perf(dspark): unblock upstream ncu; the DRAM attribution is CORROBORATED (#442) - #562

Merged

localai-bot merged 21 commits into

mainfrom

row/SPEC-DSPARK-MARLIN-NCU

Aug 14, 2026

Commits on Aug 12, 2026

docs(dspark): storage refuted; ratio stable at ~0.966 over three sessions (#442)
mudler
committed
b0fd2af

Commits on Aug 13, 2026

merge: origin/main into row/SPEC-DSPARK-LOADER
mudler
committed
73088f0
perf(dspark): unblock upstream ncu, and REFUTE the DRAM-bound attribution (#442)
mudler
committed
b87b6a5
fix(record): retract the "latency-bound" reading -- 6z's DRAM attribution stands (#442)
mudler
committed
44c0a48
perf(dspark): the Marlin kernel is NOT the gap -- 6x's localisation REFUTED (#442)
mudler
committed
e60efef
perf(dspark): the in-situ denominator was WRONG -- experts, not blocks (#442)
mudler
committed
cabaf17
fix(record): ours touches FEWER experts, not more -- 6af's inference was inverted (#442)
mudler
committed
115b630
fix(record): the GEMM mix, and the MODE HOLE under every in-situ ratio (#442)
mudler
committed
b0fc9d8
perf(dspark): blocks_per_sm is a DEAD lever, and the box cannot currently resolve 3% (#442)
mudler
committed
1fa4c2d
merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU
mudler
committed
8c74dd1
fix(record): the standalone Marlin runs were taken UNLOCKED -- wrong lock file (#442)
mudler
committed
69e9e7a
fix(record): review FAIL repairs -- per-expert cost is NOT flat, and the bound direction was inverted (#442)
mudler
committed
d7d67be
merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU
mudler
committed
571102c
perf(dspark): the first fully-controlled paired run measures 0.9889, not 0.966 (#442)
mudler
committed
ae6a8d9
merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU
mudler
committed
84ee11d
record(dspark): the n=2 repeat FAILS its own drift gate -- ~0.98 on ONE valid run (#442)
mudler
committed
258f501

Navigation Menu

perf(dspark): unblock upstream ncu; the DRAM attribution is CORROBORATED (#442) - #562

perf(dspark): unblock upstream ncu; the DRAM attribution is CORROBORATED (#442)#562
localai-bot merged 21 commits into
mainfrom
row/SPEC-DSPARK-MARLIN-NCU

Commits on Aug 12, 2026

docs(dspark): storage refuted; ratio stable at ~0.966 over three sessions (#442)

Commits on Aug 13, 2026

merge: origin/main into row/SPEC-DSPARK-LOADER

perf(dspark): unblock upstream ncu, and REFUTE the DRAM-bound attribution (#442)

fix(record): retract the "latency-bound" reading -- 6z's DRAM attribution stands (#442)

perf(dspark): the Marlin kernel is NOT the gap -- 6x's localisation REFUTED (#442)

perf(dspark): the in-situ denominator was WRONG -- experts, not blocks (#442)

fix(record): ours touches FEWER experts, not more -- 6af's inference was inverted (#442)

fix(record): the GEMM mix, and the MODE HOLE under every in-situ ratio (#442)

perf(dspark): blocks_per_sm is a DEAD lever, and the box cannot currently resolve 3% (#442)

merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU

fix(record): the standalone Marlin runs were taken UNLOCKED -- wrong lock file (#442)

fix(record): review FAIL repairs -- per-expert cost is NOT flat, and the bound direction was inverted (#442)

merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU

perf(dspark): the first fully-controlled paired run measures 0.9889, not 0.966 (#442)

merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU

record(dspark): the n=2 repeat FAILS its own drift gate -- ~0.98 on ONE valid run (#442)

Commits on Aug 14, 2026

merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU

merge: origin/main, and repair the five blocking findings of the second review (#442)

merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU, and record the third lock spelling

merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU, and record the third lock spelling

merge: origin/main into row/SPEC-DSPARK-MARLIN-NCU, and record the third lock spelling