DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2025-12-12 21:38:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: avoid generating black images when running T5 on the GPU (#882)

Browse Source
This commit is contained in:
leejet 2025-10-13 00:01:06 +08:00 committed by GitHub
parent 9727c6bb98
commit 1c32fa03bc
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 4 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
stable-diffusion.cpp
View File

@ -338,17 +338,7 @@ public:
bool clip_on_cpu = sd_ctx_params->keep_clip_on_cpu; bool clip_on_cpu = sd_ctx_params->keep_clip_on_cpu;
{ {
clip_backend = backend; clip_backend = backend;
bool use_t5xxl = false;
if (sd_version_is_dit(version) && !sd_version_is_qwen_image(version)) {
use_t5xxl = true;
}
if (!clip_on_cpu && !ggml_backend_is_cpu(backend) && use_t5xxl) {
LOG_WARN(
"!!!It appears that you are using the T5 model. Some backends may encounter issues with it."
"If you notice that the generated images are completely black,"
"try running the T5 model on the CPU using the --clip-on-cpu parameter.");
}
if (clip_on_cpu && !ggml_backend_is_cpu(backend)) { if (clip_on_cpu && !ggml_backend_is_cpu(backend)) {
LOG_INFO("CLIP: Using CPU backend"); LOG_INFO("CLIP: Using CPU backend");
clip_backend = ggml_backend_cpu_init(); clip_backend = ggml_backend_cpu_init();

4
t5.hpp
View File

@ -504,7 +504,9 @@ public:
T5DenseGatedActDense(int64_t model_dim, int64_t ff_dim) { T5DenseGatedActDense(int64_t model_dim, int64_t ff_dim) {
blocks["wi_0"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false)); blocks["wi_0"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
blocks["wi_1"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false)); blocks["wi_1"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false)); float scale = 1.f / 32.f;
// The purpose of the scale here is to prevent NaN issues on some backends(CUDA, ...).
blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false, false, false, scale));
} }
struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) { struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {