char-fork/llama.cpp
git clone https://git.t4t.associates/char-fork/llama.cpp
e29acf74f
char/aliceai-t5-moe
1#version 450 2 3#extension GL_EXT_shader_16bit_storage : require 4#extension GL_EXT_control_flow_attributes : enable 5 6layout (push_constant) uniform parameter 7{ 8 uint ncols; 9 uint rows_per_channel; 10 uint n_past; 11} p; 12 13#include "types.glsl" 14 15layout(local_size_x = 1, local_size_y = 512, local_size_z = 1) in; 16 17layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; 18layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; 19 20void main() { 21 const uint col = gl_GlobalInvocationID.y; 22 const uint row = gl_GlobalInvocationID.x; 23 24 if (col >= p.ncols) { 25 return; 26 } 27 28 const uint i = row*p.ncols + col; 29 if (col > p.n_past + row % p.rows_per_channel) { 30 data_d[i] = D_TYPE(uintBitsToFloat(0xFF800000)); 31 } else { 32 data_d[i] = D_TYPE(data_a[i]); 33 } 34}