feat: Control Net support + Textual Inversion (embeddings)

.gitignore

@@ -10,5 +10,4 @@ test/
 *.gguf
 output*.png
 models*
-!taesd-model.gguf
 *.log

CMakeLists.txt

@@ -28,17 +28,13 @@ option(SD_CUBLAS                     "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
 option(SD_FLASH_ATTN                 "sd: use flash attention for x4 less memory usage" OFF)
-option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(BUILD_SHARED_LIBS             "sd: build shared libs" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
 if(SD_CUBLAS)
 	message("Use CUBLAS as backend stable-diffusion")
     set(GGML_CUBLAS ON)
     add_definitions(-DSD_USE_CUBLAS)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
-    endif()
 endif()
 
 if(SD_METAL)

README.md

@@ -31,6 +31,7 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 - Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
 - Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
 - VAE tiling processing for reduce memory usage
+- Control Net support with SD 1.5
 - Sampling method
     - `Euler A`
     - `Euler`
@@ -53,9 +54,7 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 - [ ] More sampling methods
 - [ ] Make inference faster
     - The current implementation of ggml_conv_2d is slow and has high memory usage
-    - Implement Winograd Convolution 2D for 3x3 kernel filtering
 - [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
-- [ ] Implement Textual Inversion (embeddings)
 - [ ] Implement Inpainting support
 - [ ] k-quants support
 
@@ -159,16 +158,20 @@ arguments:
   -m, --model [MODEL]                path to model
   --vae [VAE]                        path to vae
   --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
+  --control-net [CONTROL_PATH]       path to control net model
+  --embd-dir [EMBEDDING_PATH]        path to embeddings.
   --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now.
   --type [TYPE]                      weight type (f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0)
                                      If not specified, the default is the type of the weight file.
   --lora-model-dir [DIR]             lora model directory
   -i, --init-img [IMAGE]             path to the input image, required by img2img
+  --control-image [IMAGE]            path to image condition, control net
   -o, --output OUTPUT                path to write result image to (default: ./output.png)
   -p, --prompt [PROMPT]              the prompt to render
   -n, --negative-prompt PROMPT       the negative prompt (default: "")
   --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)
   --strength STRENGTH                strength for noising/unnoising (default: 0.75)
+  --control-strength STRENGTH        strength to apply Control Net (default: 0.9)
                                      1.0 corresponds to full destruction of information in init image
   -H, --height H                     image height, in pixel space (default: 512)
   -W, --width W                      image width, in pixel space (default: 512)
@@ -182,6 +185,7 @@ arguments:
   --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)
                                      <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   --vae-tiling                       process vae in tiles to reduce memory usage
+  --control-net-cpu                  keep controlnet in cpu (for low vram)
   -v, --verbose                      print extra info
 ```
 

clip.hpp

@@ -2,6 +2,7 @@
 #define __CLIP_HPP__
 
 #include "ggml_extend.hpp"
+#include "model.h"
 
 /*================================================== CLIPTokenizer ===================================================*/
 
@@ -67,6 +68,9 @@ std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
 }
 
 // Ref: https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py
+
+typedef std::function<bool(std::string&, std::vector<int32_t>&)> on_new_token_cb_t;
+
 class CLIPTokenizer {
 private:
     SDVersion version = VERSION_1_x;
@@ -234,8 +238,11 @@ class CLIPTokenizer {
         return result;
     }
 
-    std::vector<int> tokenize(std::string text, size_t max_length = 0, bool padding = false) {
-        std::vector<int32_t> tokens = encode(text);
+    std::vector<int> tokenize(std::string text,
+                              on_new_token_cb_t on_new_token_cb,
+                              size_t max_length = 0,
+                              bool padding = false) {
+        std::vector<int32_t> tokens = encode(text, on_new_token_cb);
         tokens.insert(tokens.begin(), BOS_TOKEN_ID);
         if (max_length > 0) {
             if (tokens.size() > max_length - 1) {
@@ -255,7 +262,7 @@ class CLIPTokenizer {
         return tokens;
     }
 
-    std::vector<int> encode(std::string text) {
+    std::vector<int> encode(std::string text, on_new_token_cb_t on_new_token_cb) {
         std::string original_text = text;
         std::vector<int32_t> bpe_tokens;
         text = whitespace_clean(text);
@@ -268,6 +275,10 @@ class CLIPTokenizer {
         std::string str = text;
         std::vector<std::string> token_strs;
         while (std::regex_search(str, matches, pat)) {
+            bool skip = on_new_token_cb(str, bpe_tokens);
+            if (skip) {
+                continue;
+            }
             for (auto& token : matches) {
                 std::string token_str = token.str();
                 std::u32string utf32_token;
@@ -444,13 +455,13 @@ struct ResidualAttentionBlock {
     struct ggml_tensor* ln2_b;  // [hidden_size, ]
 
     size_t calculate_mem_size(ggml_type wtype) {
-        double mem_size = 0;
-        mem_size += 4 * hidden_size * hidden_size * ggml_type_sizef(wtype);        // q_w/k_w/v_w/out_w
-        mem_size += 8 * hidden_size * ggml_type_sizef(GGML_TYPE_F32);              // q_b/k_b/v_b/out_b/ln1_w/ln1_b/ln2_w/ln2_b
-        mem_size += 2 * hidden_size * intermediate_size * ggml_type_sizef(wtype);  // fc1_w/fc2_w
-        mem_size += intermediate_size * ggml_type_sizef(GGML_TYPE_F32);            // fc1_b
-        mem_size += hidden_size * ggml_type_sizef(GGML_TYPE_F32);                  // fc2_b
-        return static_cast<size_t>(mem_size);
+        size_t mem_size = 0;
+        mem_size += 4 * ggml_row_size(wtype, hidden_size * hidden_size);        // q_w/k_w/v_w/out_w
+        mem_size += 8 * ggml_row_size(GGML_TYPE_F32, hidden_size);              // q_b/k_b/v_b/out_b/ln1_w/ln1_b/ln2_w/ln2_b
+        mem_size += 2 * ggml_row_size(wtype, hidden_size * intermediate_size);  // fc1_w/fc2_w
+        mem_size += ggml_row_size(GGML_TYPE_F32, intermediate_size);            // fc1_b
+        mem_size += ggml_row_size(GGML_TYPE_F32, hidden_size);                  // fc2_b
+        return mem_size;
     }
 
     void init_params(struct ggml_context* ctx, ggml_allocr* alloc, ggml_type wtype) {
@@ -597,13 +608,17 @@ struct CLIPTextModel {
     struct ggml_tensor* position_ids;
     struct ggml_tensor* token_embed_weight;
     struct ggml_tensor* position_embed_weight;
+    struct ggml_tensor* token_embed_custom;
 
     // transformer
     std::vector<ResidualAttentionBlock> resblocks;
     struct ggml_tensor* final_ln_w;
     struct ggml_tensor* final_ln_b;
 
     struct ggml_tensor* text_projection;
+    std::string embd_dir;
+    int32_t num_custom_embeddings = 0;
+    std::vector<std::string> readed_embeddings;
 
     CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
                   int clip_skip       = -1,
@@ -642,18 +657,21 @@ struct CLIPTextModel {
     }
 
     size_t calculate_mem_size(ggml_type wtype) {
-        double mem_size = 0;
-        mem_size += hidden_size * max_position_embeddings * ggml_type_sizef(GGML_TYPE_I32);  // position_ids
-        mem_size += hidden_size * vocab_size * ggml_type_sizef(wtype);                       // token_embed_weight
-        mem_size += hidden_size * max_position_embeddings * ggml_type_sizef(wtype);          // position_embed_weight
+        size_t mem_size = 0;
+        mem_size += ggml_row_size(GGML_TYPE_I32, hidden_size * max_position_embeddings);  // position_ids
+        mem_size += ggml_row_size(wtype, hidden_size * vocab_size);                       // token_embed_weight
+        mem_size += ggml_row_size(wtype, hidden_size * max_position_embeddings);          // position_embed_weight
+        if(version == OPENAI_CLIP_VIT_L_14) {
+            mem_size += ggml_row_size(wtype, hidden_size * max_position_embeddings);      // token_embed_custom
+        }
         for (int i = 0; i < num_hidden_layers; i++) {
             mem_size += resblocks[i].calculate_mem_size(wtype);
         }
-        mem_size += 2 * hidden_size * ggml_type_sizef(GGML_TYPE_F32);  // final_ln_w/b
+        mem_size += 2 * ggml_row_size(GGML_TYPE_F32, hidden_size);  // final_ln_w/b
         if (version == OPEN_CLIP_VIT_BIGG_14) {
-            mem_size += hidden_size * projection_dim * ggml_type_sizef(GGML_TYPE_F32);  // text_projection
+            mem_size += ggml_row_size(GGML_TYPE_F32, hidden_size * projection_dim);  // text_projection
         }
-        return static_cast<size_t>(mem_size);
+        return mem_size;
     }
 
     void map_by_name(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
@@ -670,14 +688,48 @@ struct CLIPTextModel {
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx0, struct ggml_tensor* input_ids, size_t max_token_idx = 0, bool return_pooled = false) {
+    bool load_embedding(std::string embd_name, std::string embd_path, std::vector<int32_t> &bpe_tokens) {
+        // the order matters
+        ModelLoader model_loader;
+        if(!model_loader.init_from_file(embd_path)) {
+            LOG_ERROR("embedding '%s' failed", embd_name.c_str());
+            return false;
+        }
+        struct ggml_init_params params;
+        params.mem_size = 32 * 1024; // max for custom embeddings 32 KB
+        params.mem_buffer = NULL;
+        params.no_alloc = false;
+        struct ggml_context* embd_ctx = ggml_init(params);
+        struct ggml_tensor* embd = NULL;
+        auto on_load = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) {
+            if(tensor_storage.ne[0] != hidden_size) {
+                LOG_DEBUG("embedding wrong hidden size, got %i, expected %i", tensor_storage.ne[0], hidden_size);
+                return false;
+            }
+            embd = ggml_new_tensor_2d(embd_ctx, token_embed_weight->type, hidden_size, tensor_storage.n_dims > 1 ? tensor_storage.ne[1] : 1);
+            *dst_tensor = embd;
+            return true;
+        };
+        model_loader.load_tensors(on_load, NULL);
+        ggml_backend_tensor_set(token_embed_custom, embd->data, num_custom_embeddings * hidden_size * ggml_type_size(token_embed_custom->type), ggml_nbytes(embd));
+        readed_embeddings.push_back(embd_name);
+        for(int i = 0; i < embd->ne[1]; i++) {
+            bpe_tokens.push_back(vocab_size + num_custom_embeddings);
+            // LOG_DEBUG("new custom token: %i", vocab_size + num_custom_embeddings);
+            num_custom_embeddings++;
+        }
+        LOG_DEBUG("embedding '%s' applied, custom embeddings: %i", embd_name.c_str(), num_custom_embeddings);
+        return true;
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx0, struct ggml_tensor* input_ids, struct ggml_tensor* tkn_embeddings, uint32_t max_token_idx = 0, bool return_pooled = false) {
         // input_ids: [N, n_token]
         GGML_ASSERT(input_ids->ne[0] <= position_ids->ne[0]);
 
         // token_embedding + position_embedding
         struct ggml_tensor* x;
         x = ggml_add(ctx0,
-                     ggml_get_rows(ctx0, token_embed_weight, input_ids),
+                     ggml_get_rows(ctx0, tkn_embeddings == NULL ? token_embed_weight : tkn_embeddings, input_ids),
                      ggml_get_rows(ctx0,
                                    position_embed_weight,
                                    ggml_view_1d(ctx0, position_ids, input_ids->ne[0], 0)));  // [N, n_token, hidden_size]
@@ -723,6 +775,10 @@ struct CLIPTextModel {
 
         final_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
 
+        if(version == OPENAI_CLIP_VIT_L_14) {
+            token_embed_custom = ggml_new_tensor_2d(ctx, wtype, hidden_size, max_position_embeddings);
+        }
+
         if (version == OPEN_CLIP_VIT_BIGG_14) {
             text_projection = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, projection_dim, hidden_size);
         }
@@ -805,11 +861,11 @@ struct FrozenCLIPEmbedderWithCustomWords : public GGMLModule {
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx0, struct ggml_tensor* input_ids, struct ggml_tensor* input_ids2, size_t max_token_idx = 0, bool return_pooled = false) {
+    struct ggml_tensor* forward(struct ggml_context* ctx0, struct ggml_tensor* input_ids, struct ggml_tensor* input_ids2, struct ggml_tensor* embeddings, size_t max_token_idx = 0, bool return_pooled = false) {
         if (return_pooled) {
-            return text_model2.forward(ctx0, input_ids2, max_token_idx, return_pooled);
+            return text_model2.forward(ctx0, input_ids2, NULL, max_token_idx, return_pooled);
         }
-        auto hidden_states = text_model.forward(ctx0, input_ids);  // [N, n_token, hidden_size]
+        auto hidden_states = text_model.forward(ctx0, input_ids, embeddings);  // [N, n_token, hidden_size]
         // LOG_DEBUG("hidden_states: %d %d %d %d %d", hidden_states->n_dims, hidden_states->ne[0], hidden_states->ne[1], hidden_states->ne[2], hidden_states->ne[3]);
         if (version == VERSION_XL) {
             hidden_states = ggml_reshape_4d(ctx0,
@@ -820,7 +876,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public GGMLModule {
                                             hidden_states->ne[3]);
             hidden_states = ggml_cont(ctx0, ggml_permute(ctx0, hidden_states, 2, 0, 1, 3));
 
-            auto hidden_states2 = text_model2.forward(ctx0, input_ids2);  // [N, n_token, hidden_size2]
+            auto hidden_states2 = text_model2.forward(ctx0, input_ids2, NULL);  // [N, n_token, hidden_size2]
             hidden_states2      = ggml_reshape_4d(ctx0,
                                              hidden_states2,
                                              hidden_states2->ne[0],
@@ -857,12 +913,36 @@ struct FrozenCLIPEmbedderWithCustomWords : public GGMLModule {
             LOG_DEBUG("parse '%s' to %s", text.c_str(), ss.str().c_str());
         }
 
+        auto on_new_token_cb = [&] (std::string& str, std::vector<int32_t> &bpe_tokens) -> bool {
+            size_t word_end = str.find(",");
+            std::string embd_name = word_end == std::string::npos ? str : str.substr(0, word_end);
+            embd_name = trim(embd_name);
+            std::string embd_path = get_full_path(text_model.embd_dir, embd_name + ".pt");
+            if(embd_path.size() == 0) {
+                embd_path = get_full_path(text_model.embd_dir, embd_name + ".ckpt");
+            }
+            if(embd_path.size() == 0) {
+                embd_path = get_full_path(text_model.embd_dir, embd_name + ".safetensors");
+            }
+            if(embd_path.size() > 0) {
+                if(text_model.load_embedding(embd_name, embd_path, bpe_tokens)) {
+                    if(word_end != std::string::npos) {
+                        str = str.substr(word_end);
+                    } else {
+                        str = "";
+                    }
+                    return true;
+                }
+            }
+            return false;
+        };
+
         std::vector<int> tokens;
         std::vector<float> weights;
         for (const auto& item : parsed_attention) {
             const std::string& curr_text = item.first;
             float curr_weight            = item.second;
-            std::vector<int> curr_tokens = tokenizer.encode(curr_text);
+            std::vector<int> curr_tokens = tokenizer.encode(curr_text, on_new_token_cb);
             tokens.insert(tokens.end(), curr_tokens.begin(), curr_tokens.end());
             weights.insert(weights.end(), curr_tokens.size(), curr_weight);
         }
@@ -951,7 +1031,26 @@ struct FrozenCLIPEmbedderWithCustomWords : public GGMLModule {
             }
         }
 
-        struct ggml_tensor* hidden_states = forward(ctx0, input_ids, input_ids2, max_token_idx, return_pooled);
+        struct ggml_tensor* embeddings = NULL;
+
+        if(text_model.num_custom_embeddings > 0 && version != VERSION_XL) {
+            embeddings = ggml_new_tensor_2d(ctx0, wtype, text_model.hidden_size, text_model.vocab_size + text_model.num_custom_embeddings /* custom placeholder */);
+            ggml_allocr_alloc(allocr, embeddings);
+            if (!ggml_allocr_is_measure(allocr)) {
+                // really bad, there is memory inflexibility (this is for host<->device memory conflicts)
+                void* freeze_data = malloc(ggml_nbytes(text_model.token_embed_weight));
+                ggml_backend_tensor_get_and_sync(backend, text_model.token_embed_weight, freeze_data, 0, ggml_nbytes(text_model.token_embed_weight));
+                ggml_backend_tensor_set(embeddings, freeze_data, 0, ggml_nbytes(text_model.token_embed_weight));
+                free(freeze_data);
+                // concatenate custom embeddings
+                void* custom_data = malloc(ggml_nbytes(text_model.token_embed_custom));
+                ggml_backend_tensor_get_and_sync(backend, text_model.token_embed_custom, custom_data, 0, ggml_nbytes(text_model.token_embed_custom));
+                ggml_backend_tensor_set(embeddings, custom_data, ggml_nbytes(text_model.token_embed_weight), text_model.num_custom_embeddings * text_model.hidden_size * ggml_type_size(wtype));
+                free(custom_data);
+            }
+        }
+
+        struct ggml_tensor* hidden_states = forward(ctx0, input_ids, input_ids2, embeddings, max_token_idx, return_pooled);
 
         ggml_build_forward_expand(gf, hidden_states);
         ggml_free(ctx0);