Run-5 ran 34 cycles (c24-c57) with ZERO perf-keeps. All 6 structuralSwingIdeas from the run-4 plan were directly DISPROVEN: #1 gate+up split-K on K=4096: c37 -3.1%/-6.1%; c55 -1.8% #2 last-WG-does-norm fusion: c54 bit-correct but -1.4% #3 split_merge micro-wins: c42/43/45/52 flat #5 hidden-dim-rotated dispatch: c44 -0.34% #6 gate+up LDS-input broadcast: c35 -13.8% Plus 25+ smaller direct attacks on gate+up shader, down_proj, flash_attn, split_merge — all flat-or-regress. 108.05 tok/s is now confirmed as a hard plateau without coopmat. Updates: - knownFlatCategories: comprehensive consolidation. Marks each disproved swing with the cycle data + failure reason. Adds 14+ new run-5 dead-ends. Documents that run-4's headline +8.7% was mostly system-state shift; real code-driven gains were ~+2.2%. - structuralSwingIdeas: refactored around the realistic remaining levers. Top: flash_attn_cm1.comp coopmat port (multi-week ceiling, only untried structural lever). Then: Br=2+split-K combined, profile-phase decomp of "other" 33%, conditional split_merge dispatch (unblocks cycle-56's +0.41), q5k LM-head row-swap, TPB=128 q6k LM head, gate+up LDS+NUM_ROWS=4 unit. - Effort doc TL;DR rewritten with the 5-run trajectory: 31.19 → 108.05 = 3.46×. All 5 runs tabulated. Disproved swings table. Process gaps documented (no-op-keeps, multi-modal blockers, run-to-run baseline drift). - New criterion: ≥110 (coopmat-bypass via small remainder wins) or ≥130 (with coopmat port). Down-revised from the 115 target since every cheap structural swing is now confirmed dead.

Windows 11 Build Fails #1

Description

Activity

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions