feat: Triton fused SoftSignGLU kernel for LYNXNet2 - #312

Merged

yqzhishen merged 23 commits into

openvpi:mainfrom

KakaruHayate:feat/fused-softsign-glu

Aug 13, 2026

Commits on Jul 2, 2026

feat: Triton fused SoftSignGLU kernel for LYNXNet2
KakaruHayate
committed
d8b261c
perf: skip variance fusion, realistic warmup M, fewer autotune configs
KakaruHayate
committed
e357c71
fix: remove useless warmup from __init__ (model on CPU, no compilation)
KakaruHayate
committed
99dad24
fix: disable variance fusion by default (small backbones, no benefit)
KakaruHayate
committed
06cc355
fix: remove useless warmup from __init__ (model on CPU, no compilation)
KakaruHayate
committed
6058ec5
fix: address PR review issues
KakaruHayate
committed
9256c0b
fix: elem kernel extra N arg; variance nested glu_type
KakaruHayate
committed
d29f868

Commits on Jul 3, 2026

fix: backward kernel dtype not hardcoded fp16
KakaruHayate
committed
ade184c

Commits on Jul 7, 2026

Commits on Jul 14, 2026

review: fix hard blocks — TF32 removal, glu_type revert, robustness
KakaruHayate
committed
b336d78

Commits on Jul 15, 2026

Commits on Jul 21, 2026

fix: fail fast when Triton is unavailable
KakaruHayate
committed
f8056c2

Commits on Aug 3, 2026

fix: warm only patched variance backbones
KakaruHayate
committed
c13588f

Commits on Aug 6, 2026

fix: address CodeRabbit fused kernel review
KakaruHayate
committed
222a72b
fix: fail fast when fused kernels lack Triton
KakaruHayate
committed
6133bb5
refactor: finalize fused SoftSignGLU integration
KakaruHayate
committed
8cd54bf

Commits on Aug 7, 2026

fix: isolate warmup RNG, restore 1D input rank
KakaruHayate
committed
6716397

Commits on Aug 13, 2026

Harden LYNXNet2Block layout check in fusion wrapper
KakaruHayate
committed
b30828d