char-fork/llama.cpp

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

Diego Devesaggml : fix fallback to CPU for ununsupported ops (#15118)0d8831543

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
1.2 KiB36 linesraw
1#include "traits.h"
2
3#include "ggml-backend-impl.h"
4#include "ggml-backend.h"
5
6namespace ggml::cpu {
7tensor_traits::~tensor_traits() {}
8
9extra_buffer_type::~extra_buffer_type() {}
10}  // namespace ggml::cpu
11
12bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
13    for (auto extra : ggml_backend_cpu_get_extra_buffer_types()) {
14        if (extra && extra->context) {
15            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
16            auto tensor_traits = buf_extra->get_tensor_traits(op);
17            if (tensor_traits && tensor_traits->compute_forward(params, op)) {
18                return true;
19            }
20        }
21    }
22    return false;
23}
24
25bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
26    for (auto extra : ggml_backend_cpu_get_extra_buffer_types()) {
27        if (extra && extra->context) {
28            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
29            auto tensor_traits = buf_extra->get_tensor_traits(op);
30            if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
31                return true;
32            }
33        }
34    }
35    return false;
36}