mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2026-05-19 08:00:25 +00:00
sd: sync to master-596-90e87bc (#2204)
* sd: reuse source lists between make and cmake * sd: sync to master-596-90e87bc * Update source file path for sdtype_adapter.cpp --------- Co-authored-by: LostRuins Concedo <39025047+LostRuins@users.noreply.github.com>
This commit is contained in:
Wagner Bruna
GitHub
parent
cc82c3164e
commit
bfe9548fd5
29 changed files with 1751 additions and 116 deletions
|
|
@ -148,6 +148,7 @@ public:
|
|||
std::string taesd_path;
|
||||
sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0, 0};
|
||||
bool offload_params_to_cpu = false;
|
||||
float max_vram = 0.f;
|
||||
bool use_pmid = false;
|
||||
|
||||
bool is_using_v_parameterization = false;
|
||||
|
|
@ -208,6 +209,7 @@ public:
|
|||
vae_decode_only = sd_ctx_params->vae_decode_only;
|
||||
free_params_immediately = sd_ctx_params->free_params_immediately;
|
||||
offload_params_to_cpu = sd_ctx_params->offload_params_to_cpu;
|
||||
max_vram = sd_ctx_params->max_vram;
|
||||
|
||||
bool use_tae = false;
|
||||
|
||||
|
|
@ -575,6 +577,10 @@ public:
|
|||
|
||||
bool clip_on_cpu = sd_ctx_params->keep_clip_on_cpu;
|
||||
|
||||
const size_t max_graph_vram_bytes = max_vram <= 0.f
|
||||
? 0
|
||||
: static_cast<size_t>(static_cast<double>(max_vram) * 1024.0 * 1024.0 * 1024.0);
|
||||
|
||||
{
|
||||
clip_backend = backend;
|
||||
if (clip_on_cpu && !ggml_backend_is_cpu(backend)) {
|
||||
|
|
@ -664,6 +670,7 @@ public:
|
|||
clip_vision = std::make_shared<FrozenCLIPVisionEmbedder>(backend,
|
||||
offload_params_to_cpu,
|
||||
tensor_storage_map);
|
||||
clip_vision->set_max_graph_vram_bytes(max_graph_vram_bytes);
|
||||
clip_vision->alloc_params_buffer();
|
||||
clip_vision->get_param_tensors(tensors);
|
||||
}
|
||||
|
|
@ -740,9 +747,11 @@ public:
|
|||
}
|
||||
}
|
||||
|
||||
cond_stage_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
|
||||
cond_stage_model->alloc_params_buffer();
|
||||
cond_stage_model->get_param_tensors(tensors);
|
||||
|
||||
diffusion_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
|
||||
diffusion_model->alloc_params_buffer();
|
||||
diffusion_model->get_param_tensors(tensors);
|
||||
|
||||
|
|
@ -751,6 +760,7 @@ public:
|
|||
}
|
||||
|
||||
if (high_noise_diffusion_model) {
|
||||
high_noise_diffusion_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
|
||||
high_noise_diffusion_model->alloc_params_buffer();
|
||||
high_noise_diffusion_model->get_param_tensors(tensors);
|
||||
}
|
||||
|
|
@ -823,16 +833,19 @@ public:
|
|||
} else if (use_tae && !tae_preview_only) {
|
||||
LOG_INFO("using TAE for encoding / decoding");
|
||||
first_stage_model = create_tae();
|
||||
first_stage_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
|
||||
first_stage_model->alloc_params_buffer();
|
||||
first_stage_model->get_param_tensors(tensors, "tae");
|
||||
} else {
|
||||
LOG_INFO("using VAE for encoding / decoding");
|
||||
first_stage_model = create_vae();
|
||||
first_stage_model->set_max_graph_vram_bytes(max_graph_vram_bytes);
|
||||
first_stage_model->alloc_params_buffer();
|
||||
first_stage_model->get_param_tensors(tensors, "first_stage_model");
|
||||
if (use_tae && tae_preview_only) {
|
||||
LOG_INFO("using TAE for preview");
|
||||
preview_vae = create_tae();
|
||||
preview_vae->set_max_graph_vram_bytes(max_graph_vram_bytes);
|
||||
preview_vae->alloc_params_buffer();
|
||||
preview_vae->get_param_tensors(tensors, "tae");
|
||||
}
|
||||
|
|
@ -1332,8 +1345,13 @@ public:
|
|||
cond_stage_lora_models.push_back(lora);
|
||||
}
|
||||
}
|
||||
auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(cond_stage_lora_models);
|
||||
cond_stage_model->set_weight_adapter(multi_lora_adapter);
|
||||
// Only attach the adapter when there are LoRAs targeting the cond_stage model.
|
||||
// An empty MultiLoraAdapter still routes every linear/conv through
|
||||
// forward_with_lora() instead of the direct kernel path — slower for no benefit.
|
||||
if (!cond_stage_lora_models.empty()) {
|
||||
auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(cond_stage_lora_models);
|
||||
cond_stage_model->set_weight_adapter(multi_lora_adapter);
|
||||
}
|
||||
}
|
||||
if (diffusion_model) {
|
||||
std::vector<std::shared_ptr<LoraModel>> lora_models;
|
||||
|
|
@ -1364,10 +1382,12 @@ public:
|
|||
diffusion_lora_models.push_back(lora);
|
||||
}
|
||||
}
|
||||
auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(diffusion_lora_models);
|
||||
diffusion_model->set_weight_adapter(multi_lora_adapter);
|
||||
if (high_noise_diffusion_model) {
|
||||
high_noise_diffusion_model->set_weight_adapter(multi_lora_adapter);
|
||||
if (!diffusion_lora_models.empty()) {
|
||||
auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(diffusion_lora_models);
|
||||
diffusion_model->set_weight_adapter(multi_lora_adapter);
|
||||
if (high_noise_diffusion_model) {
|
||||
high_noise_diffusion_model->set_weight_adapter(multi_lora_adapter);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -1400,8 +1420,10 @@ public:
|
|||
first_stage_lora_models.push_back(lora);
|
||||
}
|
||||
}
|
||||
auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(first_stage_lora_models);
|
||||
first_stage_model->set_weight_adapter(multi_lora_adapter);
|
||||
if (!first_stage_lora_models.empty()) {
|
||||
auto multi_lora_adapter = std::make_shared<MultiLoraAdapter>(first_stage_lora_models);
|
||||
first_stage_model->set_weight_adapter(multi_lora_adapter);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -2387,6 +2409,7 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
|
|||
sd_ctx_params->prediction = PREDICTION_COUNT;
|
||||
sd_ctx_params->lora_apply_mode = LORA_APPLY_AUTO;
|
||||
sd_ctx_params->offload_params_to_cpu = false;
|
||||
sd_ctx_params->max_vram = 0.f;
|
||||
sd_ctx_params->enable_mmap = false;
|
||||
sd_ctx_params->keep_clip_on_cpu = false;
|
||||
sd_ctx_params->keep_control_net_on_cpu = false;
|
||||
|
|
@ -2428,6 +2451,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
|
|||
"sampler_rng_type: %s\n"
|
||||
"prediction: %s\n"
|
||||
"offload_params_to_cpu: %s\n"
|
||||
"max_vram: %.3f\n"
|
||||
"keep_clip_on_cpu: %s\n"
|
||||
"keep_control_net_on_cpu: %s\n"
|
||||
"keep_vae_on_cpu: %s\n"
|
||||
|
|
@ -2460,6 +2484,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
|
|||
sd_rng_type_name(sd_ctx_params->sampler_rng_type),
|
||||
sd_prediction_name(sd_ctx_params->prediction),
|
||||
BOOL_STR(sd_ctx_params->offload_params_to_cpu),
|
||||
sd_ctx_params->max_vram,
|
||||
BOOL_STR(sd_ctx_params->keep_clip_on_cpu),
|
||||
BOOL_STR(sd_ctx_params->keep_control_net_on_cpu),
|
||||
BOOL_STR(sd_ctx_params->keep_vae_on_cpu),
|
||||
|
|
@ -3677,9 +3702,13 @@ SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* s
|
|||
std::unique_ptr<UpscalerGGML> hires_upscaler;
|
||||
if (request.hires.upscaler == SD_HIRES_UPSCALER_MODEL) {
|
||||
LOG_INFO("hires fix: loading model upscaler from '%s'", request.hires.model_path);
|
||||
hires_upscaler = std::make_unique<UpscalerGGML>(sd_ctx->sd->n_threads,
|
||||
hires_upscaler = std::make_unique<UpscalerGGML>(sd_ctx->sd->n_threads,
|
||||
false,
|
||||
request.hires.upscale_tile_size);
|
||||
const size_t max_graph_vram_bytes = sd_ctx->sd->max_vram <= 0.f
|
||||
? 0
|
||||
: static_cast<size_t>(static_cast<double>(sd_ctx->sd->max_vram) * 1024.0 * 1024.0 * 1024.0);
|
||||
hires_upscaler->set_max_graph_vram_bytes(max_graph_vram_bytes);
|
||||
if (!hires_upscaler->load_from_file(request.hires.model_path,
|
||||
sd_ctx->sd->offload_params_to_cpu,
|
||||
sd_ctx->sd->n_threads)) {
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue