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
45 lines
1.5 KiB
Plaintext
45 lines
1.5 KiB
Plaintext
#include "types.comp"
|
|
|
|
#extension GL_EXT_shader_16bit_storage : require
|
|
|
|
layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
|
|
|
|
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
|
|
layout (binding = 1) readonly buffer Y {int data_pos[];};
|
|
layout (binding = 2) readonly buffer Z {float data_ff[];};
|
|
layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
|
|
|
|
layout (push_constant) uniform parameter {
|
|
uint ncols;
|
|
uint n_dims;
|
|
float freq_scale;
|
|
uint p_delta_rows;
|
|
float freq_base;
|
|
float ext_factor;
|
|
float attn_factor;
|
|
float corr_dims[2];
|
|
float theta_scale;
|
|
uint has_ff;
|
|
} p;
|
|
|
|
float rope_yarn_ramp(const float low, const float high, const uint i0) {
|
|
const float y = (i0 / 2 - low) / max(0.001f, high - low);
|
|
return 1.0f - min(1.0f, max(0.0f, y));
|
|
}
|
|
|
|
void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta) {
|
|
float mscale = p.attn_factor;
|
|
// Get n-d rotational scaling corrected for extrapolation
|
|
float theta_interp = p.freq_scale * theta_extrap;
|
|
float theta = theta_interp;
|
|
if (p.ext_factor != 0.0f) {
|
|
float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
|
|
theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
|
|
|
|
// Get n-d magnitude scaling corrected for interpolation
|
|
mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
|
|
}
|
|
cos_theta = cos(theta) * mscale;
|
|
sin_theta = sin(theta) * mscale;
|
|
}
|