char-fork/llama.cpp

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

Zijun Yuopenvino: OV 2026.2, context-shift, Q5_1 support, gemma4 dense/embedding, and -fa off (#24503)890f1a27e

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
1.6 KiB48 linesraw
1#include "../node_context.h"
2#include "../op_table.h"
3#include "../utils.h"
4
5#include <memory>
6#include <openvino/frontend/exception.hpp>
7#include <openvino/op/concat.hpp>
8#include <openvino/op/convert.hpp>
9
10namespace ov {
11namespace frontend {
12namespace ggml {
13namespace op {
14
15OutputVector translate_concat(const NodeContext & context) {
16    num_inputs_check(context, 2, 2);
17
18    const int32_t * op_params = context.get_output_op_params();
19    FRONT_END_CHECK_IMPLEMENTED(op_params != nullptr, "CONCAT requires output op params");
20
21    const auto output_shape = context.get_output_shape();
22    FRONT_END_CHECK_IMPLEMENTED(output_shape.rank().is_static(), "CONCAT requires static output rank");
23
24    const auto rank = output_shape.rank().get_length();
25    const int32_t ggml_dim = op_params[0];
26    FRONT_END_CHECK_IMPLEMENTED(ggml_dim >= 0 && ggml_dim < rank, "CONCAT axis is out of range");
27
28    auto input_0 = process_view_input_new(context, 0);
29    auto input_1 = process_view_input_new(context, 1);
30    const auto output_type = context.get_output_type();
31
32    if (input_0.get_element_type() != output_type) {
33        input_0 = std::make_shared<ov::op::v0::Convert>(input_0, output_type);
34    }
35    if (input_1.get_element_type() != output_type) {
36        input_1 = std::make_shared<ov::op::v0::Convert>(input_1, output_type);
37    }
38
39    const auto axis = static_cast<int64_t>(rank - 1 - ggml_dim);
40    auto res = std::make_shared<ov::op::v0::Concat>(OutputVector{input_0, input_1}, axis);
41
42    return rename_outputs_with_suffix({res}, context.get_name());
43}
44
45}  // namespace op
46}  // namespace ggml
47}  // namespace frontend
48}  // namespace ov