char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Jeff Bolzvulkan: Support Q1_0 (#21539)7b69125332026-04-10

#version 450

#include "dequant_head.glsl"

layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;

layout (binding = 0) readonly buffer A {block_q1_0 data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};

void main() {
    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;

    const uint tid = gl_LocalInvocationID.x % 64;
    const uint il  = tid / 4;
    const uint ir  = tid % 4;
    const uint ib = 4*i + ir;
    if (ib >= p.nel / 128) {
        return;
    }

    const uint b_idx = 512*i + 128*ir + 8*il;

    const float d = float(data_a[ib].d);
    const uint bits = uint(data_a[ib].qs[il]);

    [[unroll]] for (uint l = 0; l < 8; ++l) {
        data_b[b_idx + l] = D_TYPE((bits & (1u << l)) != 0u ? d : -d);
    }
}