llama.cpp

opencl: perf optimization for mamba2 ssm_scan by folding 4 dim rows into one workgroup and add more coverage

#27775

Open