DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 19:36:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

refactor: route all runner params through model manager (#1649)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2026-06-14 02:05:23 +08:00
1 parent 9b0fceb41b
commit 8d4c7af95b
36 files changed
+596 -677
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/model/te/clip.hpp
+6 -6
View File
@@ -1,4 +1,4 @@
#ifndef __SD_MODEL_TE_CLIP_HPP__
#ifndef __SD_MODEL_TE_CLIP_HPP__
#define __SD_MODEL_TE_CLIP_HPP__
#include "core/ggml_extend.hpp"
@@ -469,13 +469,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
std::vector<float> attention_mask_vec;
CLIPTextModelRunner(ggml_backend_t backend,
ggml_backend_t params_backend,
const String2TensorStorage& tensor_storage_map,
const std::string prefix,
CLIPVersion version = OPENAI_CLIP_VIT_L_14,
bool with_final_ln = true,
bool force_clip_f32 = false)
: GGMLRunner(backend, params_backend) {
CLIPVersion version = OPENAI_CLIP_VIT_L_14,
bool with_final_ln = true,
bool force_clip_f32 = false,
std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
: GGMLRunner(backend, weight_manager) {
bool proj_in = false;
for (const auto& [name, tensor_storage] : tensor_storage_map) {
if (!starts_with(name, prefix)) {
src/model/te/llm.hpp
+22 -31
View File
@@ -1,4 +1,4 @@
#ifndef __SD_MODEL_TE_LLM_HPP__
#ifndef __SD_MODEL_TE_LLM_HPP__
#define __SD_MODEL_TE_LLM_HPP__
#include <algorithm>
@@ -22,6 +22,7 @@
#include "json.hpp"
#include "model/common/rope.hpp"
#include "model_loader.h"
#include "model_manager.h"
#include "tokenizers/bpe_tokenizer.h"
#include "tokenizers/gemma_tokenizer.h"
#include "tokenizers/gpt_oss_tokenizer.h"
@@ -1571,11 +1572,11 @@ namespace LLM {
public:
LLMRunner(LLMArch arch,
ggml_backend_t backend,
ggml_backend_t params_backend,
const String2TensorStorage& tensor_storage_map,
const std::string prefix,
bool enable_vision_ = false)
: GGMLRunner(backend, params_backend),
bool enable_vision_ = false,
std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
: GGMLRunner(backend, weight_manager),
config(LLMConfig::detect_from_weights(tensor_storage_map, prefix, arch)),
enable_vision(enable_vision_) {
if (enable_vision && !config.have_vision_weight) {
@@ -1822,11 +1823,11 @@ namespace LLM {
LLMEmbedder(LLMArch arch,
ggml_backend_t backend,
ggml_backend_t params_backend,
const String2TensorStorage& tensor_storage_map = {},
const std::string prefix = "",
bool enable_vision = false)
: model(arch, backend, params_backend, tensor_storage_map, prefix, enable_vision) {
const String2TensorStorage& tensor_storage_map = {},
const std::string prefix = "",
bool enable_vision = false,
std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
: model(arch, backend, tensor_storage_map, prefix, enable_vision, weight_manager) {
if (arch == LLMArch::MISTRAL_SMALL_3_2 || arch == LLMArch::MINISTRAL_3_3B) {
tokenizer = std::make_shared<MistralTokenizer>();
} else if (arch == LLMArch::GPT_OSS_20B) {
@@ -1840,13 +1841,6 @@ namespace LLM {
model.get_param_tensors(tensors, prefix);
}
bool alloc_params_buffer() {
if (!model.alloc_params_buffer()) {
return false;
}
return true;
}
std::tuple<std::vector<int>, std::vector<float>> tokenize(std::string text,
std::pair<int, int> attn_range,
size_t max_length = 0,
@@ -2062,7 +2056,8 @@ namespace LLM {
ggml_backend_t backend = sd_backend_cpu_init();
ggml_type model_data_type = GGML_TYPE_COUNT;
ModelLoader model_loader;
auto model_manager = std::make_shared<ModelManager>();
ModelLoader& model_loader = model_manager->loader();
if (!model_loader.init_from_file_and_convert_name(file_path, "text_encoders.llm.")) {
LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
return;
@@ -2080,24 +2075,20 @@ namespace LLM {
LLMArch arch = LLMArch::QWEN3;
std::shared_ptr<LLMEmbedder> llm = std::make_shared<LLMEmbedder>(arch,
backend,
backend,
tensor_storage_map,
"text_encoders.llm",
true);
true,
model_manager);
if (!llm->alloc_params_buffer()) {
LOG_ERROR("llm model allocation failed");
return;
}
std::map<std::string, ggml_tensor*> tensors;
llm->get_param_tensors(tensors, "text_encoders.llm");
bool success = model_loader.load_tensors(tensors);
if (!success) {
LOG_ERROR("load tensors from model loader failed");
if (!model_manager->register_runner_params("LLM test",
*llm,
"text_encoders.llm",
ModelManager::ResidencyMode::Resident,
backend,
backend) ||
!model_manager->validate_registered_tensors()) {
LOG_ERROR("register llm tensors with model manager failed");
return;
}
src/model/te/t5.hpp
+21 -29
View File
@@ -1,4 +1,4 @@
#ifndef __SD_MODEL_TE_T5_HPP__
#ifndef __SD_MODEL_TE_T5_HPP__
#define __SD_MODEL_TE_T5_HPP__
#include <cfloat>
@@ -12,6 +12,7 @@
#include "core/ggml_extend.hpp"
#include "model_loader.h"
#include "model_manager.h"
#include "tokenizers/t5_unigram_tokenizer.h"
struct T5Config {
@@ -334,11 +335,11 @@ struct T5Runner : public GGMLRunner {
std::vector<int> relative_position_bucket_vec;
T5Runner(ggml_backend_t backend,
ggml_backend_t params_backend,
const String2TensorStorage& tensor_storage_map,
const std::string prefix,
bool is_umt5 = false)
: GGMLRunner(backend, params_backend),
bool is_umt5 = false,
std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
: GGMLRunner(backend, weight_manager),
config(T5Config::detect_from_weights(tensor_storage_map, prefix, is_umt5)) {
model = T5(config);
model.init(params_ctx, tensor_storage_map, prefix);
@@ -477,24 +478,17 @@ struct T5Embedder {
T5Runner model;
T5Embedder(ggml_backend_t backend,
ggml_backend_t params_backend,
const String2TensorStorage& tensor_storage_map = {},
const std::string prefix = "",
bool is_umt5 = false)
: model(backend, params_backend, tensor_storage_map, prefix, is_umt5), tokenizer(is_umt5) {
const String2TensorStorage& tensor_storage_map = {},
const std::string prefix = "",
bool is_umt5 = false,
std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
: model(backend, tensor_storage_map, prefix, is_umt5, weight_manager), tokenizer(is_umt5) {
}
void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
model.get_param_tensors(tensors, prefix);
}
bool alloc_params_buffer() {
if (!model.alloc_params_buffer()) {
return false;
}
return true;
}
std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
size_t max_length = 0,
bool padding = false) {
@@ -579,7 +573,8 @@ struct T5Embedder {
ggml_backend_t backend = sd_backend_cpu_init();
ggml_type model_data_type = GGML_TYPE_F16;
ModelLoader model_loader;
auto model_manager = std::make_shared<ModelManager>();
ModelLoader& model_loader = model_manager->loader();
if (!model_loader.init_from_file_and_convert_name(file_path)) {
LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
return;
@@ -592,19 +587,16 @@ struct T5Embedder {
}
}
std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, backend, tensor_storage_map, "", true);
std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, tensor_storage_map, "", true, model_manager);
if (!t5->alloc_params_buffer()) {
LOG_ERROR("t5 params buffer allocation failed");
return;
}
std::map<std::string, ggml_tensor*> tensors;
t5->get_param_tensors(tensors, "");
bool success = model_loader.load_tensors(tensors);
if (!success) {
LOG_ERROR("load tensors from model loader failed");
if (!model_manager->register_runner_params("T5 test",
*t5,
"",
ModelManager::ResidencyMode::Resident,
backend,
backend) ||
!model_manager->validate_registered_tensors()) {
LOG_ERROR("register t5 tensors with model manager failed");
return;
}