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
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52

#version 450

// Tiled matrix multiply: C[i,j] = sum_k A[i,k] * B[k,j]
// A: [m, K] row-major, B: [K, n] row-major, C: [m, n] row-major
// Uses shared memory tiles for cache efficiency.

layout(local_size_x = 16, local_size_y = 16) in;

layout(set = 0, binding = 0) readonly buffer InputA { float a[]; };
layout(set = 0, binding = 1) readonly buffer InputB { float b[]; };
layout(set = 0, binding = 2) writeonly buffer Output { float result[]; };

layout(push_constant) uniform Params {
    int m;  // rows of A / rows of C
    int K;  // cols of A / rows of B
    int n;  // cols of B / cols of C
};

shared float tileA[16][16];
shared float tileB[16][16];

void main() {
    uint row = gl_GlobalInvocationID.y;
    uint col = gl_GlobalInvocationID.x;
    uint localRow = gl_LocalInvocationID.y;
    uint localCol = gl_LocalInvocationID.x;

    float acc = 0.0;
    uint numTiles = (uint(K) + 15u) / 16u;

    for (uint t = 0u; t < numTiles; t++) {
        uint aCol = t * 16u + localCol;
        uint bRow = t * 16u + localRow;

        tileA[localRow][localCol] = (row < uint(m) && aCol < uint(K))
            ? a[row * uint(K) + aCol] : 0.0;
        tileB[localRow][localCol] = (bRow < uint(K) && col < uint(n))
            ? b[bRow * uint(n) + col] : 0.0;

        barrier();

        for (uint i = 0u; i < 16u; i++) {
            acc += tileA[localRow][i] * tileB[i][localCol];
        }

        barrier();
    }

    if (row < uint(m) && col < uint(n)) {
        result[row * uint(n) + col] = acc;
    }
}