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 →
869 B29 linesraw
1#include "../node_context.h"
2#include "../op_table.h"
3#include "../utils.h"
4
5#include <openvino/op/constant.hpp>
6#include <openvino/op/cum_sum.hpp>
7
8namespace ov {
9namespace frontend {
10namespace ggml {
11namespace op {
12
13// GGML cumsum computes prefix sum along dim 0 (the innermost/fastest dimension).
14// In OV layout the dims are reversed: ggml [ne0, ne1, ne2, ne3] → OV [ne3, ne2, ne1, ne0],
15// so ggml dim 0 maps to OV axis 3 (last axis).
16OutputVector translate_cumsum(const NodeContext & context) {
17    num_inputs_check(context, 1, 1);
18
19    auto x    = context.get_input(0);
20    auto axis = ov::op::v0::Constant::create(ov::element::i64, {}, {3});
21    auto res  = std::make_shared<ov::op::v0::CumSum>(x, axis);
22
23    return rename_outputs_with_suffix({res}, context.get_name());
24}
25
26}  // namespace op
27}  // namespace ggml
28}  // namespace frontend
29}  // namespace ov