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 →
2.6 KiB61 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/group_conv.hpp>
7#include <openvino/op/reshape.hpp>
8#include <openvino/op/squeeze.hpp>
9#include <openvino/op/transpose.hpp>
10#include <openvino/op/unsqueeze.hpp>
11
12namespace ov {
13namespace frontend {
14namespace ggml {
15namespace op {
16
17OutputVector translate_ssm_conv(const NodeContext & context) {
18    num_inputs_check(context, 2, 2);
19
20    auto sx = context.get_input(0);                         // conv state + input: OV shape [1, n_s, d_inner, ncs]
21    auto c = context.get_input(1);                          // conv1d weight:      OV shape [1, 1, d_inner, d_conv]
22
23    auto sx_shape = context.get_input_shape(0).to_shape();  // [1, n_s, d_inner, ncs]
24    auto c_shape = context.get_input_shape(1).to_shape();   // [1, 1, d_inner, d_conv]
25
26    // int64_t n_s  = sx_shape[1];
27    int64_t d_inner = sx_shape[2];
28    // int64_t ncs  = sx_shape[3];  // d_conv - 1 + n_t
29    int64_t d_conv  = c_shape[3];
30    // int64_t n_t  = ncs - d_conv + 1;
31
32    // Reshape sx from [1, n_s, d_inner, ncs] to [n_s, d_inner, ncs] for 1D GroupConvolution
33    auto sx_reshaped =
34        std::make_shared<ov::op::v0::Squeeze>(sx, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
35
36    // Reshape c from [1, 1, d_inner, d_conv] to [d_inner, 1, 1, d_conv]
37    // GroupConvolution filter: [groups, out_channels/groups, in_channels/groups, kernel_size]
38    auto c_new_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, std::vector<int64_t>{d_inner, 1, 1, d_conv});
39    auto c_reshaped = std::make_shared<ov::op::v1::Reshape>(c, c_new_shape, false);
40
41    // Depthwise 1D convolution: groups=d_inner, stride=1, no padding, no dilation
42    // Input: [n_s, d_inner, ncs], Filter: [d_inner, 1, 1, d_conv]
43    // Output: [n_s, d_inner, n_t]
44    auto conv = std::make_shared<ov::op::v1::GroupConvolution>(
45        sx_reshaped, c_reshaped, ov::Strides{1}, ov::CoordinateDiff{0}, ov::CoordinateDiff{0}, ov::Strides{1});
46
47    // Transpose from [n_s, d_inner, n_t] to [n_s, n_t, d_inner]
48    auto perm = ov::op::v0::Constant::create(ov::element::i64, {3}, std::vector<int64_t>{0, 2, 1});
49    auto transposed = std::make_shared<ov::op::v1::Transpose>(conv, perm);
50
51    // Reshape to output shape [1, n_s, n_t, d_inner]
52    auto res =
53        std::make_shared<ov::op::v0::Unsqueeze>(transposed, ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
54
55    return rename_outputs_with_suffix({res}, context.get_name());
56}
57
58}  // namespace op
59}  // namespace ggml
60}  // namespace frontend
61}  // namespace ov