upstream-archive/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-08-03 16:50:45 -05:00
Code Issues Packages Projects Releases Wiki Activity

perf: --eager-load to pre-load params at model-load time (#1687)

Browse Source
This commit is contained in:
fszontagh
2026-06-22 16:10:09 +02:00
committed by GitHub
parent b12098f5d0
commit 787d229d84
6 changed files with 34 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
src/model_manager.cpp
View File

@@ -147,6 +147,17 @@ bool ModelManager::register_param_tensors(const std::string& desc,
return true;
}
bool ModelManager::load_all_params_eagerly() {
std::vector<TensorState*> all_states;
all_states.reserve(tensor_states_.size());
for (const auto& s : tensor_states_) {
if (s != nullptr) {
all_states.push_back(s.get());
}
}
return load_tensors_to_params_backend(all_states);
}
bool ModelManager::validate_registered_tensors() {
bool ok = true;
for (const auto& state : tensor_states_) {

1
src/model_manager.h
View File

@@ -158,6 +158,7 @@ public:
}
bool validate_registered_tensors();
bool load_all_params_eagerly();
bool prepare_params(const std::vector<ggml_tensor*>& tensors) override;
void release_compute_backend_params(const std::vector<ggml_tensor*>& tensors) override;

15
src/stable-diffusion.cpp
View File

@@ -199,6 +199,7 @@ public:
bool enable_mmap = false;
sd::ggml_graph_cut::MaxVramAssignment max_vram_assignment;
bool stream_layers = false;
bool eager_load = false;
std::string backend_spec;
std::string params_backend_spec;
@@ -342,6 +343,7 @@ public:
n_threads = sd_ctx_params->n_threads;
enable_mmap = sd_ctx_params->enable_mmap;
stream_layers = sd_ctx_params->stream_layers;
eager_load = sd_ctx_params->eager_load;
backend_spec = SAFE_STR(sd_ctx_params->backend);
params_backend_spec = SAFE_STR(sd_ctx_params->params_backend);
max_vram_assignment.reset(0.f);
@@ -1153,7 +1155,15 @@ public:
return false;
}
LOG_DEBUG("model metadata validated; weights will be prepared lazily");
if (eager_load) {
if (!model_manager->load_all_params_eagerly()) {
LOG_ERROR("model params eager load failed");
return false;
}
LOG_DEBUG("model metadata validated; weights pre-loaded to params backend");
} else {
LOG_DEBUG("model metadata validated; weights will be prepared lazily");
}
{
size_t total_params_ram_size = 0;
@@ -2696,6 +2706,7 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
sd_ctx_params->lora_apply_mode = LORA_APPLY_AUTO;
sd_ctx_params->max_vram = nullptr;
sd_ctx_params->stream_layers = false;
sd_ctx_params->eager_load = false;
sd_ctx_params->enable_mmap = false;
sd_ctx_params->diffusion_flash_attn = false;
sd_ctx_params->circular_x = false;
@@ -2742,6 +2753,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
"prediction: %s\n"
"max_vram: %s\n"
"stream_layers: %s\n"
"eager_load: %s\n"
"backend: %s\n"
"params_backend: %s\n"
"flash_attn: %s\n"
@@ -2777,6 +2789,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
sd_prediction_name(sd_ctx_params->prediction),
SAFE_STR(sd_ctx_params->max_vram),
BOOL_STR(sd_ctx_params->stream_layers),
BOOL_STR(sd_ctx_params->eager_load),
SAFE_STR(sd_ctx_params->backend),
SAFE_STR(sd_ctx_params->params_backend),
BOOL_STR(sd_ctx_params->flash_attn),