alphard/stable-diffusion.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: fix rank 1 loras (#575)

Browse Source
This commit is contained in:
stduhpf 2025-02-05 09:13:17 +01:00 committed by GitHub
parent a3cbdf6dcb
commit e500d95abd
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 6 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
lora.hpp
View File

@ -615,9 +615,12 @@ struct LoraModel : public GGMLRunner {
scale_value *= multiplier; scale_value *= multiplier;
// flat lora tensors to multiply it // flat lora tensors to multiply it
int64_t lora_up_rows = lora_up->ne[ggml_n_dims(lora_up) - 1]; 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); 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]; 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];
lora_down = ggml_reshape_2d(compute_ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows); lora_down = ggml_reshape_2d(compute_ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
// ggml_mul_mat requires tensor b transposed // ggml_mul_mat requires tensor b transposed