Repository navigation
feat(dcu-kernel-auto-opt): harness planner, gate policy and DKAO tooling - #36
Merged
myrfy001 merged 1 commit intoSep 28, 2026
Conversation
…ainer Updates the task plugin to the latest container content (/root/zth_agent/MetaInfer/metainfer/tasks/dcu_kernel_auto_opt), applied as an incremental 3-way merge so the PR-side alignment fixes stay intact (init_or_resume(task_id), CPU-only-CI torch guards, hermetic dsh tests). Added: - orchestrator/contention.py, orchestrator/round_strategy.py and harness_default/systemprompt/round_strategy.yaml - tools/retest_variant.py, tools/baseline/tp8_m8_* (Triton m8 baselines) - reports/: down-proj diagnostic scripts and tooling checks - tests: test_contention, test_execution_profile, test_gpu_mode, test_retest_tool, test_round_strategy, test_shape_catalog_panel, test_ahe_child_embed Changed: planner/gate_policy/gpu_preflight/harness_io/config/prompts, gen_and_opt_pipeline, w8a8_pipeline, w8a8_baselines, server/routes, static detail/shape panels, form.yaml, WORKFLOW.md, bridge/dsh. Dependency guards (tests only): Three features measured by this plugin live outside this PR's scope for now — metainfer.orchestrator.gpu_broker (GPU broker leases), the DCU toolchain (/opt/dtk/bin/hipprof PMC profiling) and the harness_evolve sibling task. Eleven tests that require them are marked with pytest.mark.skipif via importlib.util.find_spec / path checks (same convention as the torch guards), so the suite stays green in CI and they re-enable automatically once those modules land.
fms-zth
force-pushed
the
feat/dcu-kernel-auto-opt-20260911
branch
from
September 20, 2026 13:42
0345ba7 to
20b8dc2
Compare
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
2 participants
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
Brings the
dcu-kernel-auto-opt(DKAO) task plugin up to the state of the DCU development container, as two commits:6a22b60— harness planner, gate policy and DKAO tooling0345ba7— refresh from the container (contention/round-strategy, retest tooling, m8 baselines, 7 new test modules)Scope is limited to
metainfer/tasks/dcu_kernel_auto_opt/— 132 files, 0 files outside the plugin, 0 deletions.What's in it
Orchestrator / harness
planner.py+harness_default/{planner_catalog,planner_policy,manifest,gates}.yamlandharness_default/systemprompt/round_strategy.yamlgate_policy.py,gpu_preflight.py,harness_io.py,predictions.py,validation_budget.py,variant_promote.pycontention.py,round_strategy.py,w8a8_baselines.py; updates toconfig.py,prompts.py,gen_and_opt_pipeline.py,w8a8_pipeline.py,real_pipeline.py,skill_store.py,variant_store.pyTooling / evidence
tools/retest_variant.py,tools/baseline/Triton INT8 benchmarks + TP8 M8 graph JSONsreports/best-variant tables (csv/json/xlsx) plus the down-proj diagnostic scriptsskills/INT8 W8A8 GEMM skill packs (foundations / decode / prefill)Tests: 11 new modules (planner, gate policy, GPU preflight, harness IO, predictions, validation budget, variant promote, contention, execution profile, GPU mode, round strategy, shape catalog panel, retest tool).
PR-side fixes from the previous round are preserved:
init_or_resume(task_id)alignment, CPU-only-CI torch guards, hermetic dsh agent tests.Dependency guards
Three capabilities this plugin measures are not part of this PR and do not exist on
next-appyet:metainfer.orchestrator.gpu_brokerconfig.py,gen_and_opt_pipeline.py— the latter already falls back viatry/except)pytest.mark.skipif(importlib.util.find_spec(...) is None)on 10 tests/opt/dtk/bin/hipprofskipifon the probe test (DCU toolchain only)metainfer.tasks.harness_evolvetools/retest_variant.py, AHE child embeddingskipifon 1 testSame convention as the existing torch guards: the suites stay green in CI and re-enable automatically once those modules land. Follow-ups: land
gpu_broker.py(+harness_evolve) as their own PRs; no runtime path in the plugin is broken by their absence (all imports are lazy/guarded).Tests
HuangPuStar/MetaInfer:next-app(baseline)The 10 failures are identical in both trees and all outside this plugin (2 plugin-registry/form checks, 2
gen_infer_framework_cpp, 2sglang_trace_analyze, 4tests/e2e). This PR therefore introduces no new failures and +181 passing tests.Plugin suite on its own:
417 passed, 14 skipped(13 dependency/environment guards + 1 pre-existing PyTorch guard).