llama-gguf 0.14.0

A high-performance Rust implementation of llama.cpp - LLM inference engine with full GGUF support
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

#version 450

layout(local_size_x = 256) in;

layout(set = 0, binding = 0) readonly buffer Input { float x[]; };
layout(set = 0, binding = 1) writeonly buffer Output { float result[]; };

layout(push_constant) uniform Params {
    int n;
};

void main() {
    uint idx = gl_GlobalInvocationID.x;
    if (idx < n) {
        float val = x[idx];
        // GELU approximation: 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))
        float c = 0.7978845608; // sqrt(2/pi)
        float inner = c * (val + 0.044715 * val * val * val);
        inner = clamp(inner, -10.0, 10.0);
        result[idx] = 0.5 * val * (1.0 + tanh(inner));
    }
}