Skip to content

refactor: remove obsolete unused tensor filtering #1984

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
leejet merged 1 commit into from
Sep 18, 2026
+2 −80
Merged

refactor: remove obsolete unused tensor filtering #1984

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion src/core/backend_fit.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,7 +102,7 @@ namespace sd::backend_fit {
for (const auto& [name, stored_tensor] : loader.get_tensor_storage_map()) {
TensorStorage ts = stored_tensor;
ComponentKind kind;
if (is_unused_tensor(ts.name) || !classify_tensor(ts.name, kind)) {
if (!classify_tensor(ts.name, kind)) {
continue;
}
if (ts.expected_type != GGML_TYPE_COUNT) {
Expand Down
2 changes: 1 addition & 1 deletion src/model/adapter/lora.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,7 +67,7 @@ struct LoraModel : public GGMLRunner {
std::map<std::string, ggml_tensor*> scalars;
std::set<std::string> scalar_names;
for (const auto& [name, source] : sources) {
if (is_unused_tensor(name) || (filter && !filter(name)))
if (filter && !filter(name))
continue;
const bool scalar = source.nelements() == 1 && (ends_with(name, ".alpha") || ends_with(name, ".scale"));
auto* tensor = ggml_new_tensor(params_ctx, scalar ? GGML_TYPE_F32 : source.type, source.n_dims, source.ne);
Expand Down
76 changes: 0 additions & 76 deletions src/model_loader.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,52 +35,6 @@

/*================================================= Preprocess ==================================================*/

const char* unused_tensors[] = {
"betas",
"alphas_cumprod_prev",
"sqrt_alphas_cumprod",
"sqrt_one_minus_alphas_cumprod",
"log_one_minus_alphas_cumprod",
"sqrt_recip_alphas_cumprod",
"sqrt_recipm1_alphas_cumprod",
"posterior_variance",
"posterior_log_variance_clipped",
"posterior_mean_coef1",
"posterior_mean_coef2",
"cond_stage_model.transformer.text_model.embeddings.position_ids",
"cond_stage_model.1.model.text_model.embeddings.position_ids",
"cond_stage_model.transformer.vision_model.embeddings.position_ids",
"cond_stage_model.model.logit_scale",
"conditioner.embedders.0.transformer.text_model.embeddings.position_ids",
"conditioner.embedders.0.model.logit_scale",
"conditioner.embedders.1.model.logit_scale",
"model.diffusion_model.time_embedding.cond_proj.weight",
"unet.time_embedding.cond_proj.weight",
"model_ema.decay",
"model_ema.num_updates",
"model_ema.diffusion_model",
"embedding_manager",
"denoiser.sigmas",
"text_encoders.t5xxl.transformer.encoder.embed_tokens.weight", // only used during training
"ztsnr", // Found in some SDXL vpred models
"edm_vpred.sigma_min", // Found in CosXL
// TODO: find another way to avoid the "unknown tensor" for these two
// "edm_vpred.sigma_max", // Used to detect CosXL
// "v_pred", // Used to detect SDXL vpred models
"text_encoders.llm.output.weight",
"text_encoders.llm.lm_head.",
"language_model.lm_head.",
};

bool is_unused_tensor(const std::string& name) {
for (size_t i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
if (starts_with(name, unused_tensors[i])) {
return true;
}
}
return false;
}

void f64_to_f32_vec(double* src, float* dst, int64_t n) {
// support inplace op
for (int64_t i = 0; i < n; i++) {
Expand Down Expand Up @@ -284,10 +238,6 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
size_t file_index = add_file_path(file_path);

for (auto& tensor_storage : tensor_storages) {
if (is_unused_tensor(tensor_storage.name)) {
continue;
}

if (!starts_with(tensor_storage.name, prefix)) {
tensor_storage.name = prefix + tensor_storage.name;
}
Expand Down Expand Up @@ -356,10 +306,6 @@ bool ModelLoader::init_from_torch_legacy_file(const std::string& file_path, cons
size_t file_index = add_file_path(file_path);

for (auto& tensor_storage : tensor_storages) {
if (is_unused_tensor(tensor_storage.name)) {
continue;
}

if (!starts_with(tensor_storage.name, prefix)) {
tensor_storage.name = prefix + tensor_storage.name;
}
Expand Down Expand Up @@ -674,10 +620,6 @@ SDVersion ModelLoader::get_sd_version() const {
std::map<ggml_type, uint32_t> ModelLoader::get_wtype_stat() const {
std::map<ggml_type, uint32_t> wtype_stat;
for (auto& [name, tensor_storage] : tensor_storage_map) {
if (is_unused_tensor(tensor_storage.name)) {
continue;
}

auto iter = wtype_stat.find(tensor_storage.type);
if (iter != wtype_stat.end()) {
iter->second++;
Expand All @@ -691,10 +633,6 @@ std::map<ggml_type, uint32_t> ModelLoader::get_wtype_stat() const {
std::map<ggml_type, uint32_t> ModelLoader::get_conditioner_wtype_stat() const {
std::map<ggml_type, uint32_t> wtype_stat;
for (auto& [name, tensor_storage] : tensor_storage_map) {
if (is_unused_tensor(tensor_storage.name)) {
continue;
}

if ((tensor_storage.name.find("text_encoders") == std::string::npos &&
tensor_storage.name.find("cond_stage_model") == std::string::npos &&
tensor_storage.name.find("te.text_model.") == std::string::npos &&
Expand All @@ -715,10 +653,6 @@ std::map<ggml_type, uint32_t> ModelLoader::get_conditioner_wtype_stat() const {
std::map<ggml_type, uint32_t> ModelLoader::get_diffusion_model_wtype_stat() const {
std::map<ggml_type, uint32_t> wtype_stat;
for (auto& [name, tensor_storage] : tensor_storage_map) {
if (is_unused_tensor(tensor_storage.name)) {
continue;
}

if (tensor_storage.name.find("model.diffusion_model.") == std::string::npos && tensor_storage.name.find("unet.") == std::string::npos) {
continue;
}
Expand All @@ -736,10 +670,6 @@ std::map<ggml_type, uint32_t> ModelLoader::get_diffusion_model_wtype_stat() cons
std::map<ggml_type, uint32_t> ModelLoader::get_vae_wtype_stat() const {
std::map<ggml_type, uint32_t> wtype_stat;
for (auto& [name, tensor_storage] : tensor_storage_map) {
if (is_unused_tensor(tensor_storage.name)) {
continue;
}

if (tensor_storage.name.find("vae.") == std::string::npos &&
tensor_storage.name.find("first_stage_model") == std::string::npos) {
continue;
Expand Down Expand Up @@ -823,9 +753,6 @@ void ModelLoader::process_model_files(bool enable_mmap, bool writable_mmap) {

std::vector<TensorStorage> processed_tensor_storages;
for (const auto& [name, tensor_storage] : tensor_storage_map) {
if (is_unused_tensor(tensor_storage.name)) {
continue;
}
processed_tensor_storages.push_back(tensor_storage);
}

Expand Down Expand Up @@ -1551,9 +1478,6 @@ int64_t ModelLoader::get_params_mem_size(ggml_backend_t backend, ggml_type type)
int64_t mem_size = 0;
std::vector<TensorStorage> processed_tensor_storages;
for (auto [name, tensor_storage] : tensor_storage_map) {
if (is_unused_tensor(tensor_storage.name)) {
continue;
}
if (tensor_should_be_converted(tensor_storage, type)) {
tensor_storage.type = type;
}
Expand Down
2 changes: 0 additions & 2 deletions src/model_loader.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,8 +28,6 @@ struct MmapTensorStore {
std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
};

bool is_unused_tensor(const std::string& name);

class ModelLoader {
public:
using FileId = uint64_t;
Expand Down
Loading