llama.cpp

ggml-webgpu: compute pass batching and removing profiling overhead

#21873

Merged