Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-10 23:09:24 +02:00
Code Issues Packages Projects Releases Wiki Activity

merge sd fix

Browse Source
This commit is contained in:
Concedo
2025-02-21 11:16:26 +08:00
parent 6d7ef10671
commit 5f74ee3c3b
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
otherarch/sdcpp/lora.hpp
+6 -4
View File
@@ -140,10 +140,12 @@ struct LoraModel : public GGMLRunner {
scale_value *= multiplier;
// flat lora tensors to multiply it
int64_t lora_up_rows = lora_up->ne[ggml_n_dims(lora_up) - 1];
lora_up = ggml_reshape_2d(compute_ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
int64_t lora_down_rows = lora_down->ne[ggml_n_dims(lora_down) - 1];
lora_down = ggml_reshape_2d(compute_ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
int64_t lora_up_rows = lora_up->ne[ggml_n_dims(lora_up) - 1];
lora_up = ggml_reshape_2d(compute_ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
auto lora_down_n_dims = ggml_n_dims(lora_down);
// assume n_dims should always be a multiple of 2 (otherwise rank 1 doesn't work)
lora_down_n_dims = (lora_down_n_dims + lora_down_n_dims % 2);
int64_t lora_down_rows = lora_down->ne[lora_down_n_dims - 1];
// ggml_mul_mat requires tensor b transposed
lora_down = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, lora_down));