spike(QUANT-CT-MXFP4): native MXFP4 keep-quant + Qwen e2e — W0 vehicle + W1 target pinned - #38

Closed

localai-bot wants to merge 12 commits into

mainfrom

row/QUANT-CT-MXFP4

Commits on Aug 5, 2026

spike(quant-mxfp4): Qwen W0 vehicle + W1 kernel target pinned (FlashInfer W4A4)
mudler
committed
1d6c0d8
spike(quant-mxfp4): W0 golden RUNS + W1 runtime correction (Marlin W4A16, not FlashInfer W4A4)
mudler
committed
6ff8335
spike(quant-mxfp4): pin the W2 code surface (Marlin W4A16 mxf4) for the continuation
mudler
committed
74b1525
feat(quant-mxfp4): W2 native Marlin W4A16 mxf4 keep-quant path — builds+runs, e2e numerics bug localized
mudler
committed
7068dca
docs(quant-mxfp4): record W2 native mxf4 keep-quant lands (builds+runs), e2e RED with bug localized
mudler
committed
8fef047
test(quant-mxfp4): W3 device unit gate — MXFP4 Marlin GEMM matches CPU dequant ref (M=1+M=8)
mudler
committed
8469e33
test(quant-mxfp4): extend MXFP4 unit gate to real Qwen3-8B shapes (run-pending)
mudler
committed
e28130e
docs(quant-mxfp4): W3 unit gate GREEN — compute proven correct; e2e residual localized off-compute
mudler
committed
26881bb
test(quant-mxfp4): model-facing MXFP4 Apply gate GREEN — BuildMarlinDenseResident correct
mudler
committed
0241045
record(quant-mxfp4): e2e RESOLVED — compute correct (async-off 3/4 token-exact), async-default is a pre-existing non-MXFP4 bug
mudler
committed
1c5ee09
record(quant-mxfp4): quantization-matrix row — native compute + gates + e2e token-exact (async-off)
mudler
committed
151028c
Merge remote-tracking branch 'origin/main' into mxfp4-merge
mudler
committed
c1f3512