Skip to content

docs(M74): record forward_qwen3_moe_traced_with_plan SHIPPED — M-MOE-SUB-2 step (a)#60

Merged
noahgift merged 1 commit intomainfrom
docs/m74-record-pr-1516-moe-sub-2-step-a
May 5, 2026
Merged

docs(M74): record forward_qwen3_moe_traced_with_plan SHIPPED — M-MOE-SUB-2 step (a)#60
noahgift merged 1 commit intomainfrom
docs/m74-record-pr-1516-moe-sub-2-step-a

Conversation

@noahgift
Copy link
Copy Markdown
Contributor

@noahgift noahgift commented May 5, 2026

Summary

Records aprender PR #1516 MERGED on aprender main 2026-05-05 as squash `3138d134d` — second substantive code in the M-GPU-MOE-1.4 traced-bisection chain.

Cross-reference bumps

  • README status block: M0–M73 → M0–M74
  • CONTRIBUTING status footer: M0–M73 → M0–M74
  • Spec status header: M0–M73 → M0–M74 + M74 narrative
  • Spec status snapshot: M0–M73 → M0–M74
  • Run history Run 1 end-M: M1–M73 → M1–M74

Verification

  • `bash scripts/check-doc-drift.sh` PASS — sub-milestones tail M74
  • `pv validate contracts/claude-code-parity-apr-v1.yaml` 0/0

Refs: aprender #1516 MERGED

🤖 Generated with Claude Code

…SUB-2 step (a)

aprender PR #1516 MERGED on aprender main 2026-05-05 as squash
3138d134d. Companion-only spec record.

**Second substantive code in the M-GPU-MOE-1.4 traced-bisection
chain** (M68 was the helper, M74 wires it). Wires
SaveTensorStage::MoeRouter + MoeFfnOut emission into the CPU
traced MoE forward path per
contracts/trace-moe-gpu-sub-stages-v1.yaml v1.1.0 step (a).

Adds new method `forward_qwen3_moe_traced_with_plan` accepting
`Option<&SaveTensorPlan>`; existing `forward_qwen3_moe_traced`
becomes a thin one-line delegate passing None. Public API
unchanged.

When plan selects MoeRouter or MoeFfnOut for a layer, the last
sequence position's MoE forward dispatches through
moe_ffn_forward_layer_with_router (M68 helper); other positions
use production moe_ffn_forward_layer.

Production forward_qwen3_moe / forward_qwen3_moe_cuda hot paths
byte-unchanged (additive-purity invariant pinned in v1.1.0).

Discharges FALSIFY-MOE-SUB-002 partially at the CPU-traced
surface. Full discharge needs:
  - M-MOE-SUB-2 step (b): GPU sibling forward_qwen3_moe_cuda_traced
  - M-MOE-SUB-3: live bisection on lambda-vector RTX 4090
  - M-GPU-MOE-1.4: fix at bisected stage

Cross-reference bumps:

- README status block:                M0–M73 → M0–M74
- CONTRIBUTING status footer:         M0–M73 → M0–M74
- Spec status header (line 5):        M0–M73 → M0–M74 + M74 narrative
- Spec status snapshot (line 311):    M0–M73 → M0–M74
- Run history Run 1 end-M (line 746): M1–M73 → M1–M74

Drift detector PASS — sub-milestones tail M74, gate count 13,
contract v1.23.0, corpus 30/30.

Refs: aprender PR #1516 MERGED 2026-05-05 (squash 3138d134d),
      contracts/trace-moe-gpu-sub-stages-v1.yaml v1.1.0 step (a),
      M68 helper PR #1507 (squash 0f22c7841)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@noahgift noahgift enabled auto-merge (squash) May 5, 2026 15:17
@noahgift noahgift merged commit 0ef9264 into main May 5, 2026
1 check passed
@noahgift noahgift deleted the docs/m74-record-pr-1516-moe-sub-2-step-a branch May 5, 2026 15:35
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant