Repository navigation
Next app rebased - #38
Merged
Merged
Conversation
Adds `pi` (https://pi.dev) as a third agent backend alongside claude and codex. The orchestrator invokes `pi --mode json -p` and reuses the operator's local ~/.pi provider settings; MetaInfer stores no API keys. orchestrator/subagent_manager.py - AgentBackend Literal gains "pi"; _normalize_agent_backend accepts pi / pi-coding-agent / earendil. - pi_bin constructor param; _build_pi_command maps the orchestrator effort knob onto pi --thinking (low/medium/high/max) and reuses the operator's configured provider/model. - pi session isolation: when state_dir is provided, --session-dir is scoped to <state_dir>/pi-sessions so orchestrator sessions are invisible to a user running `pi --continue` from their shell (pi indexes sessions by cwd; without this a user could attach to a production orchestrator session mid-flight and corrupt the context that later --session <id> --continue turns rely on). The session id is captured from pi's leading events for resume. orchestrator/_bootstrap.py - make_subagent_manager gains pi_bin and state_dir; both resolve from args or env (METAINFER_PI_BIN, METAINFER_AGENT_BACKEND) with claude remaining the default. state_dir is forwarded only to the pi backend. orchestrator/token_budget.py - usage_from_result_event tolerates pi's camelCase usage names (input / output / cacheRead / cacheWrite / cost.total) so a raw pi event is accounted correctly. tasks - Every task orchestrator that builds a SubAgentManager now passes state_dir=state_dir so the pi backend can scope its session store per task (calc_value, find_low_hanging_kernel, gen_cpp_infer_framework, gen_infer_framework, gen_infer_framework_cpp, opt_kernel, dcu_kernel_auto_opt, evolve_kernel). opt_GEMM_kernel and port_model already wired it. Other backends ignore the kwarg. docs - README (en/zh): document the pi backend, install, env vars and the effort->thinking mapping. tests - test_subagent_manager covers pi command building, thinking-level mapping, session-dir scoping and resume-id capture. - test_token_budget covers pi camelCase usage parsing.
The detail page could open a library kernel's code, but the code was
rendered as a plain <pre> with no line numbers and no way to see what a
kernel actually changed relative to a baseline — the key question when
reviewing an evolved kernel is "what did it change, vs the original or
vs the kernel it was derived from?".
Add a diff endpoint and an inspector panel:
- _state_readers.py: read_kernel_source() resolves one library kernel by
id; read_kernel_diff() produces a unified diff against the original
reference kernel or against the kernel's parent_id. Missing kernels, a
root kernel with no parent, and an evicted parent (library caps at
MAX_LIBRARY_SIZE) return exists=false with a reason instead of raising.
- routes.py: GET /kernels/{kernel_id}/diff?base=reference|parent.
- ok-runtime-api.js: getKernelDiff() client.
- ok-detail.js: KernelInspector panel with Source / Diff vs reference /
Diff vs parent modes, a line-numbered source renderer, and a colored
unified-diff renderer with +/- stats. The parent mode is disabled when
the kernel has no parent.
- ok.css: .ok-source* / .ok-diff* styles.
- test_kernel_diff.py: 11 tests covering both readers and their
non-raising failure paths.
Co-Authored-By: deepseek/deepseek-v4-flash-0731 <deepseek-ai@claude-code-best.win>
Content imported from the DCU dev container tree (/root/zth_agent/MetaInfer), 3-way merged onto the merged PR #26 head so the next-app alignment and CI fixes stay in place. Scope is limited to this task plugin. Added: - orchestrator: planner.py (+ planner_policy/manifest/gates), gate_policy.py, gpu_preflight.py, harness_io.py, predictions.py, validation_budget.py, variant_promote.py - harness_default/: planner_catalog.yaml, planner_policy.yaml, manifest.yaml, gates.yaml - reports/: best-variant tables (csv/json/xlsx) and the summary scripts - skills/: int8-w8a8-gemm foundations/decode/prefill skill packs - tools/baseline/: Triton INT8 benchmark harness, baseline graph JSONs - 11 new test modules (planner, gate policy, gpu preflight, harness io, predictions, validation budget, variant promote, plan render, prompt schema, planner wiring, planner policy) - retuned variant kernels for glm52/TP8 and hy3/TP4, refreshed WORKFLOW.md Changed: - gen_and_opt_pipeline.py, w8a8_pipeline.py, real_pipeline.py, pipeline.py, skill_store.py, variant_store.py, server/routes.py, static (detail view, shape input, css), form.yaml, bridge/dsh/dsh_agent.py, api/assets helpers - PR-side fixes kept: init_or_resume(task_id) alignment (73c213c) and the CPU-only-CI torch guards (3d70f58) / hermetic dsh tests (6a8f235) Excluded: *.bak-* kernel backups.
…tainer
Import metainfer/tasks/harness_evolve/ verbatim from the DCU development
container, so the AHE outer loop lives in the repository alongside the
DKAO task it drives.
Source: 10.8.160.17:/root/zth_agent/MetaInfer/metainfer/tasks/harness_evolve
73 files, imported without local modification.
What this adds:
- FLOW.md / README.md: the fixed variant-round protocol (HE only launches
DKAO tasks; measurement, GPU gate and contamination stay inside DKAO).
- orchestrator/: pipeline, decision_engine, evolve, rounds, promotion,
champion_selection, supervisor, process_control, variant, pool.
- server/ + static/: the HE web surface (routes, plugin, he-api/detail JS).
- tests/: 33 tests including test_flow_contract.py, which keeps FLOW.md
and the code in sync.
Scope note: this commit is material-only. The package is NOT wired into
the framework yet and is not expected to import cleanly on this branch —
it depends on host APIs that are not part of the DKAO-only PR lineage:
- metainfer.orchestrator.gpu_broker (not in this tree)
- metainfer.server.* (StateStore / forms / registry, next-app lineage)
- metainfer.tasks.dcu_kernel_auto_opt.orchestrator.{harness_io,variant_promote}
plus 11 hardcoded /root/zth_agent/... paths carried over from the container.
Follow-up: reconcile the host API differences and drop the skipif guards
that currently stand in for this module (see the DKAO sync guide).
The metric cards only surface the *last* iteration's speedup and the library's current best, so a run that is stuck — or that regressed after a good iteration — looks identical to one that is climbing, and the per-iteration "goal" (what the optimizer set out to do) was not visible anywhere. Add an "Iteration Progress" panel rendered from the already-fetched `/iterations` payload: one row per iteration with status, exec time, speedup, library size, duration and goal. Rows that set a new best exec time are marked green with a "▲" tag — the "is it actually improving?" signal. - export `formatMs` from ok-kernel-library.js so both tables share one formatter instead of duplicating it - coerce `perf.speedup` with Number() before toFixed: the value is stored straight from an LLM-generated perf harness report with no numeric validation, so a malformed row now degrades to "—" instead of throwing - add tests/test_iteration_progress.py pinning the reader fields the table renders, plus that a partially written iteration record does not take the whole list down
…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.
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
4 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.
No description provided.