char-fork/llama.cpp

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

Zijun YuOpenVINO: Qwen3.5, memory optimization, and test-recurrent-state-rollback (#26952)aee56b3ab

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#include "ggml.h"
5
6#include <memory>
7#include <openvino/op/broadcast.hpp>
8#include <openvino/op/constant.hpp>
9#include <openvino/op/divide.hpp>
10#include <openvino/op/shape_of.hpp>
11#include <openvino/op/tile.hpp>
12#include <vector>
13
14namespace ov {
15namespace frontend {
16namespace ggml {
17namespace op {
18
19// GGML_OP_REPEAT tiles src[0] to fill the destination shape. Every destination
20// dimension is an integer multiple of the corresponding source dimension.
21OutputVector translate_repeat(const NodeContext & context) {
22    num_inputs_check(context, 1, 2);
23
24    auto input = process_view_input_new(context, 0);
25
26    const auto input_shape = context.get_input_shape(0).to_shape();
27    const auto output_shape = context.get_output_shape().to_shape();
28
29    std::vector<int64_t> repeats(4, 1);
30    for (size_t axis = 0; axis < 4; ++axis) {
31        const int64_t input_dim = input_shape[axis];
32        const int64_t output_dim = output_shape[axis];
33
34        FRONT_END_OP_CONVERSION_CHECK(input_dim > 0 && output_dim > 0 && output_dim % input_dim == 0,
35                                      "REPEAT input shape ", input_shape, " cannot tile to match ", output_shape);
36
37        repeats[axis] = output_dim / input_dim;
38    }
39
40    auto repeats_node = ov::op::v0::Constant::create(ov::element::i64, {repeats.size()}, repeats);
41    ov::Output<ov::Node> res = std::make_shared<ov::op::v0::Tile>(input, repeats_node);
42    return rename_outputs_with_suffix({res}, context.get_name());
43}
44
45}  // namespace op
46}  // namespace ggml
47}  // namespace frontend
48}  // namespace ov