refector: remove some useless code

leejet · Apr 14, 2024 · ec82d52 · ec82d52
1 parent afea457
commit ec82d52
Show file tree

Hide file tree

Showing 10 changed files with 23 additions and 126 deletions.
diff --git a/clip.hpp b/clip.hpp
@@ -939,22 +939,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public GGMLModule {
         return "clip";
     }
 
-    size_t get_params_mem_size() {
-        size_t params_mem_size = text_model.get_params_mem_size();
-        if (version == VERSION_XL) {
-            params_mem_size += text_model2.get_params_mem_size();
-        }
-        return params_mem_size;
-    }
-
-    size_t get_params_num() {
-        size_t params_num = text_model.get_params_num();
-        if (version == VERSION_XL) {
-            params_num += text_model2.get_params_num();
-        }
-        return params_num;
-    }
-
     void set_clip_skip(int clip_skip) {
         text_model.set_clip_skip(clip_skip);
         if (version == VERSION_XL) {
@@ -1385,14 +1369,6 @@ struct FrozenCLIPVisionEmbedder : public GGMLModule {
         return "clip_vision";
     }
 
-    size_t get_params_mem_size() {
-        return vision_model.get_params_mem_size();
-    }
-
-    size_t get_params_num() {
-        return vision_model.get_params_num();
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         vision_model.get_param_tensors(tensors, prefix + "transformer");
     }

diff --git a/control.hpp b/control.hpp
@@ -369,14 +369,6 @@ struct ControlNet : public GGMLModule {
         return "control_net";
     }
 
-    size_t get_params_mem_size() {
-        return control_net.get_params_mem_size();
-    }
-
-    size_t get_params_num() {
-        return control_net.get_params_num();
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         control_net.get_param_tensors(tensors, prefix);
     }

diff --git a/esrgan.hpp b/esrgan.hpp
@@ -152,14 +152,6 @@ struct ESRGAN : public GGMLModule {
         return "esrgan";
     }
 
-    size_t get_params_mem_size() {
-        return rrdb_net.get_params_mem_size();
-    }
-
-    size_t get_params_num() {
-        return rrdb_net.get_params_num();
-    }
-
     bool load_from_file(const std::string& file_path) {
         LOG_INFO("loading esrgan from '%s'", file_path.c_str());
 

diff --git a/ggml_extend.hpp b/ggml_extend.hpp
@@ -752,18 +752,15 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_timestep_embedding(
     return ggml_timestep_embedding(ctx, timesteps, dim, max_period);
 }
 
-// struct GGMLComputeGraph {
-//     virtual void init(struct ggml_context* ctx, ggml_type wtype) = 0;
-//     virtual std::string get_desc() = 0;
-//     virtual size_t get_params_mem_size()   = 0;
-//     virtual size_t get_params_num() = 0;
-//     virtual struct ggml_cgraph* get_ggml_cgraph() = 0;
-// };
-
-/*
-#define MAX_PARAMS_TENSOR_NUM 10240
-#define MAX_GRAPH_SIZE 10240
-*/
+
+__STATIC_INLINE__ size_t ggml_tensor_num(ggml_context * ctx) {
+    size_t num = 0;
+    for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+        num++;
+    }
+    return num;
+}
+
 /* SDXL with LoRA requires more space */
 #define MAX_PARAMS_TENSOR_NUM 15360
 #define MAX_GRAPH_SIZE 15360
@@ -854,8 +851,6 @@ struct GGMLModule {
     }
 
 public:
-    virtual size_t get_params_mem_size() = 0;
-    virtual size_t get_params_num()      = 0;
     virtual std::string get_desc()       = 0;
 
     GGMLModule(ggml_backend_t backend, ggml_type wtype = GGML_TYPE_F32)
@@ -876,7 +871,7 @@ struct GGMLModule {
     }
 
     bool alloc_params_buffer() {
-        size_t num_tensors = get_params_num();
+        size_t num_tensors = ggml_tensor_num(params_ctx);
         params_buffer      = ggml_backend_alloc_ctx_tensors(params_ctx, backend);
         if (params_buffer == NULL) {
             LOG_ERROR("%s alloc params backend buffer failed", get_desc().c_str());
@@ -898,6 +893,13 @@ struct GGMLModule {
         }
     }
 
+    size_t get_params_buffer_size() {
+        if (params_buffer != NULL) {
+            return ggml_backend_buffer_get_size(params_buffer);
+        }
+        return 0;
+    }
+
     void free_compute_buffer() {
         if (compute_allocr != NULL) {
             ggml_gallocr_free(compute_allocr);
@@ -968,19 +970,6 @@ struct GGMLModule {
 };
 
 class GGMLBlock {
-private:
-    static char temp_buffer[1024 * 1024 * 10];
-    ggml_context* get_temp_ctx() {
-        struct ggml_init_params params;
-        params.mem_size   = sizeof(temp_buffer);
-        params.mem_buffer = temp_buffer;
-        params.no_alloc   = true;
-
-        ggml_context* temp_ctx = ggml_init(params);
-        GGML_ASSERT(temp_ctx != NULL);
-        return temp_ctx;
-    }
-
 protected:
     typedef std::unordered_map<std::string, struct ggml_tensor*> ParameterMap;
     typedef std::unordered_map<std::string, std::shared_ptr<GGMLBlock>> GGMLBlockMap;
@@ -1003,14 +992,6 @@ class GGMLBlock {
         init_params(ctx, wtype);
     }
 
-    std::tuple<size_t, size_t> get_params_info(ggml_type wtype) {
-        ggml_context* temp_ctx = get_temp_ctx();
-        init(temp_ctx, wtype);
-        size_t num_tensors = get_params_num();
-        size_t mem_size    = get_params_mem_size();
-        return {num_tensors, mem_size};
-    }
-
     size_t get_params_num() {
         size_t num_tensors = params.size();
         for (auto& pair : blocks) {

diff --git a/lora.hpp b/lora.hpp
@@ -27,14 +27,6 @@ struct LoraModel : public GGMLModule {
         return "lora";
     }
 
-    size_t get_params_num() {
-        return LORA_GRAPH_SIZE;
-    }
-
-    size_t get_params_mem_size() {
-        return model_loader.get_params_mem_size(NULL);
-    }
-
     bool load_from_file(bool filter_tensor = false) {
         LOG_INFO("loading LoRA from '%s'", file_path.c_str());
 

diff --git a/pmid.hpp b/pmid.hpp
@@ -186,16 +186,6 @@ struct PhotoMakerIDEncoder : public GGMLModule {
         return "pmid";
     }
 
-    size_t get_params_mem_size() {
-        size_t params_mem_size = id_encoder.get_params_mem_size();
-        return params_mem_size;
-    }
-
-    size_t get_params_num() {
-        size_t params_num = id_encoder.get_params_num();
-        return params_num;
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         id_encoder.get_param_tensors(tensors, prefix);
     }

diff --git a/stable-diffusion.cpp b/stable-diffusion.cpp
@@ -42,8 +42,6 @@ const char* sampling_methods_str[] = {
     "LCM",
 };
 
-char GGMLBlock::temp_buffer[1024 * 1024 * 10];
-
 /*================================================== Helper Functions ================================================*/
 
 void calculate_alphas_cumprod(float* alphas_cumprod,
@@ -353,27 +351,27 @@ class StableDiffusionGGML {
             // first_stage_model->test();
             // return false;
         } else {
-            size_t clip_params_mem_size = cond_stage_model->get_params_mem_size();
-            size_t unet_params_mem_size = diffusion_model->get_params_mem_size();
+            size_t clip_params_mem_size = cond_stage_model->get_params_buffer_size();
+            size_t unet_params_mem_size = diffusion_model->get_params_buffer_size();
             size_t vae_params_mem_size  = 0;
             if (!use_tiny_autoencoder) {
-                vae_params_mem_size = first_stage_model->get_params_mem_size();
+                vae_params_mem_size = first_stage_model->get_params_buffer_size();
             } else {
                 if (!tae_first_stage->load_from_file(taesd_path)) {
                     return false;
                 }
-                vae_params_mem_size = tae_first_stage->get_params_mem_size();
+                vae_params_mem_size = tae_first_stage->get_params_buffer_size();
             }
             size_t control_net_params_mem_size = 0;
             if (control_net) {
                 if (!control_net->load_from_file(control_net_path)) {
                     return false;
                 }
-                control_net_params_mem_size = control_net->get_params_mem_size();
+                control_net_params_mem_size = control_net->get_params_buffer_size();
             }
             size_t pmid_params_mem_size = 0;
             if (stacked_id) {
-                pmid_params_mem_size = pmid_model->get_params_mem_size();
+                pmid_params_mem_size = pmid_model->get_params_buffer_size();
             }
 
             size_t total_params_ram_size  = 0;

diff --git a/tae.hpp b/tae.hpp
@@ -200,14 +200,6 @@ struct TinyAutoEncoder : public GGMLModule {
         return "taesd";
     }
 
-    size_t get_params_mem_size() {
-        return taesd.get_params_mem_size();
-    }
-
-    size_t get_params_num() {
-        return taesd.get_params_num();
-    }
-
     bool load_from_file(const std::string& file_path) {
         LOG_INFO("loading taesd from '%s'", file_path.c_str());
         alloc_params_buffer();

diff --git a/unet.hpp b/unet.hpp
@@ -543,14 +543,6 @@ struct UNetModel : public GGMLModule {
         return "unet";
     }
 
-    size_t get_params_mem_size() {
-        return unet.get_params_mem_size();
-    }
-
-    size_t get_params_num() {
-        return unet.get_params_num();
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         unet.get_param_tensors(tensors, prefix);
     }

diff --git a/vae.hpp b/vae.hpp
@@ -526,14 +526,6 @@ struct AutoEncoderKL : public GGMLModule {
         return "vae";
     }
 
-    size_t get_params_mem_size() {
-        return ae.get_params_mem_size();
-    }
-
-    size_t get_params_num() {
-        return ae.get_params_num();
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         ae.get_param_tensors(tensors, prefix);
     }