, so one bounded resolution replaces a 159-step replay.1bit/hrx-vulkan, the tag created before the branch moves (so the commit the engine pins stays reachable), and the--force-with-leasepush.Remaining step, stated plainly
This makes the automation correct, not the divergence resolved. The next run will merge and still hit those 15 files once, because AMD has changed our own HRX files (including
flash_attention_decode_split_f32_f16_wmma.loom, the kernel from #314). Resolving them is a one-off manual job that needs HRX validation on Strix Halo — it is not on the release-critical path, because the release needs only the narrow #315 pin bump, which does not move to AMD's pin at all.