happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

remove redundant reshape

This commit is contained in:
Xuan Son Nguyen 2025-12-18 23:35:47 +01:00
parent 4239bf2cde
commit d244fe1a3e
2 changed files with 13 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
convert_hf_to_gguf.py
View File

@ -9904,6 +9904,12 @@ class LFM2AudioModel(MmprojModel):
# reshape conv weights
if name.startswith("conformer.pre_encode.conv.") and name.endswith(".bias"):
data_torch = data_torch[:, None, None]
if "conv.depthwise_conv" in name and name.endswith(".weight"):
assert data_torch.shape[1] == 1
data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[2])
if "conv.pointwise_conv" in name and name.endswith(".weight"):
assert data_torch.shape[2] == 1
data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[1])
return [(self.map_tensor_name(name), data_torch)]

19
tools/mtmd/models/conformer.cpp
View File

@ -158,9 +158,7 @@ ggml_cgraph * clip_graph_conformer::build() {
// conv
{
auto * x = cur;
auto * conv_pw1_w =
ggml_reshape_2d(ctx0, layer.conv_pw1_w, layer.conv_pw1_w->ne[1], layer.conv_pw1_w->ne[2]);
x = ggml_mul_mat(ctx0, conv_pw1_w, x);
x = ggml_mul_mat(ctx0, layer.conv_pw1_w, x);
x = ggml_add(ctx0, x, layer.conv_pw1_b);
cb(x, "conformer.layers.{}.conv.pointwise_conv1", il);
@ -174,20 +172,17 @@ ggml_cgraph * clip_graph_conformer::build() {
}
// use ggml_ssm_conv for f32 precision
x = ggml_pad(ctx0, x, 4, 0, 0, 0);
x = ggml_roll(ctx0, x, 4, 0, 0, 0);
x = ggml_pad(ctx0, x, 4, 0, 0, 0);
auto * conv_dw_w = ggml_reshape_2d(ctx0, layer.conv_dw_w, layer.conv_dw_w->ne[0], layer.conv_dw_w->ne[2]);
x = ggml_ssm_conv(ctx0, x, conv_dw_w);
x = ggml_add(ctx0, x, ggml_reshape_1d(ctx0, layer.conv_dw_b, layer.conv_dw_b->ne[0]));
x = ggml_pad(ctx0, x, 4, 0, 0, 0);
x = ggml_roll(ctx0, x, 4, 0, 0, 0);
x = ggml_pad(ctx0, x, 4, 0, 0, 0);
x = ggml_ssm_conv(ctx0, x, layer.conv_dw_w);
x = ggml_add(ctx0, x, layer.conv_dw_b);
x = ggml_add(ctx0, ggml_mul(ctx0, x, layer.conv_norm_w), layer.conv_norm_b);
x = ggml_silu(ctx0, x);
// pointwise_conv2
auto * conv_pw2_w =
ggml_reshape_2d(ctx0, layer.conv_pw2_w, layer.conv_pw2_w->ne[1], layer.conv_pw2_w->ne[2]);
x = ggml_mul_mat(ctx0, conv_pw2_w, x);
x = ggml_mul_mat(ctx0, layer.conv_pw2_w, x);
x = ggml_add(ctx0, x, layer.conv_pw2_b);
cur = x;