llama.cpp

opencl: fold the gpt-oss MoE per-expert bias adds into the epilogue (op/kernel fusion)

#26431

Merged