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 */
/*
* onnx_pb.h - the protobuf wire format, as the converters of ONNX models
* (silero2jns, wespeaker2jns) read it: fields one after another, their
* numbers, wire types and bytes.
*/
#ifndef JANAS_TOOLS_ONNX_PB_H
#define JANAS_TOOLS_ONNX_PB_H
#include <stdint.h>
#include <string.h>
struct pb {
const uint8_t *p, *end;
};
static inline int varint(struct pb *b, uint64_t *v)
{
*v = 0;
for (int s = 0; s < 64 && b->p < b->end; s += 7) {
uint8_t c = *b->p++;
*v |= (uint64_t)(c & 0x7F) << s;
if (!(c & 0x80))
return 0;
}
return -1;
}
/* the next field: its number, wire type, and for length-delimited ones
its bytes in sub; 0 at the end, -1 broken */
static inline int field(struct pb *b, int *num, int *wt, uint64_t *v,
struct pb *sub)
{
if (b->p >= b->end)
return 0;
uint64_t key;
if (varint(b, &key) != 0)
return -1;
*num = (int)(key >> 3);
*wt = (int)(key & 7);
switch (*wt) {
case 0:
return varint(b, v) == 0 ? 1 : -1;
case 1:
if (b->end - b->p < 8)
return -1;
b->p += 8;
return 1;
case 5:
if (b->end - b->p < 4)
return -1;
b->p += 4;
return 1;
case 2:
if (varint(b, v) != 0 || *v > (uint64_t)(b->end - b->p))
return -1;
sub->p = b->p;
sub->end = b->p + *v;
b->p += *v;
return 1;
default:
return -1;
}
}
static inline int str_eq(const struct pb *s, const char *c)
{
size_t n = strlen(c);
return (size_t)(s->end - s->p) == n && !memcmp(s->p, c, n);
}
#endif