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 */
/*
* store_disk.h - the store's documents on disk (store_disk.c), one file
* each, so that they outlive the server. Used by store.c only.
*/
#ifndef JANAS_SERVER_STORE_DISK_H
#define JANAS_SERVER_STORE_DISK_H
#include <stddef.h>
#include <stdint.h>
struct srv_disk;
/* The directory, made if missing (0700), and locked for this process:
NULL with a reason in err when another server holds it, or on failure. */
struct srv_disk *srv_disk_open(const char *dir, char *err, size_t err_len);
void srv_disk_close(struct srv_disk *d);
/* Whether kind and id make a file name (letters, digits, '.', '_', '-',
not starting with '.'): a document whose names do not stays in memory. */
int srv_disk_name_ok(const char *s);
/* Writes the document (aux may be NULL), replacing what was there: 0 or -1. */
int srv_disk_write(struct srv_disk *d, const char *kind, const char *id,
uint64_t seq, int64_t created, const char *json,
const char *aux);
void srv_disk_remove(struct srv_disk *d, const char *kind, const char *id);
/* Calls fn for every document found, in no order; fn takes json and aux
(aux may be NULL), and returns non-zero to stop. Files that cannot be
read are skipped, and counted in the return value. */
typedef int (*srv_disk_fn)(void *arg, const char *kind, const char *id,
uint64_t seq, int64_t created, char *json,
char *aux);
int srv_disk_load(struct srv_disk *d, srv_disk_fn fn, void *arg);
#endif