llama.cpp

ggml webgpu: initial flashattention implementation

#18610

Merged