DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 03:24:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: merge LoRAs in F32 when convert changes the tensor type (#2302)

Browse Source
This commit is contained in:
Windyskr authored and GitHub committed 2026-11-09 22:16:04 +08:00
1 parent d22dcda1cf
commit cfdf3cf037
1 file changed
+7 -3
Download Patch File Download Diff File Expand all files Collapse all files
src/convert.cpp
+7 -3
View File
@@ -463,11 +463,15 @@ static bool load_model_into_memory(ModelLoader& model_loader,
const ggml_type export_type = get_export_tensor_type(model_loader, storage, output_type, type_rules); const ggml_type export_type = get_export_tensor_type(model_loader, storage, output_type, type_rules);
const bool quant_src = ggml_is_quantized(storage.type); const bool quant_src = ggml_is_quantized(storage.type);
const bool quant_dst = ggml_is_quantized(export_type); const bool quant_dst = ggml_is_quantized(export_type);
const bool half_src = storage.type == GGML_TYPE_F16 || storage.type == GGML_TYPE_BF16;
if (quant_src && quant_dst) { if (quant_src && quant_dst) {
load_type = GGML_TYPE_F16; load_type = GGML_TYPE_F16;
} else if (quant_src || quant_dst) { } else if (quant_src) {
load_type = quant_src ? export_type : storage.type; load_type = export_type;
} else if (storage.type == GGML_TYPE_F32 || export_type == GGML_TYPE_F32) { } else if (storage.type == GGML_TYPE_F32 || export_type == GGML_TYPE_F32 ||
(half_src && export_type != storage.type)) {
// LoRA deltas can be smaller than half an F16/BF16 ulp of the weight. Merge in F32
// when the export converts the tensor again, so the merged weight is rounded once.
load_type = GGML_TYPE_F32; load_type = GGML_TYPE_F32;
} }
ggml_tensor* tensor = ggml_new_tensor(ctx, load_type, storage.n_dims, storage.ne); ggml_tensor* tensor = ggml_new_tensor(ctx, load_type, storage.n_dims, storage.ne);