37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.
C avx2ccpu-inferenceggufinference-enginelinuxllmlocal-llmmixture-of-expertsmoenvmequantizationqwen3simdspeculative-decodingvulkan
/* SPDX-License-Identifier: GPL-3.0-or-later
Copyright (C) 2026 Maurizio Cammalleri */
#version 450
// A mixture's activations in the experts' order (gpu_token.c): row g of
// dst is the quantized vector of token tok[g] of src, nb blocks of 320
// bytes a vector (the product shaders' layout). A workgroup a row.
#extension GL_EXT_buffer_reference : require
#extension GL_EXT_buffer_reference_uvec2 : require
layout(local_size_x = 64) in;
layout(buffer_reference, std430, buffer_reference_align = 16) readonly buffer U4 { uvec4 v[]; };
layout(buffer_reference, std430, buffer_reference_align = 16) writeonly buffer U4O { uvec4 v[]; };
layout(buffer_reference, std430, buffer_reference_align = 4) readonly buffer U1 { uint v[]; };
layout(push_constant) uniform P { uvec2 src, dst, tok; uint nb; } p;
void main()
{
uint g = gl_WorkGroupID.x, words = p.nb * 20u; // 320 bytes = 20 uvec4
uint t = U1(p.tok).v[g];
U4 S = U4(p.src);
U4O D = U4O(p.dst);
for (uint i = gl_LocalInvocationID.x; i < words; i += 64u)
D.v[g * words + i] = S.v[t * words + i];
}