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

fix: do not force using f32 for some flux layers

Browse Source 
This sometimes leads to worse result
This commit is contained in:
leejet 2024-08-25 14:07:22 +08:00
parent 1bdc767aaf
commit f5997a1951
1 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
flux.hpp
View File

@ -13,7 +13,7 @@ namespace Flux {
struct MLPEmbedder : public UnaryBlock { struct MLPEmbedder : public UnaryBlock {
public: public:
MLPEmbedder(int64_t in_dim, int64_t hidden_dim) { MLPEmbedder(int64_t in_dim, int64_t hidden_dim) {
blocks["in_layer"] = std::shared_ptr<GGMLBlock>(new Linear(in_dim, hidden_dim, true, true)); blocks["in_layer"] = std::shared_ptr<GGMLBlock>(new Linear(in_dim, hidden_dim, true));
blocks["out_layer"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_dim, hidden_dim, true)); blocks["out_layer"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_dim, hidden_dim, true));
} }
@ -449,7 +449,7 @@ namespace Flux {
int64_t patch_size, int64_t patch_size,
int64_t out_channels) { int64_t out_channels) {
blocks["norm_final"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false)); blocks["norm_final"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
blocks["linear"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, patch_size * patch_size * out_channels, true, true)); blocks["linear"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, patch_size * patch_size * out_channels));
blocks["adaLN_modulation.1"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, 2 * hidden_size)); blocks["adaLN_modulation.1"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, 2 * hidden_size));
} }
@ -634,13 +634,13 @@ namespace Flux {
int64_t out_channels = params.in_channels; int64_t out_channels = params.in_channels;
int64_t pe_dim = params.hidden_size / params.num_heads; int64_t pe_dim = params.hidden_size / params.num_heads;
blocks["img_in"] = std::shared_ptr<GGMLBlock>(new Linear(params.in_channels, params.hidden_size, true, true)); blocks["img_in"] = std::shared_ptr<GGMLBlock>(new Linear(params.in_channels, params.hidden_size, true));
blocks["time_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(256, params.hidden_size)); blocks["time_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(256, params.hidden_size));
blocks["vector_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(params.vec_in_dim, params.hidden_size)); blocks["vector_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(params.vec_in_dim, params.hidden_size));
if (params.guidance_embed) { if (params.guidance_embed) {
blocks["guidance_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(256, params.hidden_size)); blocks["guidance_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(256, params.hidden_size));
} }
blocks["txt_in"] = std::shared_ptr<GGMLBlock>(new Linear(params.context_in_dim, params.hidden_size, true, true)); blocks["txt_in"] = std::shared_ptr<GGMLBlock>(new Linear(params.context_in_dim, params.hidden_size, true));
for (int i = 0; i < params.depth; i++) { for (int i = 0; i < params.depth; i++) {
blocks["double_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new DoubleStreamBlock(params.hidden_size, blocks["double_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new DoubleStreamBlock(params.hidden_size,