Repository navigation
ci(ck): arch-prefix dispatcher lane result JSON so gfx942 and gfx950 stop overwriting each other - #14265
Merged
illsilin merged 1 commit intoOct 28, 2026
Conversation
DISPATCHER_PERF and DISPATCHER_CORRECTNESS run gfx942 and gfx950 as parallel branches of one build and archive the same build-relative *_results.json names. Jenkins keeps one artifact per path, so the last lane to archive silently replaced the other arch's results, and tile_engine JSON has no arch field to tell them apart. Add archiveArchResults(arch, glob), which renames each file to <arch>_<name> and archives the prefixed glob, and call it from both finally blocks. It never throws (returnStatus + allowEmptyArchive), so it cannot mask the lane's real failure. Each lane starts with deleteDir(), so no stale prefixed file can be renamed twice.
✅ All Checks Passed — Ready for Review
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
|
🎉 All checks passed! This PR is ready for review. |
illsilin
approved these changes
Oct 28, 2026
assistant-librarian Bot
pushed a commit
to ROCm/composable_kernel
that referenced
this pull request
Oct 28, 2026
ci(ck): arch-prefix dispatcher lane result JSON so gfx942 and gfx950 stop overwriting each other (#14265) ## Summary `DISPATCHER_PERF` and `DISPATCHER_CORRECTNESS` run gfx942 and gfx950 as parallel branches of the same build, and both archive `projects/composablekernel/build/*_results.json` under identical names. Jenkins keeps one artifact per path, so whichever lane archives last wins and the other arch's results are dropped every night (e.g. build #2866 has a single `gemm_universal_results.json`, and all `dispatcher_*` files are gfx950). The tile_engine JSON carries no arch field, so a consumer cannot even tell which GPU a surviving file came from. ## Change One helper in `vars/ck.groovy`, `archiveArchResults(arch, glob)`, renames each matching file to `<arch>_<name>` and archives the prefixed glob. Both `finally` blocks call it: - perf lane: `archiveArchResults(gpuTarget, "*_results.json")` - correctness lane: `archiveArchResults(arch, "dispatcher_*_results.json")` Archived names become e.g. `gfx942_gemm_universal_results.json` / `gfx950_dispatcher_gemm_results.json`, so both arches survive and the arch is in the name. The existing `*_results.json` glob still matches. The helper never throws (`returnStatus: true` + `allowEmptyArchive: true`), so a cmake/ninja failure before any JSON exists is still reported as the real error rather than masked by the archive step. Each lane starts with `deleteDir()`, so no stale prefixed file can be renamed twice. ## Consumers Anything keyed on the exact old names (e.g. the CK legacy perf dashboard ingest) needs to strip the `<arch>_` prefix, and can now key on it instead of guessing. ## Testing Ran the real `runDispatcherPerfTests` / `runDispatcherCorrectnessTests` from `ck.groovy` under Groovy 4 with only the Jenkins steps mocked: `buildAndTest` writes exactly the files named by `--json` in the real `execute_cmd`, `sh` executes for real, and `archiveArtifacts` stores into one map keyed by path (one artifact per path, like a Jenkins build). Four lanes (perf/correctness x gfx942/gfx950), each in its own workspace: | ck.groovy | archived artifacts | |---|---| | develop | 22, all from gfx950 (gfx942 lost) | | this PR | 42 = 20 gfx942 + 22 gfx950 (the 2 extra are the gfx950-only `mx_gemm` / `gemm_bquant`) | | this PR, `buildAndTest` throws | the real build error propagates; archive step does not throw | End-to-end check on Jenkins needs a run of this PR with `USE_CURRENT_BRANCH_FOR_CK_GROOVY=true` (otherwise the `ck` library loads from develop and the old archive step runs) plus `RUN_DISPATCHER_PERF_TESTS=true` and `RUN_DISPATCHER_CORRECTNESS_TESTS=true`. Expected: `gfx942_*_results.json` and `gfx950_*_results.json` both archived. JIRA: AICK-2692 Co-authored-by: Muhammed Emin Ozturk <3836908+ozturkosu@users.noreply.github.com>
Contributor
TheRock Submodule Bump ActivityNewest first
|
shumway
pushed a commit
to ROCm/composable_kernel
that referenced
this pull request
Nov 4, 2026
ci(ck): arch-prefix dispatcher lane result JSON so gfx942 and gfx950 stop overwriting each other (#14265) ## Summary `DISPATCHER_PERF` and `DISPATCHER_CORRECTNESS` run gfx942 and gfx950 as parallel branches of the same build, and both archive `projects/composablekernel/build/*_results.json` under identical names. Jenkins keeps one artifact per path, so whichever lane archives last wins and the other arch's results are dropped every night (e.g. build #2866 has a single `gemm_universal_results.json`, and all `dispatcher_*` files are gfx950). The tile_engine JSON carries no arch field, so a consumer cannot even tell which GPU a surviving file came from. ## Change One helper in `vars/ck.groovy`, `archiveArchResults(arch, glob)`, renames each matching file to `<arch>_<name>` and archives the prefixed glob. Both `finally` blocks call it: - perf lane: `archiveArchResults(gpuTarget, "*_results.json")` - correctness lane: `archiveArchResults(arch, "dispatcher_*_results.json")` Archived names become e.g. `gfx942_gemm_universal_results.json` / `gfx950_dispatcher_gemm_results.json`, so both arches survive and the arch is in the name. The existing `*_results.json` glob still matches. The helper never throws (`returnStatus: true` + `allowEmptyArchive: true`), so a cmake/ninja failure before any JSON exists is still reported as the real error rather than masked by the archive step. Each lane starts with `deleteDir()`, so no stale prefixed file can be renamed twice. ## Consumers Anything keyed on the exact old names (e.g. the CK legacy perf dashboard ingest) needs to strip the `<arch>_` prefix, and can now key on it instead of guessing. ## Testing Ran the real `runDispatcherPerfTests` / `runDispatcherCorrectnessTests` from `ck.groovy` under Groovy 4 with only the Jenkins steps mocked: `buildAndTest` writes exactly the files named by `--json` in the real `execute_cmd`, `sh` executes for real, and `archiveArtifacts` stores into one map keyed by path (one artifact per path, like a Jenkins build). Four lanes (perf/correctness x gfx942/gfx950), each in its own workspace: | ck.groovy | archived artifacts | |---|---| | develop | 22, all from gfx950 (gfx942 lost) | | this PR | 42 = 20 gfx942 + 22 gfx950 (the 2 extra are the gfx950-only `mx_gemm` / `gemm_bquant`) | | this PR, `buildAndTest` throws | the real build error propagates; archive step does not throw | End-to-end check on Jenkins needs a run of this PR with `USE_CURRENT_BRANCH_FOR_CK_GROOVY=true` (otherwise the `ck` library loads from develop and the old archive step runs) plus `RUN_DISPATCHER_PERF_TESTS=true` and `RUN_DISPATCHER_CORRECTNESS_TESTS=true`. Expected: `gfx942_*_results.json` and `gfx950_*_results.json` both archived. JIRA: AICK-2692 Co-authored-by: Muhammed Emin Ozturk <3836908+ozturkosu@users.noreply.github.com>
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
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.
Summary
DISPATCHER_PERFandDISPATCHER_CORRECTNESSrun gfx942 and gfx950 as parallel branches of the same build, and both archiveprojects/composablekernel/build/*_results.jsonunder identical names. Jenkins keeps one artifact per path, so whichever lane archives last wins and the other arch's results are dropped every night (e.g. build #2866 has a singlegemm_universal_results.json, and alldispatcher_*files are gfx950). The tile_engine JSON carries no arch field, so a consumer cannot even tell which GPU a surviving file came from.Change
One helper in
vars/ck.groovy,archiveArchResults(arch, glob), renames each matching file to<arch>_<name>and archives the prefixed glob. Bothfinallyblocks call it:archiveArchResults(gpuTarget, "*_results.json")archiveArchResults(arch, "dispatcher_*_results.json")Archived names become e.g.
gfx942_gemm_universal_results.json/gfx950_dispatcher_gemm_results.json, so both arches survive and the arch is in the name. The existing*_results.jsonglob still matches.The helper never throws (
returnStatus: true+allowEmptyArchive: true), so a cmake/ninja failure before any JSON exists is still reported as the real error rather than masked by the archive step. Each lane starts withdeleteDir(), so no stale prefixed file can be renamed twice.Consumers
Anything keyed on the exact old names (e.g. the CK legacy perf dashboard ingest) needs to strip the
<arch>_prefix, and can now key on it instead of guessing.Testing
Ran the real
runDispatcherPerfTests/runDispatcherCorrectnessTestsfromck.groovyunder Groovy 4 with only the Jenkins steps mocked:buildAndTestwrites exactly the files named by--jsonin the realexecute_cmd,shexecutes for real, andarchiveArtifactsstores into one map keyed by path (one artifact per path, like a Jenkins build). Four lanes (perf/correctness x gfx942/gfx950), each in its own workspace:mx_gemm/gemm_bquant)buildAndTestthrowsEnd-to-end check on Jenkins needs a run of this PR with
USE_CURRENT_BRANCH_FOR_CK_GROOVY=true(otherwise thecklibrary loads from develop and the old archive step runs) plusRUN_DISPATCHER_PERF_TESTS=trueandRUN_DISPATCHER_CORRECTNESS_TESTS=true. Expected:gfx942_*_results.jsonandgfx950_*_results.jsonboth archived.JIRA: AICK-2692