root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-11-15 15:29:53 +00:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

llama : fix Mamba-2 conv state saving

Browse Source 
* ggml : make the ggml_mul fast broadcast path more consistently formatted
This commit is contained in:
Francis Couture-Harpin 2024-08-21 16:28:07 -04:00
parent 2bfe9de6d3
commit aff96920f9
2 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml.c
View File

@ -10226,11 +10226,11 @@ static void ggml_compute_forward_mul_f32(
if (scale == 0.0f) { if (scale == 0.0f) {
// NOTE: this also sets NANs to zero, which is not compliant with IEEE754, // NOTE: this also sets NANs to zero, which is not compliant with IEEE754,
// but it is useful when resetting the state of recurrent models. // but it is useful when resetting the state of recurrent models.
memset((char *)dst->data + ir*nb1, 0, nb1); memset((char *) dst->data + ir*nb1, 0, ne0 * sizeof(float));
} else { } else {
if (dst->data != src0->data) { if (dst->data != src0->data) {
// src0 is same shape as dst => same indices // src0 is same shape as dst => same indices
memcpy((char *)dst->data + ir*nb1, (char *)src0->data + ir*nb01, ne0 * sizeof(float)); memcpy((char *) dst->data + ir*nb1, (char *) src0->data + ir*nb01, ne0 * sizeof(float));
} }
if (scale != 1.0f) { if (scale != 1.0f) {
ggml_vec_scale_f32(ne0, (float *) ((char *) dst->data + ir*nb1), scale); ggml_vec_scale_f32(ne0, (float *) ((char *) dst->data + ir*nb1), scale);

2
src/llama.cpp
View File

@ -9335,7 +9335,7 @@ static struct ggml_tensor * llm_build_mamba2(
ggml_cpy(ctx, last_conv, ggml_cpy(ctx, last_conv,
ggml_view_1d(ctx, conv_states_all, ggml_view_1d(ctx, conv_states_all,
(d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs), (d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs),
kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all)))); kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all))));
// 1D convolution // 1D convolution
// The equivalent is to make a self-overlapping view of conv_x // The equivalent is to make a self-overlapping view of conv_x