feat(structures): ggml host adapter — llama.cpp distribution for pi0.5 on Thor (SM110) - #185

Open

LiangSu8899 wants to merge 17 commits into

perf/pi05-precision-studyfrom

feat/structures-ggml-adapter

Commits on Aug 24, 2026

feat(fp4): add SigLIP FFN GEMM pair variant with fp32 bias/residual boundaries
LiangSu8899
committed
a3d351e
feat(structures): add native ggml host adapter (Thor SM110 NVFP4 path)
LiangSu8899
committed
357dba3
feat(structures): pipeline binding and qualification gates for the ggml pi0.5 host
LiangSu8899
committed
7b7e8fc
feat(ggml-adapter): fold the Gemma-style RMS norm chain into the modulate kernel
LiangSu8899
committed
0992c52
feat(ggml-adapter): prefill fused QKV window
LiangSu8899
committed
0990445
feat(fp4): fp32-bias fp16-out GEMM variant; vision QKV window emits f16 K/V directly
LiangSu8899
committed
e5cf117
feat(ggml-adapter): decomposed tiny-M decode attention
LiangSu8899
committed
b70ee45
chore(ggml-adapter): refresh the e2e action golden
LiangSu8899
committed
29ef196

Commits on Aug 25, 2026