char-fork/llama.cpp

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

xctanggml-cpu : split arch-specific implementations (#13892)f470bc36b

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
2.0 KiB55 linesraw
1#ifdef GGML_USE_CPU_HBM
2
3#include "ggml-backend.h"
4#include "ggml-backend-impl.h"
5#include "ggml-cpu.h"
6#include "ggml-impl.h"
7
8#include "hbm.h"
9
10// buffer type HBM
11
12#include <hbwmalloc.h>
13
14static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
15    return "CPU_HBM";
16
17    GGML_UNUSED(buft);
18}
19
20static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
21    hbw_free(buffer->context);
22}
23
24static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
25                                                                           size_t                     size) {
26    void * ptr;
27    int    result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
28    if (result != 0) {
29        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
30        return NULL;
31    }
32
33    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
34    buffer->buft                 = buft;
35    buffer->iface.free_buffer    = ggml_backend_cpu_hbm_buffer_free_buffer;
36
37    return buffer;
38}
39
40ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
41    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
42        /* .iface    = */ {
43                           /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
44                           /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
45                           /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
46                           /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
47                           /* .get_alloc_size   = */ nullptr,  // defaults to ggml_nbytes
48                           /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
49                           },
50        /* .context  = */ nullptr,
51    };
52
53    return &ggml_backend_cpu_buffer_type_hbm;
54}
55#endif