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
23
24
25
26
27
28
29
30
31
32
33

#version 450

// Find max value per workgroup, write to partial_max[gl_WorkGroupID.x]
layout(local_size_x = 256) in;

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

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

shared float sdata[256];

void main() {
    uint tid = gl_LocalInvocationID.x;
    uint idx = gl_GlobalInvocationID.x;

    sdata[tid] = (idx < n) ? x[idx] : -3.402823466e+38;

    barrier();

    for (uint s = gl_WorkGroupSize.x / 2; s > 0; s >>= 1) {
        if (tid < s) {
            sdata[tid] = max(sdata[tid], sdata[tid + s]);
        }
        barrier();
    }

    if (tid == 0) {
        partial_max[gl_WorkGroupID.x] = sdata[0];
    }
}