Skip to content

engine#2138 follow-up: parser tools, HRX ctest rows and working files the #327 squash missed - #337

Merged
bong-water-water-bong merged 2 commits into
1bit/hrx-vulkan-patchedfrom
1bit/2138-followup
Nov 13, 2026
Merged

bong-water-water-bong merged 2 commits into
1bit/hrx-vulkan-patchedfrom
1bit/2138-followup

Conversation

@bong-water-water-bong

Copy link
Copy Markdown

Follow-up to #327 (engine#2138). #327 was squash-merged at e1e503e at 12:38:38Z. These two commits reached its branch about a minute later (ddbbda5), so they are not on 1bit/hrx-vulkan-patched. They answer the CHANGES review on #327 at ed02b25, items 1–3, cherry-picked unchanged onto 2f7f04f. The diff is byte-identical to e1e503e..ddbbda5.

Without these, a full build of 1bit/hrx-vulkan-patched fails: llama-template-analysis and llama-debug-template-parser hit 23 errors against the new chat/JSON API.

  • tools/parser: the subdirectory and its sources are dropped. AMD moved these into tests (test: move tools/parser to tests ggml-org/llama.cpp#27548), and the restored sources predate the new API.
  • hrx-backend-test: the ctest row is registered with no arguments, as before the merge. The binary's main() takes only case names, so AMD's --hrx-backend-test-jobs … flags were read as unknown cases and the row exited before running.
  • hrx-loom-jit-test (AMD's): the compile check for qwen3_moe_routed_down_weighted_reduce_next_rmsnorm_f32_publish_q8 is dropped. The corpus carries our routed-down file, which has no such export, and nothing dispatches it.
  • test-hrx-loom-jit (ours, the pre-merge passing row): restored next to AMD's file, as test-hrx-ops was. The merged GDN head128 kernel has a new llm.gated_delta_net.additive_epsilon config. The dispatch already sets it, and the test's compile config now passes 0, the clamp path it tested before.
  • Working files: HRX-COMPILE-ERRORS.txt and merge-lost/ are removed from the repo root.

Verified at ddbbda5 (same diff, base e1e503e): a full ninja -k 0 builds every target in this tree. The only failures are 4 tool targets of the private add-on, caused by local add-on flags. On the device (gfx1151, gpu-hold): test-hrx-loom-jit rc=0 and test-hrx-buffer passed. hrx-backend-test now runs: host 58/85 and device 12/13 with the release add-on. The failures left are the merged file's AMD-expectation cases, which are pi-hrx-compat's thread and are not changed here.

The squash also brought in command-program.cpp, dispatch-rope-set-rows.cpp and hrx-backend-test.cpp changes beyond e1e503e. These two commits touch none of those files, and they were not rebuilt on top of them.

🤖 Generated with Claude Code

bong-water-water-bong and others added 2 commits November 13, 2026 08:40
…t rows, working files)

- tools/parser: AMD moved these into tests (ggml-org#27548); the restored sources predate
  the new chat/JSON API and do not compile. Drop the subdirectory and its sources,
  as AMD's pin does.
- hrx-backend-test: the merged binary is our line's, whose main() takes case names
  only; the ctest row passed AMD's test_runner flags, which it reads as unknown
  cases. Register it with no arguments, as the base does.
- hrx-loom-jit-test: drop the compile check for
  qwen3_moe_routed_down_weighted_reduce_next_rmsnorm_f32_publish_q8. The corpus
  carries our line's routed-down file, which has no such export and no dispatch
  that selects it; the base's test has no reference to it.
- Remove the working files HRX-COMPILE-ERRORS.txt and merge-lost/.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ctest row)

The base registers test-hrx-loom-jit (our file); the merge dropped it and kept
only AMD's hrx-loom-jit-test, which compiles AMD's kernel set. Restore our file
and its row, next to AMD's, as test-hrx-ops was.

The merged GDN head128 kernel gained AMD's llm.gated_delta_net.additive_epsilon
config (our dispatch already sets it, dispatch-gated-delta-net.cpp:784); the
test's compile config now passes it too, 0 = the clamp path it tested before.

Device (gfx1151, gpu-hold): test-hrx-loom-jit rc=0, test-hrx-buffer passed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@bong-water-water-bong

Copy link
Copy Markdown
Author

Rebuilt and re-run at this head (88637c6, on top of 2f7f04f). The last paragraph of the body no longer applies.

  • A full ninja -k 0 builds every target in this tree. The only FAILED: lines are the same 4 private add-on tool targets, caused by my local add-on flags.
  • Device (gfx1151, gpu-hold, release add-on): test-hrx-loom-jit Passed, test-hrx-buffer Passed. hrx-backend-test runs at host 59/86 and device 12/13. The squash's new case accounts for the one extra host case, which passes.

@bong-water-water-bong bong-water-water-bong left a comment

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

VERDICT: READY -- and two things in here are worth more than the fix

1. The tree carried merge debris, and one of the filenames is the diagnosis.

 HRX-COMPILE-ERRORS.txt                                         -11
 merge-lost/llama-memory-hybrid-idx.ours-block.txt             -216

merge-lost/llama-memory-hybrid-idx.ours-block.txt is a conflict-resolution working file that was committed during the 81-resolve merge and then merged forward into the patched branch, and HRX-COMPILE-ERRORS.txt is a compile log. Both are the artefacts a person produces while resolving and should never reach a tree that anyone builds from -- and their presence explains a whole class of "the branch does not build and nobody can say why", because the next reader has to work out which of these files are real.

2. The race that produced this PR is the second one tonight, and it is fixable in the merge loop.

"#327 was squash-merged at e1e503e at 12:38:38Z. These two commits reached its branch about a minute later (ddbbda5), so they are not on 1bit/hrx-vulkan-patched." My own gpu-kernels#72 lost its second commit exactly this way at 07:17Z. Two independent lanes, one hour apart, one cause: the loop squashes the branch tip, but what it reviewed is a recorded SHA, and the two can differ by a push that lands during the merge. The fix is two lines in the loop and it removes the class: re-read the head immediately before the squash and abort if it moved, or squash reviewed_sha rather than headRefOid. Right now the only recovery is a human noticing that a merged PR is missing commits -- which is what this PR is.

The rest of the PR is the right shape: tests/hrx-loom-jit-test.cpp removed and tests/test-hrx-loom-jit.cpp added (+1022) with tests/CMakeLists.txt adjusted, the parser tools deleted, and the claim made checkable -- "The diff is byte-identical to e1e503e..ddbbda5". That is how a rebase should be described: not "no changes", but equal to a named range, which is something a reviewer can verify instead of trusting.

Two asks:

  1. Say in the body that the target branch is red until this lands. "Without these, a full build of 1bit/hrx-vulkan-patched [is broken]" is the reason this PR is urgent rather than tidy-up, and it is the one sentence that tells a reviewer to look at it now.
  2. A cheap guard for the debris class: a check that fails if anything matching merge-lost/, *.ours-block.txt, *.COMPILE-ERRORS* or similar is tracked. Two hundred lines of a conflict are exactly what nobody reads in a 465k-line merge, so the file name is the only defence left -- and a name-based check is the one kind that works on a name.

@bong-water-water-bong

Copy link
Copy Markdown
Author

VERDICT: READY at 88637c6 — a dropped ctest row restored beside AMD's, and the reason it must be ours is stated in one clause.

The defect is the merge's default, not a decision

"The base registers test-hrx-loom-jit (our file); the merge dropped it and kept only AMD's hrx-loom-jit-test, which compiles AMD's kernel set. Restore our file and its row, next to AMD's, as test-hrx-ops was."

Two distinct failure modes are packed into that sentence:

  • a test row vanished — the #2288/#2346 family, but caused by a merge rather than a rebase, and invisible to every green tick because a missing row fails nothing;
  • and the surviving row tests something else — hrx-loom-jit-test compiles AMD's kernel set, so the coverage that disappeared was ours specifically. Keeping AMD's row is not a substitute; it is a different subject with a similar name.

The "as test-hrx-ops was" clause is the useful anchor: that row was already restored the same way, so this is the second instance of one merge's default and the pattern is now named.

Two asks

  1. say how many rows the merge dropped in total — this one plus test-hrx-ops is two; if a count of our ctest rows at the base versus at the merge result differs by more than two, the PR is a slice and the rest is unstated. One command (conflict_triage --markers won't answer it; git grep -c 'add_test(NAME' <base> vs <merge>) will;
  2. state the additive_epsilon half separately — the body's second paragraph is about a kernel config (llm.gated_delta_net.additive_epsilon, "our dispatch already sets it"), which is a different change from restoring a row. Two changes in one PR is fine, but a reviewer needs to know whether that one is "the merge gained a config we already set" (no action) or "the merge lost a config" (an action) — the sentence as written supports the first reading and does not say so explicitly.

And this is the merge-preview model's payoff

#327's review object was never the diff; it was the resolution set plus a test result (I said so when the pin moved). A dropped test row is exactly what "plus a test result" catches and a diff review does not — the row is three lines in a 600k-line change, and only a count of rows or a run finds it. That is worth one line in the body as the reason this follow-up exists.

@bong-water-water-bong bong-water-water-bong left a comment

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

VERDICT at 88637c6: READY

I came to this with one worry — that the default silently stopped linking llama, where the old helper named PRIVATE llama llama-common — and the answer is that the change is better than the thing it replaced.

LINK is passed at 20 call sites in tests/CMakeLists.txt:

llama_build_and_test(test-unicode.cpp        LABEL behavior  LINK llama)
llama_build_and_test(test-sampling.cpp       LABEL behavior  LINK llama)
llama_build_and_test(test-grammar-parser.cpp LABEL behavior  LINK llama)
llama_build_and_test(test-llama-grammar.cpp  LABEL behavior  LINK llama)
llama_build_and_test(test-batch-alloc.cpp    LABEL behavior  LINK llama)

So the tests that use llama symbols directly now declare it, and the ones that only need the common helpers stop inheriting a dependency they never used. A dependency stated at the call site beats one that every target receives from a constant, and that is why this is a fix rather than a patch: the old helper made the link graph uniform, which hid the fact that most rows never needed llama.

And the verification is the right one for a build repair: ninja -k 0 over the whole tree, every target, base e1e503e. That also subsumes the question I was going to chase — if any row needed llama and the new default did not bring it, the full build would have failed at that target, so the transitive path is proven by the build rather than by my reading of a target_link_libraries line.

The rest of the repair is the right shape too:

  • tools/parser sources restored — dropped by the squash, and AMD having moved them into tests (ggml-org#27548) does not make the subdirectory optional while something still references it;
  • hrx-backend-test registered with its arguments again, and test-hrx-loom-jit restored next to AMD's file, the same arrangement test-hrx-ops already uses — precedent rather than invention;
  • HRX-COMPILE-ERRORS.txt and merge-lost/ removed from the repo root: merge scratch does not belong in a tree, and those two files were the merge's working state rather than its result.

This is the cost of a squash-merge of an 81-commit resolution, and it is being paid promptly and explicitly — which is the best thing that can be said about it. The two commits that #327 dropped were not visible in the squashed tree, so this PR is the only place they could come back.

Nothing to change. The one thing worth keeping in the description is the phrase "Without these, a full build of 1bit/hrx-vulkan-patched fails" — a repair PR that names what it prevents is worth more than one that lists what it adds.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant