issue #88: Gemma v3 decision-stability physical producer — campaign terminal V3_HOLDOUT_FAIL - #30
Merged
Merged
Conversation
…terogeneous numerical qualification Implements the minimum harness required to execute the frozen issue FlashML-org#74 methodology (inferswarm@f394dc9): case-driven single-arm reference runner (RTX 3090), three-stage chain runner (2x RTX 3060 node-01 + remote last stage on node-03 via a per-case R4 wire service), RowPruningSink capture (FlashML-org#71-compatible, host-side final-row pruning of the full BF16 logits matrix), o_proj checkpoint wrappers (layer-0 in/out, global-layer-15 attention projection), pure host-float64 reducer matching the frozen REDUCER.md identity, and torch-free unit/source-contract tests. No execution/model math is added or changed: all model execution flows through the accepted R6 GemmaDenseStage replay-prefill semantics.
…the full corpus sentinel-subset.json carries identity hashes without prompt text; runners now cross-check subset rows against the frozen calibration corpus before execution.
…stage 1's bundle)
…ness Cherry-picks the FlashML-org#76/FlashML-org#81 execution harness verbatim (PR #29, b1389e3^..9f06d81; zero tree diff over benchmarks/inferswarm_76) and adds the issue-FlashML-org#88 v3 semantic layer only (benchmarks/inferswarm_88): - frozen argmax/tie-break rule twin + per-decision executor rule proofs; - reference-only decision domain D(r) construction (reference-top-1024-with-cutoff-ties/1) with canonical membership hashes; - candidate teacher-forcing with mechanical prefix-identity proof before each of all 8 decisions (fail closed); - retention of the candidate actual full-vocabulary FP32 winner per canonical-prefix row (decision-<i>.f32 on the last-stage node, sha256-bound, never free-run); - all-8-decision reference FP32 rows retained for E_full-class evidence and decision_local_error; - unchanged 15-envelope capture at 0/1/3/7 from the FlashML-org#76 harness. 28 CPU-pure producer tests freeze the contract before any physical execution (prefix identity, tie-break rule incl. exact ties, 8-rows- exactly-once, 15-envelope completeness, candidate-cannot-influence-D(r), identity binding, torch-free semantic layer).
…alid-attempt defect) phaseC-1 on inferswarm04 chained arm_full_capture inside the per-case loop: o_proj wrappers re-bound each case, duplicating capture records per case (40 -> 604 records by case 48) until / filled (89 GB of duplicated bundles) and torch.save failed. Duplicate host-side copies only — each wrapper layer passes the exact tensor through, no device math change; Phase B logits-derived margins/selected-eight unaffected (decision-.f32 rows come from the pre-wrapper logits return value and their sha256s are recorded per decision). Fix: arm ONCE at startup (the accepted FlashML-org#76 pattern); per-case isolation via sink swap only. Attempt phaseC-1 retained and classified producer-capture-defect + infrastructure disk exhaustion (invalid).
…seD-stress-1 defect) The FlashML-org#88 chain runner only forwarded capture_step at the frozen envelope positions 0/1/3/7, so the last-stage service retained decision rows (and returned rule proofs) for only those four decisions. v3 needs the ACTUAL full-vocab winner at ALL 8 decisions. Fix: capture_step is sent on every decision (the last-stage keys its retained decision-<i>.f32 rows off this value); the 15-envelope capture positions themselves remain the frozen 0/1/3/7 (envelope reduction unchanged; stages 1-2 emit their per-position sinks exactly as before — positions 2/5/6 now also emit, which the reducer's per-(position, checkpoint) grouping already handles by ignoring non-frozen positions... no: the reducer iterates CAPTURE_POSITIONS only, so extra positions are inert). Attempt phaseD-stress-1 retained as invalid (incomplete decision evidence: 4/8 rows per case).
ezutfen
changed the title
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.
Physical producer for InferSwarm issue FlashML-org#88 (execute frozen v3 decision-stability qualification).
Methodology: inferswarm@a8ec98a9fb9b673c93de5100d784ea772395efdb (issue FlashML-org#86 / PR FlashML-org#87).
Left unmerged until maintainer campaign acceptance per issue FlashML-org#88 evidence discipline.