mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-11-14 23:09:53 +00:00
7c7836d9d4
* Refactor shaders, extract GLSL code from ggml_vk_generate_shaders.py into vulkan-shaders directory * Improve debug log code * Add memory debug output option * Fix flake8 * Fix unnecessary high llama-3 VRAM use
38 lines
901 B
Plaintext
38 lines
901 B
Plaintext
#version 450
|
|
|
|
#include "rope_head.comp"
|
|
|
|
void main() {
|
|
const uint col = gl_GlobalInvocationID.y * 2;
|
|
const uint row = gl_GlobalInvocationID.x;
|
|
|
|
if (col >= p.ncols) {
|
|
return;
|
|
}
|
|
|
|
if (col >= p.n_dims) {
|
|
const uint i = row*p.ncols + col;
|
|
|
|
data_d[i + 0] = data_a[i + 0];
|
|
data_d[i + 1] = data_a[i + 1];
|
|
|
|
return;
|
|
}
|
|
|
|
const uint i = row*p.ncols + col;
|
|
const uint i2 = row/p.p_delta_rows;
|
|
|
|
const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f);
|
|
|
|
const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f;
|
|
|
|
float cos_theta, sin_theta;
|
|
rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta);
|
|
|
|
const float x0 = float(data_a[i + 0]);
|
|
const float x1 = float(data_a[i + 1]);
|
|
|
|
data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
|
|
data_d[i + 1] = D_TYPE(x0*sin_theta + x1*cos_theta);
|
|
}
|