feat: integrate MTPLX 2.4.0 M5 runtime stack - #5

Closed

PhilipJohnBasile wants to merge 29 commits into

mainfrom

codex/moe-dflash-integration-20260730

Commits on Jul 30, 2026

docs: MTP+DFlash coordination design ('Priced Union') + 2026-07-29 drafter market scan

PhilipJohnBasile
and
claude
committed
e41662d
docs+harness: Codex external-review deltas (drafter-cache transactions, full-cycle cost, Phase-0 upgrades) + baseline runner fixes

PhilipJohnBasile
and
claude
committed
9cc9657
Phase 0 executed: tau gate GREEN (4.31 greedy / 3.92 sampled), true AR baselines, T_V(M) curves on all three bodies

PhilipJohnBasile
and
claude
committed
a6a2ac1
Phase 0 corrections per Codex audit: narrow the gate claim, fix tau semantics, add MoE varied-token curve + GDN cache finding

PhilipJohnBasile
and
claude
committed
f433d93
Phase 0: 4-bit prize lane measured — tau is quant-independent, block 8 wins, but reference DFlash lands BELOW MTP D3 on both quant widths

PhilipJohnBasile
and
claude
committed
6a3ea11
docs: mark the section-7 envelope as partly superseded by Phase-0 measurements

PhilipJohnBasile
and
claude
committed
6e85be9
RETRACT the GDN conv 'blocker': it was a probe bug (empty verify rows), plus the real finding and the corrected 4-bit curve

PhilipJohnBasile
and
claude
committed
bfdc589
Phase 0 final: all curves re-measured clean; Qwen DFlash lane heading NO-GO, Laguna is the surviving case

PhilipJohnBasile
and
claude
committed
40f6e4f
FINAL: Qwen DFlash NO-GO confirmed at every block size on a same-body same-suite head-to-head

PhilipJohnBasile
and
claude
committed
982a00a
RETRACT the depth-5 spin-off: repeats show depths 3-5 are indistinguishable

PhilipJohnBasile
and
claude
committed
24097be
THE REVERSAL: DFlash wins ~1.9x over MTP on MoE, protocol-matched and repeat-confirmed

PhilipJohnBasile
and
claude
committed
2c9f420
MoE block sweep: optimum is B5-B6 at ~194 tok/s (1.85x AR), not the reference default of 16

PhilipJohnBasile
and
claude
committed
b6f1d55
CORRECTED per Codex audit: fix a buggy aggregator and a thinking-protocol mismatch; MoE direction holds, numbers restated

PhilipJohnBasile
and
claude
committed
51048a3
Dense arms had the same thinking mismatch: DFlash 1.55x -> 1.89x, NO-GO holds but margin narrows to ~18%

PhilipJohnBasile
and
claude
committed
9a327e7

Commits on Jul 31, 2026

Commits on Aug 1, 2026

Commits on Aug 10, 2026

feat: add guarded OoO-Spec semantic hint lane
PhilipJohnBasile
committed
5262443

Navigation Menu

feat: integrate MTPLX 2.4.0 M5 runtime stack - #5

feat: integrate MTPLX 2.4.0 M5 runtime stack#5
PhilipJohnBasile wants to merge 29 commits into
mainfrom
codex/moe-dflash-integration-20260730

Commits on Jul 30, 2026

docs: MTP+DFlash coordination design ('Priced Union') + 2026-07-29 drafter market scan

docs+harness: Codex external-review deltas (drafter-cache transactions, full-cycle cost, Phase-0 upgrades) + baseline runner fixes

Phase 0 executed: tau gate GREEN (4.31 greedy / 3.92 sampled), true AR baselines, T_V(M) curves on all three bodies

Phase 0 corrections per Codex audit: narrow the gate claim, fix tau semantics, add MoE varied-token curve + GDN cache finding

Phase 0: 4-bit prize lane measured — tau is quant-independent, block 8 wins, but reference DFlash lands BELOW MTP D3 on both quant widths

docs: mark the section-7 envelope as partly superseded by Phase-0 measurements

RETRACT the GDN conv 'blocker': it was a probe bug (empty verify rows), plus the real finding and the corrected 4-bit curve

Phase 0 final: all curves re-measured clean; Qwen DFlash lane heading NO-GO, Laguna is the surviving case

FINAL: Qwen DFlash NO-GO confirmed at every block size on a same-body same-suite head-to-head

RETRACT the depth-5 spin-off: repeats show depths 3-5 are indistinguishable

THE REVERSAL: DFlash wins ~1.9x over MTP on MoE, protocol-matched and repeat-confirmed

MoE block sweep: optimum is B5-B6 at ~194 tok/s (1.85x AR), not the reference default of 16

CORRECTED per Codex audit: fix a buggy aggregator and a thinking-protocol mismatch; MoE direction holds, numbers restated

Dense arms had the same thinking mismatch: DFlash 1.55x -> 1.89x, NO-GO holds but margin narrows to ~18%

Commits on Jul 31, 2026

Merge upstream main before MoE DFlash integration

Add experimental MoE DFlash integration and evidence gates

Honor forged MTP contracts in benchmark gates

Merge remote-tracking branch 'upstream/main' into codex/moe-dflash-integration-20260730

Implement exact sampled-q DFlash reference lane

Commits on Aug 1, 2026

registry: DeepSeek V4 has no MTP head — describe the entry as the DSpark drafter it is

docs+bench: superseded divergence analysis, and a guarded (non-runnable) scaffold

docs: preliminary release-gate proposal, pre-registered and externally audited before the deciding measurement

tests: local joint-law prototype for the K1 acceptance math -- NOT engine evidence

docs: actually canonicalize the audit corrections the previous commit claimed were folded in

Commits on Aug 10, 2026

feat: add guarded OoO-Spec semantic hint lane

Commits on Aug 11, 2026

feat: add exact sparse checkpoint placement planner

feat: add SessionBank checkpoint replay telemetry

Merge pull request #3 from PhilipJohnBasile/codex/sparse-complete-state-checkpoints-20260811

Merge pull request #4 from PhilipJohnBasile/codex/session-overlap-trace-gate-20260811