compiler: SO-3 — the MegaMoE producer consumes inferred dependence edges - #625
Merged
Merged
Conversation
Queue order 3's first producer adoption. Before this, the R3 composition candidate executed MegaMoE's hand-written `_action_dependencies` edges directly. Now: * `megamoe_graph_function` represents the chunk pipeline as Graph IR using REGISTERED semantics — dispatch/combine are the registered ordered collectives (`tessera.moe_dispatch`/`moe_combine`), expert compute is pure `tessera.matmul` over its own chunk's dispatch result. No edges are written here. * `megamoe_inferred_composition` derives the production edges through `infer_action_dag` (W2.1 dataflow + W2.2 effects) and returns the candidate, the inferred DAG, and the parity result. * The hand-authored plan DAG becomes the declared coverage ORACLE (#31): `compare_inferred_action_dag` must show generated edges cover every hand edge or construction fails closed, so a generation regression can never silently weaken the executed schedule. Additional conservative edges (the ordered-collective total order) report separately, per the SO-3 acceptance clause. * The candidate carries the content-addressed `ScheduleObject` built from the inferred edges; its digest is deterministic across constructions. * `megamoe_issue_order` is the one deterministic topological order every consumer indexes (Kahn + (chunk, phase) tie-break). `composition_candidate_for_megamoe_plan` keeps its signature and now routes through the inferred path, so `prune_megamoe_overlap_plans` and every existing consumer are unchanged at the call site. Tests (test_megamoe_inferred_dag.py): deterministic topological order with the overlap property visible; registered-collective semantics in the Graph representation; coverage of the hand oracle with extras reported; deterministic ScheduleObject digest; a generation-regression control that drops the collective semantics and must fail closed; and the pre-existing total-evidence guard. Gates: MegaMoE + composition + schedule-object suites 65 pass; the x86/gfx1151/CPU composition conformance lanes 42 pass (numerics unchanged, per the SO-3 acceptance); mypy 0; ruff clean; dashboards regenerated. Note: a concurrent session is editing this worktree (the spectral half of SO-3 plus pipeline-stage work), so this commit is path-limited to the MegaMoE slice and the whole-suite figure is not reported here — the targeted lanes above are the evidence. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 92385f9fe5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
# Conflicts: # docs/audit/compiler/INTEGRATED_COMPILER_PLAN.md
…an identity, record cross-backend P1 (overlap erased) — CONFIRMED and worse than reported. A 12-action MegaMoE plan inferred all 66 edges of the complete order, so R3 would estimate and prune overlap plans as sequential pipelines. Measured three distinct over-conservatisms in the SHARED inference, each fixed with the sound rule: 1. ordered collectives serialized against every surrounding op. They constrain the order OF COLLECTIVES (all ranks issue them in the same relative order), not unrelated local work — data through a collective is still ordered by SSA/alias/memory edges. Now collective-to- collective only. 2. a registered op declaring `aliasing="none"` had its declaration discarded whenever it was effectful, so every collective result was assumed to alias hidden state and `unknown_alias_fact` became an all-pairs barrier. A declared fact now has its consumer (#29/#30). 3. `has_memory_dependence` returned top for ANY effectful pair. A registered pure op touches no hidden state, so with a disjoint alias set it cannot depend through memory; two pure ops sharing a read-only operand (the weights every expert reads) likewise get no alias edge. Result: 66 → 36 edges, still covering every hand-oracle edge, with the extras now collective-ordering only. `combine:0 -> compute:1` and `compute:0 -> compute:1` are gone; `dispatch:0 -> dispatch:1` and `combine:0 -> combine:1` remain. An existing test asserted the old blanket rule for collectives; it is deliberately replaced by an explicit ordering test (positive: collectives keep relative order; negative: independent local work floats) plus an alias-gated memory-dependence test. The other three barrier classes keep their all-pairs assertion unchanged. P2 (digest identity) — CONFIRMED: action ids and graph shape are functions of the chunk COUNT, so plans differing in capacities/buffers/token ranges/in-flight limit content-addressed identically. The schedule object now binds `plan.artifact_digest`, which already addresses exactly those fields; `candidate_id` stays the plan id so pruning keys are unchanged. P1 (cross-backend) — sync key SO3-INFER-EDGES-2026-08-24 in all four queues: rocm/x86 parity validated (analysis is prune/rank-only; no generated code or numerics change), nvidia/apple not-applicable-today with the inherited-on-adoption reason. Also resolves the merge conflict with #624 by keeping both queue rows. Gates: composition-cost + graph-dataflow + MegaMoE suites 71 pass; mypy 0; ruff clean; generated docs in sync. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…' into tmp/so3-merge # Conflicts: # docs/audit/compiler/INTEGRATED_COMPILER_PLAN.md
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
1 participant
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.
Queue order 3 (SO-3 + W5.2e-PRODUCER-1), first producer adoption: the collective/MoE producer stops executing hand-written edges.
What changed
megamoe_graph_functionrepresents the chunk pipeline as Graph IR using registered semantics — dispatch/combine are the registered ordered collectives, expert compute is puretessera.matmulover its own chunk's dispatch result. No edges are written here.megamoe_inferred_compositionderives the production edges viainfer_action_dag(W2.1 dataflow + W2.2 effects).ScheduleObjectbuilt from inferred edges; digest is deterministic.composition_candidate_for_megamoe_plankeeps its signature, soprune_megamoe_overlap_plansand all existing consumers are unchanged at the call site.Tests
test_megamoe_inferred_dag.py— deterministic topological order (with the overlap property visible in the order itself); registered-collective semantics; oracle coverage with extras reported; deterministic digest; a generation-regression control that strips the collective semantics and must fail closed; plus the pre-existing total-evidence guard.Evidence
MegaMoE + composition + schedule-object suites 71 pass; x86/gfx1151/CPU composition conformance lanes 42 pass (numerics unchanged — the SO-3 acceptance clause); mypy 0; ruff clean; dashboards regenerated.
Coordination note (please read)
A concurrent session is editing this same worktree — it has
scheduled_spectral.py,test_spectral_inferred_dag.py, and the pipeline-stage passes modified, i.e. the spectral half of this very queue row. This PR is therefore path-limited to the MegaMoE slice, and I deliberately do not report a whole-suite figure (the ~23 spectral/matmul/fp8 failures in my runs are that session's in-progress work, not this change). The targeted lanes above are this PR's evidence.Remaining on order 3 after this: the spectral producer (in flight elsewhere), digest stamping through the physical lowering seam, and the scalar-pipeline-reconstruction deletion.
🤖 Generated with Claude Code