Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 01:05:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

sd: sync to master-560-e8323ca (#2082)

Browse Source 
* sd: sync to master-540-f16a110

* tae post-merge fixes

* build fixes

* restore image mask for non-inpainting models

* sd: sync to master-551-99c1de3

* avoid nlohmann/json.hpp include diffs

* Euler A now works on Flux

* sd: sync to master-555-7397dda

avi_writer.h got removed upstream, but I've simply kept the local
copy for now.

* sd: sync to master-558-8afbeb6

* sd: sync to master-560-e8323ca
This commit is contained in:
Wagner Bruna
2026-04-09 03:44:59 -03:00
committed by GitHub
parent 2b1282a664
commit f371bb14d4
59 changed files with 15571 additions and 12384 deletions
Download Patch File Download Diff File Expand all files Collapse all files
otherarch/sdcpp/model.cpp
+27 -51
View File
@@ -176,43 +176,7 @@ uint16_t f8_e4m3_to_f16(uint8_t f8) {
}
uint16_t f8_e5m2_to_f16(uint8_t fp8) {
uint8_t sign = (fp8 >> 7) & 0x1;
uint8_t exponent = (fp8 >> 2) & 0x1F;
uint8_t mantissa = fp8 & 0x3;
uint16_t fp16_sign = sign << 15;
uint16_t fp16_exponent;
uint16_t fp16_mantissa;
if (exponent == 0 && mantissa == 0) { // zero
return fp16_sign;
}
if (exponent == 0x1F) { // NAN and INF
fp16_exponent = 0x1F;
fp16_mantissa = mantissa ? (mantissa << 8) : 0;
return fp16_sign | (fp16_exponent << 10) | fp16_mantissa;
}
if (exponent == 0) { // subnormal numbers
fp16_mantissa = (mantissa << 8);
return fp16_sign | fp16_mantissa;
}
// normal numbers
int16_t true_exponent = (int16_t)exponent - 15 + 15;
if (true_exponent <= 0) {
fp16_exponent = 0;
fp16_mantissa = (mantissa << 8);
} else if (true_exponent >= 0x1F) {
fp16_exponent = 0x1F;
fp16_mantissa = 0;
} else {
fp16_exponent = (uint16_t)true_exponent;
fp16_mantissa = mantissa << 8;
}
return fp16_sign | (fp16_exponent << 10) | fp16_mantissa;
return static_cast<uint16_t>(fp8) << 8;
}
void f8_e4m3_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
@@ -301,7 +265,7 @@ void ModelLoader::add_tensor_storage(const TensorStorage& tensor_storage) {
}
bool is_zip_file(const std::string& file_path) {
struct zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
if (zip == nullptr) {
return false;
}
@@ -467,9 +431,9 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
size_t total_size = 0;
size_t data_offset = gguf_get_data_offset(ctx_gguf_);
for (int i = 0; i < n_tensors; i++) {
std::string name = gguf_get_tensor_name(ctx_gguf_, i);
struct ggml_tensor* dummy = ggml_get_tensor(ctx_meta_, name.c_str());
size_t offset = data_offset + gguf_get_tensor_offset(ctx_gguf_, i);
std::string name = gguf_get_tensor_name(ctx_gguf_, i);
ggml_tensor* dummy = ggml_get_tensor(ctx_meta_, name.c_str());
size_t offset = data_offset + gguf_get_tensor_offset(ctx_gguf_, i);
// LOG_DEBUG("%s", name.c_str());
@@ -828,7 +792,7 @@ struct PickleTensorReader {
}
}
void read_string(const std::string& str, struct zip_t* zip, std::string dir) {
void read_string(const std::string& str, zip_t* zip, std::string dir) {
if (str == "storage") {
read_global_type = true;
} else if (str != "state_dict") {
@@ -1011,7 +975,7 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
file_paths_.push_back(file_path);
size_t file_index = file_paths_.size() - 1;
struct zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
if (zip == nullptr) {
LOG_ERROR("failed to open '%s'", file_path.c_str());
return false;
@@ -1130,10 +1094,12 @@ SDVersion ModelLoader::get_sd_version() {
tensor_storage.name.find("unet.mid_block.resnets.1.") != std::string::npos) {
has_middle_block_1 = true;
}
if (tensor_storage.name.find("model.diffusion_model.output_blocks.3.1.transformer_blocks.1") != std::string::npos) {
if (tensor_storage.name.find("model.diffusion_model.output_blocks.3.1.transformer_blocks.1") != std::string::npos ||
tensor_storage.name.find("unet.up_blocks.1.attentions.0.transformer_blocks.1") != std::string::npos) {
has_output_block_311 = true;
}
if (tensor_storage.name.find("model.diffusion_model.output_blocks.7.1") != std::string::npos) {
if (tensor_storage.name.find("model.diffusion_model.output_blocks.7.1") != std::string::npos ||
tensor_storage.name.find("unet.up_blocks.2.attentions.1") != std::string::npos) {
has_output_block_71 = true;
}
if (tensor_storage.name == "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight" ||
@@ -1371,6 +1337,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
std::atomic<int64_t> memcpy_time_ms(0);
std::atomic<int64_t> copy_to_backend_time_ms(0);
std::atomic<int64_t> convert_time_ms(0);
std::atomic<uint64_t> bytes_processed(0);
int num_threads_to_use = n_threads_p > 0 ? n_threads_p : sd_get_num_physical_cores();
LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
@@ -1437,7 +1404,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
for (int i = 0; i < n_threads; ++i) {
workers.emplace_back([&, file_path, is_zip]() {
std::ifstream file;
struct zip_t* zip = nullptr;
zip_t* zip = nullptr;
if (is_zip) {
zip = zip_open(file_path.c_str(), 0, 'r');
if (zip == nullptr) {
@@ -1582,6 +1549,8 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
t1 = ggml_time_ms();
copy_to_backend_time_ms.fetch_add(t1 - t0);
}
bytes_processed.fetch_add((uint64_t)nbytes_to_read);
}
if (zip != nullptr) {
zip_close(zip);
@@ -1594,8 +1563,12 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
if (current_idx >= file_tensors.size() || failed) {
break;
}
size_t curr_num = total_tensors_processed + current_idx;
pretty_progress(static_cast<int>(curr_num), static_cast<int>(total_tensors_to_process), (ggml_time_ms() - t_start) / 1000.0f / (curr_num + 1e-6f));
size_t curr_num = total_tensors_processed + current_idx;
float elapsed_seconds = (ggml_time_ms() - t_start) / 1000.0f;
pretty_bytes_progress(static_cast<int>(curr_num),
static_cast<int>(total_tensors_to_process),
bytes_processed.load(),
elapsed_seconds);
std::this_thread::sleep_for(std::chrono::milliseconds(200));
}
@@ -1608,7 +1581,10 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
break;
}
total_tensors_processed += file_tensors.size();
pretty_progress(static_cast<int>(total_tensors_processed), static_cast<int>(total_tensors_to_process), (ggml_time_ms() - t_start) / 1000.0f / (total_tensors_processed + 1e-6f));
pretty_bytes_progress(static_cast<int>(total_tensors_processed),
static_cast<int>(total_tensors_to_process),
bytes_processed.load(),
(ggml_time_ms() - t_start) / 1000.0f);
if (total_tensors_processed < total_tensors_to_process) {
printf("\n");
}
@@ -1625,7 +1601,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
return success;
}
bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
bool ModelLoader::load_tensors(std::map<std::string, ggml_tensor*>& tensors,
std::set<std::string> ignore_tensors,
int n_threads,
bool enable_mmap) {
@@ -1639,7 +1615,7 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
tensor_names_in_file.insert(name);
}
struct ggml_tensor* real;
ggml_tensor* real;
if (tensors.find(name) != tensors.end()) {
real = tensors[name];
} else {