fix: resolve embedding loading issue when calling generate_image multiple times (#1078)

conditioner.hpp

@@ -60,7 +60,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
     int32_t num_custom_embeddings   = 0;
     int32_t num_custom_embeddings_2 = 0;
     std::vector<uint8_t> token_embed_custom;
-    std::vector<std::string> readed_embeddings;
+    std::map<std::string, std::pair<int, int>> embedding_pos_map;
 
     FrozenCLIPEmbedderWithCustomWords(ggml_backend_t backend,
                                       bool offload_params_to_cpu,
@@ -123,14 +123,17 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
     }
 
     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_and_convert_name(embd_path)) {
             LOG_ERROR("embedding '%s' failed", embd_name.c_str());
             return false;
         }
-        if (std::find(readed_embeddings.begin(), readed_embeddings.end(), embd_name) != readed_embeddings.end()) {
+        auto iter = embedding_pos_map.find(embd_name);
+        if (iter != embedding_pos_map.end()) {
             LOG_DEBUG("embedding already read in: %s", embd_name.c_str());
+            for (int i = iter->second.first; i < iter->second.second; i++) {
+                bpe_tokens.push_back(text_model->model.vocab_size + i);
+            }
             return true;
         }
         struct ggml_init_params params;
@@ -161,7 +164,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             return true;
         };
         model_loader.load_tensors(on_load, 1);
-        readed_embeddings.push_back(embd_name);
+        int pos_start = num_custom_embeddings;
         if (embd) {
             int64_t hidden_size = text_model->model.hidden_size;
             token_embed_custom.resize(token_embed_custom.size() + ggml_nbytes(embd));
@@ -188,6 +191,11 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             }
             LOG_DEBUG("embedding '%s' applied, custom embeddings: %i (text model 2)", embd_name.c_str(), num_custom_embeddings_2);
         }
+        int pos_end = num_custom_embeddings;
+        if (pos_end == pos_start) {
+            return false;
+        }
+        embedding_pos_map[embd_name] = std::pair{pos_start, pos_end};
         return true;
     }
 

ggml_extend.hpp

@@ -60,6 +60,14 @@
 #define SD_UNUSED(x) (void)(x)
 #endif
 
+__STATIC_INLINE__ int align_up_offset(int n, int multiple) {
+    return (multiple - n % multiple) % multiple;
+}
+
+__STATIC_INLINE__ int align_up(int n, int multiple) {
+    return n + align_up_offset(n, multiple);
+}
+
 __STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void*) {
     switch (level) {
         case GGML_LOG_LEVEL_DEBUG:
@@ -1392,10 +1400,14 @@ __STATIC_INLINE__ void ggml_ext_backend_tensor_get_and_sync(ggml_backend_t backe
 }
 
 __STATIC_INLINE__ float ggml_ext_backend_tensor_get_f32(ggml_tensor* tensor) {
-    GGML_ASSERT(tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_I32);
+    GGML_ASSERT(tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_I32 || tensor->type == GGML_TYPE_BF16);
     float value;
     if (tensor->type == GGML_TYPE_F32) {
         ggml_backend_tensor_get(tensor, &value, 0, sizeof(value));
+    } else if (tensor->type == GGML_TYPE_BF16) {
+        ggml_bf16_t bf16_value;
+        ggml_backend_tensor_get(tensor, &bf16_value, 0, sizeof(bf16_value));
+        value = ggml_bf16_to_fp32(bf16_value);
     } else if (tensor->type == GGML_TYPE_F16) {
         ggml_fp16_t f16_value;
         ggml_backend_tensor_get(tensor, &f16_value, 0, sizeof(f16_value));

stable-diffusion.cpp

@@ -1898,6 +1898,18 @@ public:
         return vae_scale_factor;
     }
 
+    int get_diffusion_model_down_factor() {
+        int down_factor = 8;  // unet
+        if (sd_version_is_dit(version)) {
+            if (sd_version_is_wan(version)) {
+                down_factor = 2;
+            } else {
+                down_factor = 1;
+            }
+        }
+        return down_factor;
+    }
+
     int get_latent_channel() {
         int latent_channel = 4;
         if (sd_version_is_dit(version)) {
@@ -3133,22 +3145,19 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
     sd_ctx->sd->vae_tiling_params = sd_img_gen_params->vae_tiling_params;
     int width                     = sd_img_gen_params->width;
     int height                    = sd_img_gen_params->height;
+
     int vae_scale_factor            = sd_ctx->sd->get_vae_scale_factor();
-    if (sd_version_is_dit(sd_ctx->sd->version)) {
-        if (width % 16 || height % 16) {
-            LOG_ERROR("Image dimensions must be must be a multiple of 16 on each axis for %s models. (Got %dx%d)",
-                      model_version_to_str[sd_ctx->sd->version],
-                      width,
-                      height);
-            return nullptr;
-        }
-    } else if (width % 64 || height % 64) {
-        LOG_ERROR("Image dimensions must be must be a multiple of 64 on each axis for %s models. (Got %dx%d)",
-                  model_version_to_str[sd_ctx->sd->version],
-                  width,
-                  height);
-        return nullptr;
+    int diffusion_model_down_factor = sd_ctx->sd->get_diffusion_model_down_factor();
+    int spatial_multiple            = vae_scale_factor * diffusion_model_down_factor;
+
+    int width_offset  = align_up_offset(width, spatial_multiple);
+    int height_offset = align_up_offset(height, spatial_multiple);
+    if (width_offset > 0 || height_offset > 0) {
+        width += width_offset;
+        height += height_offset;
+        LOG_WARN("align up %dx%d to %dx%d (multiple=%d)", sd_img_gen_params->width, sd_img_gen_params->height, width, height, spatial_multiple);
     }
+
     LOG_DEBUG("generate_image %dx%d", width, height);
     if (sd_ctx == nullptr || sd_img_gen_params == nullptr) {
         return nullptr;
@@ -3422,9 +3431,19 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
     int frames       = sd_vid_gen_params->video_frames;
     frames           = (frames - 1) / 4 * 4 + 1;
     int sample_steps = sd_vid_gen_params->sample_params.sample_steps;
-    LOG_INFO("generate_video %dx%dx%d", width, height, frames);
 
     int vae_scale_factor            = sd_ctx->sd->get_vae_scale_factor();
+    int diffusion_model_down_factor = sd_ctx->sd->get_diffusion_model_down_factor();
+    int spatial_multiple            = vae_scale_factor * diffusion_model_down_factor;
+
+    int width_offset  = align_up_offset(width, spatial_multiple);
+    int height_offset = align_up_offset(height, spatial_multiple);
+    if (width_offset > 0 || height_offset > 0) {
+        width += width_offset;
+        height += height_offset;
+        LOG_WARN("align up %dx%d to %dx%d (multiple=%d)", sd_vid_gen_params->width, sd_vid_gen_params->height, width, height, spatial_multiple);
+    }
+    LOG_INFO("generate_video %dx%dx%d", width, height, frames);
 
     enum sample_method_t sample_method = sd_vid_gen_params->sample_params.sample_method;
     if (sample_method == SAMPLE_METHOD_COUNT) {