Repository navigation
bump-hrx: hold hrx-system by config and handle AMD's history rewrite - #391
Conversation
The scheduled run failed every day from 2026-10-09 at the mirror step: AMD-Ecosystem/llama.cpp squashed 141 commits into 53d443cd0 (same tree as their old tip 8108322b7), so our mirror 1bit/hrx-vulkan (02f2880c) is no longer an ancestor of AMD's pin d9e2dc49. None of those commits are ours. The fork side moves to scripts/bump-hrx-sync.sh, which runs locally with DRY_RUN=1 and pushes nothing: - Mirror: when AMD rewrote its history, check that every mirror commit is still on an AMD branch (fail and list them if not), tag the old tip hrx-vulkan-<sha>, and move the mirror with a lease. The merge treats the squash commit as a child of its pre-squash equivalent, so only AMD's new changes are merged (conflicts drop from 108 files, 3 outside ggml-hrx, to 114 files all inside ggml-hrx). - Hold: config/hrx-system-hold.config names 98d05d94. While it does, hrx-system is not merged and only AMD's llama.cpp changes outside the HRX backend paths are taken, as a plain commit, so AMD's HRX changes still arrive in full when the hold is lifted. AMD's d9e2dc49 changes only HRX files (#139 moves the kernel corpus to multiple_of, which 98d05d94 cannot parse), so today nothing is merged and the run reports that. Conflicts outside ggml-hrx, and any hrx-system conflict when not held, still fail the run and name the files. Fixes #388 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PR Reviewer Guide 🔍Here are some key observations to aid the review process:
|
…rvable the engine currently lacks (#2468) Main pins 669178cb0, one commit behind the fork head. The missing commit is #391: aae78ecdf common/sampling: log whether the fast greedy path was taken, so eligibility drift is observable (#391) 669178cb0 hrx: one-token GDN decode updates the recurrent state in place (no 12.6 MB state copy per token) (#390) 6d6cb0f32 hrx: strict refuses a misplacement, not a placement the code calls by design (#389) Why this one commit matters beyond being current: #391 exists to be READ from a served binary's log. engine#2452's review asked for an engine-level greedy comparison and it cannot be written while the engine has no observable -- `git show 669178cb0:common/sampling.cpp | grep -c LLAMA_LOG_FAST_GREEDY` = 0 -- so the arm stays owed for a reason that is one line away from disappearing. Verified at this pin (engine build, submodule checked out at aae78ecd and the sha printed by the build): BUILD_RC=0 0 real compile errors 3 objects recompiled (llama-graph / sampling / relink) <build>/hrx/llama/bin/llama-server relinked registry: regenerated for the new pin, --check-pins rc 0. Not claimed: the engine-level fast-greedy arm itself. It becomes writable with this pin and is still owed on the next device-or-CPU serve; what this PR removes is the reason it could not be written at all. Co-authored-by: bcloud <agent@local>
Why
bump-hrx has failed on every scheduled run since 2026-10-09 at "Sync the fork and merge AMD's pin into our line" with "1bit/hrx-vulkan has commits that AMD's pin does not contain".
Those commits are not ours. AMD-Ecosystem/llama.cpp rewrote its history:
53d443cd0"ggml-hrx: squash development through 8108322b7" replaces 141 commits and has the same tree as their old tip8108322b7. Our mirror1bit/hrx-vulkansits at their previous pin02f2880c, which is now on their branchhrx-graph-develop-v2--backup-for-squash-onto-33a1f2b23but not in the new pind9e2dc49.Separately, the owner keeps the engine on our HRX line (2026-10-10): hrx-system stays at
98d05d94because AMD's newer hrx-system miscompiles long-context attention (ROCm/hrx-system#1535). The workflow had no way to express that hold.What
scripts/bump-hrx-sync.sh(new) does the fork side. The workflow calls it; it also runs locally withDRY_RUN=1and prints pushes instead of doing them.hrx-vulkan-<sha>and the mirror moves with--force-with-lease. For the merge, the squash commit is grafted (locally) onto its pre-squash equivalent, found by tree equality, so only AMD's new changes are merged. Without that, mergingd9e2dc49into our line conflicts in 108 files, 3 of them outside ggml-hrx; with it, 114 files, all inside ggml-hrx.config/hrx-system-hold.config(new) names98d05d94. While it names a commit, hrx-system is not merged,third_party/hrx-systemdoes not move, and only AMD's llama.cpp changes outside the HRX backend paths (ggml/src/ggml-hrx,ggml/include/ggml-hrx.h,tests/*hrx*) are taken, as a plain commit on1bit/hrx-vulkan-patched. AMD is not a merge parent there, so its HRX changes still arrive in full after the hold is removed. AMD's ggml-hrx is written against AMD's hrx-system:d9e2dc49(serve --mmproj (images in chat); Zyphra Zamba2-VL on Vulkan (pin llama.cpp-vulkan 2c9d3983) #139) moves 44 kernel-corpus files to themultiple_ofpredicate, which98d05d94cannot parse.Conflicts outside ggml-hrx, an unmappable history rewrite, and hrx-system conflicts when not held still fail the run and name the files or commits.
Checked
Dry runs against the real repositories (no token, nothing pushed):
Other paths, same data:
HOLD_FILE=no-hold): llama.cpp merges with 114 conflicts, all in ggml-hrx, resolved to ours, exit 0; hrx-system fails loudly withconflicts in: loom/src/loom/target/arch/amdgpu/planning/wait_plan.c, exit 1.hrx-integration): fails with "AMD rewrote its history in a way this job cannot map", exit 1.d9e2dc49+ one README line yields a tree that differs from our line only in README.md.yaml.safe_loadparses the workflow;tools/copyright.py --checkpasses.When this merges the next run will open one PR moving
third_party/llama.cppfrom0e255d77to2bdb04a0(fork #97, the loom-jit port already on our line) with hrx-system unchanged. That PR needs the Strix Halo checks before merging, as usual.Fixes #388
🤖 Generated with Claude Code