Gemma 4 engine gaps: gemma4 registries, KernelDispatch self-heal, and dense FP32 matmul (fixes #1220) - #1221

Merged

michalharakal merged 8 commits into

fix/gemma4-registries-and-fp32-kernels

Aug 31, 2026

Commits on Aug 30, 2026

fix(io,backend): gemma4 tokenizer/arch registry, streaming decodeToken, observable dispatch fallback

michalharakal
and
claude
committed
220c77d

Commits on Aug 31, 2026

feat(kernel): KernelDispatch self-heals via a discoverable ViewKernelPack SPI

michalharakal
and
claude
committed
3734006
perf(cpu): cache the dense weight transpose, and land it where the fast kernels can read it

michalharakal
and
claude
committed
8d4456f
perf(cpu): serve small dense FP32 matmuls directly — the tiled kernel is overhead-bound there

michalharakal
and
claude
committed
558a31a
docs(kernel): record why small dense FP32 matmuls bypass the native kernel

michalharakal
and
claude
committed
019c664
test(kernel): benchmark the FP32/BF16/FP16 matmul kernels at decode shape

michalharakal
and
claude
committed
b840530
perf(kernel): PanamaVectorMatmulKernel was slower than scalar at m=1 — add a GEMV path

michalharakal
and
claude
committed
a63d0e2
fix(build): androidMain cannot see KernelServiceLoader; keep the transpose cache private

michalharakal
and
claude
committed
8f3fe16