char-fork/llama.cpp
git clone https://git.t4t.associates/char-fork/llama.cpp
e29acf74f
char/aliceai-t5-moe
1#version 450 2 3#include "types.glsl" 4#include "generic_unary_head.glsl" 5 6layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; 7 8uint src0_idx_mod(uint idx) { 9 const uint i13 = idx / (p.ne12*p.ne11*p.ne10); 10 const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10; 11 const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10); 12 const uint i12_offset = i12*p.ne11*p.ne10; 13 const uint i11 = (idx - i13_offset - i12_offset) / p.ne10; 14 const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10; 15 return (i13 % p.ne03)*p.nb03 + (i12 % p.ne02)*p.nb02 + (i11 % p.ne01)*p.nb01 + (i10 % p.ne00)*p.nb00; 16} 17 18void main() { 19 const uint idx = get_idx(); 20 21 if (idx >= p.ne) { 22 return; 23 } 24 25 data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx_mod(idx)]); 26}