A single day of rapid iteration took the Gemma 4 port from a baseline to validated QKV fusion, invalidating several assumptions along the way.
Fusing QKV, inlining norm-weight application, and pushing block loads through the Q4_K path.
The standalone backend finally produced inference after fixing ABI layout, shared-memory buffers, and a simple CPU fallback.
47 commits and a progressive enablement series that ended with full local decode dispatch at 7.0 t/s and 100% correctness.
SLM-tiled GEMM, SPIR-V decompilation, and a corrected Q4_K dequant that finally produced coherent output.
A 12-hour debugging session documented commit by commit until I isolated the EP bug to a fused aggregation kernel.