fix: make flash attn work with high noise diffusion model (#1111)

examples/cli/README.md

@@ -31,6 +31,7 @@ Context Options:
   --high-noise-diffusion-model <string>    path to the standalone high noise diffusion model
   --vae <string>                           path to standalone vae model
   --taesd <string>                         path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
+  --tae <string>                           alias of --taesd
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory

examples/cli/main.cpp

@@ -31,7 +31,6 @@ struct SDCliParams {
     std::string output_path = "output.png";
 
     bool verbose          = false;
-    bool version          = false;
     bool canny_preprocess = false;
 
     preview_t preview_method = PREVIEW_NONE;
@@ -74,10 +73,6 @@ struct SDCliParams {
              "--verbose",
              "print extra info",
              true, &verbose},
-            {"",
-             "--version",
-             "print stable-diffusion.cpp version",
-             true, &version},
             {"",
              "--color",
              "colors the logging tags according to level",
@@ -106,9 +101,8 @@ struct SDCliParams {
                     }
                 }
                 if (mode_found == -1) {
-                    fprintf(stderr,
-                            "error: invalid mode %s, must be one of [%s]\n",
-                            mode_c_str, SD_ALL_MODES_STR);
+                    LOG_ERROR("error: invalid mode %s, must be one of [%s]\n",
+                              mode_c_str, SD_ALL_MODES_STR);
                     exit(1);
                 }
                 mode = (SDMode)mode_found;
@@ -128,8 +122,7 @@ struct SDCliParams {
                 }
             }
             if (preview_found == -1) {
-                fprintf(stderr, "error: preview method %s\n",
-                        preview);
+                LOG_ERROR("error: preview method %s", preview);
                 return -1;
             }
             preview_method = (preview_t)preview_found;
@@ -161,7 +154,7 @@ struct SDCliParams {
 
     bool process_and_check() {
         if (output_path.length() == 0) {
-            fprintf(stderr, "error: the following arguments are required: output_path\n");
+            LOG_ERROR("error: the following arguments are required: output_path");
             return false;
         }
 
@@ -219,18 +212,6 @@ void parse_args(int argc, const char** argv, SDCliParams& cli_params, SDContextP
     }
 }
 
-static std::string sd_basename(const std::string& path) {
-    size_t pos = path.find_last_of('/');
-    if (pos != std::string::npos) {
-        return path.substr(pos + 1);
-    }
-    pos = path.find_last_of('\\');
-    if (pos != std::string::npos) {
-        return path.substr(pos + 1);
-    }
-    return path;
-}
-
 std::string get_image_params(const SDCliParams& cli_params, const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed) {
     std::string parameter_string = gen_params.prompt_with_lora + "\n";
     if (gen_params.negative_prompt.size() != 0) {
@@ -288,47 +269,9 @@ std::string get_image_params(const SDCliParams& cli_params, const SDContextParam
     return parameter_string;
 }
 
-/* Enables Printing the log level tag in color using ANSI escape codes */
 void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
     SDCliParams* cli_params = (SDCliParams*)data;
-    int tag_color;
-    const char* level_str;
-    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
-
-    if (!log || (!cli_params->verbose && level <= SD_LOG_DEBUG)) {
-        return;
-    }
-
-    switch (level) {
-        case SD_LOG_DEBUG:
-            tag_color = 37;
-            level_str = "DEBUG";
-            break;
-        case SD_LOG_INFO:
-            tag_color = 34;
-            level_str = "INFO";
-            break;
-        case SD_LOG_WARN:
-            tag_color = 35;
-            level_str = "WARN";
-            break;
-        case SD_LOG_ERROR:
-            tag_color = 31;
-            level_str = "ERROR";
-            break;
-        default: /* Potential future-proofing */
-            tag_color = 33;
-            level_str = "?????";
-            break;
-    }
-
-    if (cli_params->color == true) {
-        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
-    } else {
-        fprintf(out_stream, "[%-5s] ", level_str);
-    }
-    fputs(log, out_stream);
-    fflush(out_stream);
+    log_print(level, log, cli_params->verbose, cli_params->color);
 }
 
 bool load_images_from_dir(const std::string dir,
@@ -338,7 +281,7 @@ bool load_images_from_dir(const std::string dir,
                           int max_image_num   = 0,
                           bool verbose        = false) {
     if (!fs::exists(dir) || !fs::is_directory(dir)) {
-        fprintf(stderr, "'%s' is not a valid directory\n", dir.c_str());
+        LOG_ERROR("'%s' is not a valid directory\n", dir.c_str());
         return false;
     }
 
@@ -360,14 +303,12 @@ bool load_images_from_dir(const std::string dir,
         std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
 
         if (ext == ".jpg" || ext == ".jpeg" || ext == ".png" || ext == ".bmp") {
-            if (verbose) {
-                printf("load image %zu from '%s'\n", images.size(), path.c_str());
-            }
+            LOG_DEBUG("load image %zu from '%s'", images.size(), path.c_str());
             int width             = 0;
             int height            = 0;
             uint8_t* image_buffer = load_image_from_file(path.c_str(), width, height, expected_width, expected_height);
             if (image_buffer == nullptr) {
-                fprintf(stderr, "load image from '%s' failed\n", path.c_str());
+                LOG_ERROR("load image from '%s' failed", path.c_str());
                 return false;
             }
 
@@ -408,9 +349,6 @@ int main(int argc, const char* argv[]) {
     SDGenerationParams gen_params;
 
     parse_args(argc, argv, cli_params, ctx_params, gen_params);
-    if (cli_params.verbose || cli_params.version) {
-        std::cout << version_string() << "\n";
-    }
     if (gen_params.video_frames > 4) {
         size_t last_dot_pos   = cli_params.preview_path.find_last_of(".");
         std::string base_path = cli_params.preview_path;
@@ -429,6 +367,8 @@ int main(int argc, const char* argv[]) {
         cli_params.preview_fps /= 4;
 
     sd_set_log_callback(sd_log_cb, (void*)&cli_params);
+    log_verbose = cli_params.verbose;
+    log_color   = cli_params.color;
     sd_set_preview_callback(step_callback,
                             cli_params.preview_method,
                             cli_params.preview_interval,
@@ -436,12 +376,11 @@ int main(int argc, const char* argv[]) {
                             cli_params.preview_noisy,
                             (void*)&cli_params);
 
-    if (cli_params.verbose) {
-        printf("%s", sd_get_system_info());
-        printf("%s\n", cli_params.to_string().c_str());
-        printf("%s\n", ctx_params.to_string().c_str());
-        printf("%s\n", gen_params.to_string().c_str());
-    }
+    LOG_DEBUG("version: %s", version_string().c_str());
+    LOG_DEBUG("%s", sd_get_system_info());
+    LOG_DEBUG("%s", cli_params.to_string().c_str());
+    LOG_DEBUG("%s", ctx_params.to_string().c_str());
+    LOG_DEBUG("%s", gen_params.to_string().c_str());
 
     if (cli_params.mode == CONVERT) {
         bool success = convert(ctx_params.model_path.c_str(),
@@ -450,17 +389,16 @@ int main(int argc, const char* argv[]) {
                                ctx_params.wtype,
                                ctx_params.tensor_type_rules.c_str());
         if (!success) {
-            fprintf(stderr,
-                    "convert '%s'/'%s' to '%s' failed\n",
-                    ctx_params.model_path.c_str(),
-                    ctx_params.vae_path.c_str(),
-                    cli_params.output_path.c_str());
+            LOG_ERROR("convert '%s'/'%s' to '%s' failed",
+                      ctx_params.model_path.c_str(),
+                      ctx_params.vae_path.c_str(),
+                      cli_params.output_path.c_str());
             return 1;
         } else {
-            printf("convert '%s'/'%s' to '%s' success\n",
-                   ctx_params.model_path.c_str(),
-                   ctx_params.vae_path.c_str(),
-                   cli_params.output_path.c_str());
+            LOG_INFO("convert '%s'/'%s' to '%s' success",
+                     ctx_params.model_path.c_str(),
+                     ctx_params.vae_path.c_str(),
+                     cli_params.output_path.c_str());
             return 0;
         }
     }
@@ -503,7 +441,7 @@ int main(int argc, const char* argv[]) {
         int height      = 0;
         init_image.data = load_image_from_file(gen_params.init_image_path.c_str(), width, height, gen_params.width, gen_params.height);
         if (init_image.data == nullptr) {
-            fprintf(stderr, "load image from '%s' failed\n", gen_params.init_image_path.c_str());
+            LOG_ERROR("load image from '%s' failed", gen_params.init_image_path.c_str());
             release_all_resources();
             return 1;
         }
@@ -516,7 +454,7 @@ int main(int argc, const char* argv[]) {
         int height     = 0;
         end_image.data = load_image_from_file(gen_params.end_image_path.c_str(), width, height, gen_params.width, gen_params.height);
         if (end_image.data == nullptr) {
-            fprintf(stderr, "load image from '%s' failed\n", gen_params.end_image_path.c_str());
+            LOG_ERROR("load image from '%s' failed", gen_params.end_image_path.c_str());
             release_all_resources();
             return 1;
         }
@@ -528,7 +466,7 @@ int main(int argc, const char* argv[]) {
         int height      = 0;
         mask_image.data = load_image_from_file(gen_params.mask_image_path.c_str(), width, height, gen_params.width, gen_params.height, 1);
         if (mask_image.data == nullptr) {
-            fprintf(stderr, "load image from '%s' failed\n", gen_params.mask_image_path.c_str());
+            LOG_ERROR("load image from '%s' failed", gen_params.mask_image_path.c_str());
             release_all_resources();
             return 1;
         }
@@ -536,7 +474,7 @@ int main(int argc, const char* argv[]) {
         mask_image.data = (uint8_t*)malloc(gen_params.width * gen_params.height);
         memset(mask_image.data, 255, gen_params.width * gen_params.height);
         if (mask_image.data == nullptr) {
-            fprintf(stderr, "malloc mask image failed\n");
+            LOG_ERROR("malloc mask image failed");
             release_all_resources();
             return 1;
         }
@@ -547,7 +485,7 @@ int main(int argc, const char* argv[]) {
         int height         = 0;
         control_image.data = load_image_from_file(gen_params.control_image_path.c_str(), width, height, gen_params.width, gen_params.height);
         if (control_image.data == nullptr) {
-            fprintf(stderr, "load image from '%s' failed\n", gen_params.control_image_path.c_str());
+            LOG_ERROR("load image from '%s' failed", gen_params.control_image_path.c_str());
             release_all_resources();
             return 1;
         }
@@ -568,7 +506,7 @@ int main(int argc, const char* argv[]) {
             int height            = 0;
             uint8_t* image_buffer = load_image_from_file(path.c_str(), width, height);
             if (image_buffer == nullptr) {
-                fprintf(stderr, "load image from '%s' failed\n", path.c_str());
+                LOG_ERROR("load image from '%s' failed", path.c_str());
                 release_all_resources();
                 return 1;
             }
@@ -616,7 +554,7 @@ int main(int argc, const char* argv[]) {
         num_results = 1;
         results     = (sd_image_t*)calloc(num_results, sizeof(sd_image_t));
         if (results == nullptr) {
-            printf("failed to allocate results array\n");
+            LOG_INFO("failed to allocate results array");
             release_all_resources();
             return 1;
         }
@@ -627,7 +565,7 @@ int main(int argc, const char* argv[]) {
         sd_ctx_t* sd_ctx = new_sd_ctx(&sd_ctx_params);
 
         if (sd_ctx == nullptr) {
-            printf("new_sd_ctx_t failed\n");
+            LOG_INFO("new_sd_ctx_t failed");
             release_all_resources();
             return 1;
         }
@@ -704,7 +642,7 @@ int main(int argc, const char* argv[]) {
         }
 
         if (results == nullptr) {
-            printf("generate failed\n");
+            LOG_ERROR("generate failed");
             free_sd_ctx(sd_ctx);
             return 1;
         }
@@ -721,7 +659,7 @@ int main(int argc, const char* argv[]) {
                                                         gen_params.upscale_tile_size);
 
         if (upscaler_ctx == nullptr) {
-            printf("new_upscaler_ctx failed\n");
+            LOG_ERROR("new_upscaler_ctx failed");
         } else {
             for (int i = 0; i < num_results; i++) {
                 if (results[i].data == nullptr) {
@@ -731,7 +669,7 @@ int main(int argc, const char* argv[]) {
                 for (int u = 0; u < gen_params.upscale_repeats; ++u) {
                     sd_image_t upscaled_image = upscale(upscaler_ctx, current_image, upscale_factor);
                     if (upscaled_image.data == nullptr) {
-                        printf("upscale failed\n");
+                        LOG_ERROR("upscale failed");
                         break;
                     }
                     free(current_image.data);
@@ -749,8 +687,8 @@ int main(int argc, const char* argv[]) {
             std::error_code ec;
             fs::create_directories(out_dir, ec);  // OK if already exists
             if (ec) {
-                fprintf(stderr, "failed to create directory '%s': %s\n",
-                        out_dir.string().c_str(), ec.message().c_str());
+                LOG_ERROR("failed to create directory '%s': %s",
+                          out_dir.string().c_str(), ec.message().c_str());
                 return 1;
             }
         }
@@ -780,7 +718,7 @@ int main(int argc, const char* argv[]) {
             vid_output_path = base_path + ".avi";
         }
         create_mjpg_avi_from_sd_images(vid_output_path.c_str(), results, num_results, gen_params.fps);
-        printf("save result MJPG AVI video to '%s'\n", vid_output_path.c_str());
+        LOG_INFO("save result MJPG AVI video to '%s'\n", vid_output_path.c_str());
     } else {
         // appending ".png" to absent or unknown extension
         if (!is_jpg && file_ext_lower != ".png") {
@@ -796,11 +734,11 @@ int main(int argc, const char* argv[]) {
             if (is_jpg) {
                 write_ok = stbi_write_jpg(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
                                           results[i].data, 90, get_image_params(cli_params, ctx_params, gen_params, gen_params.seed + i).c_str());
-                printf("save result JPEG image to '%s' (%s)\n", final_image_path.c_str(), write_ok == 0 ? "failure" : "success");
+                LOG_INFO("save result JPEG image to '%s' (%s)", final_image_path.c_str(), write_ok == 0 ? "failure" : "success");
             } else {
                 write_ok = stbi_write_png(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
                                           results[i].data, 0, get_image_params(cli_params, ctx_params, gen_params, gen_params.seed + i).c_str());
-                printf("save result PNG image to '%s' (%s)\n", final_image_path.c_str(), write_ok == 0 ? "failure" : "success");
+                LOG_INFO("save result PNG image to '%s' (%s)", final_image_path.c_str(), write_ok == 0 ? "failure" : "success");
             }
         }
     }

examples/common/common.hpp

@@ -86,6 +86,114 @@ static std::string argv_to_utf8(int index, const char** argv) {
 
 #endif
 
+static void print_utf8(FILE* stream, const char* utf8) {
+    if (!utf8)
+        return;
+
+#ifdef _WIN32
+    HANDLE h = (stream == stderr)
+                   ? GetStdHandle(STD_ERROR_HANDLE)
+                   : GetStdHandle(STD_OUTPUT_HANDLE);
+
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, utf8, -1, NULL, 0);
+    if (wlen <= 0)
+        return;
+
+    wchar_t* wbuf = (wchar_t*)malloc(wlen * sizeof(wchar_t));
+    MultiByteToWideChar(CP_UTF8, 0, utf8, -1, wbuf, wlen);
+
+    DWORD written;
+    WriteConsoleW(h, wbuf, wlen - 1, &written, NULL);
+
+    free(wbuf);
+#else
+    fputs(utf8, stream);
+#endif
+}
+
+static std::string sd_basename(const std::string& path) {
+    size_t pos = path.find_last_of('/');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    pos = path.find_last_of('\\');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    return path;
+}
+
+static void log_print(enum sd_log_level_t level, const char* log, bool verbose, bool color) {
+    int tag_color;
+    const char* level_str;
+    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
+
+    if (!log || (!verbose && level <= SD_LOG_DEBUG)) {
+        return;
+    }
+
+    switch (level) {
+        case SD_LOG_DEBUG:
+            tag_color = 37;
+            level_str = "DEBUG";
+            break;
+        case SD_LOG_INFO:
+            tag_color = 34;
+            level_str = "INFO";
+            break;
+        case SD_LOG_WARN:
+            tag_color = 35;
+            level_str = "WARN";
+            break;
+        case SD_LOG_ERROR:
+            tag_color = 31;
+            level_str = "ERROR";
+            break;
+        default: /* Potential future-proofing */
+            tag_color = 33;
+            level_str = "?????";
+            break;
+    }
+
+    if (color) {
+        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
+    } else {
+        fprintf(out_stream, "[%-5s] ", level_str);
+    }
+    print_utf8(out_stream, log);
+    fflush(out_stream);
+}
+
+#define LOG_BUFFER_SIZE 4096
+
+static bool log_verbose = false;
+static bool log_color   = false;
+
+static void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
+    va_list args;
+    va_start(args, format);
+
+    static char log_buffer[LOG_BUFFER_SIZE + 1];
+    int written = snprintf(log_buffer, LOG_BUFFER_SIZE, "%s:%-4d - ", sd_basename(file).c_str(), line);
+
+    if (written >= 0 && written < LOG_BUFFER_SIZE) {
+        vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
+    }
+    size_t len = strlen(log_buffer);
+    if (log_buffer[len - 1] != '\n') {
+        strncat(log_buffer, "\n", LOG_BUFFER_SIZE - len);
+    }
+
+    log_print(level, log_buffer, log_verbose, log_color);
+
+    va_end(args);
+}
+
+#define LOG_DEBUG(format, ...) log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_INFO(format, ...) log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_WARN(format, ...) log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_ERROR(format, ...) log_printf(SD_LOG_ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
+
 struct StringOption {
     std::string short_name;
     std::string long_name;
@@ -295,11 +403,11 @@ static bool parse_options(int argc, const char** argv, const std::vector<ArgOpti
         }
 
         if (invalid_arg) {
-            fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+            LOG_ERROR("error: invalid parameter for argument: %s", arg.c_str());
             return false;
         }
         if (!found_arg) {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            LOG_ERROR("error: unknown argument: %s", arg.c_str());
             return false;
         }
     }
@@ -406,6 +514,10 @@ struct SDContextParams {
              "--taesd",
              "path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)",
              &taesd_path},
+            {"",
+             "--tae",
+             "alias of --taesd",
+             &taesd_path},
             {"",
              "--control-net",
              "path to control net model",
@@ -510,8 +622,8 @@ struct SDContextParams {
             const char* arg = argv[index];
             wtype           = str_to_sd_type(arg);
             if (wtype == SD_TYPE_COUNT) {
-                fprintf(stderr, "error: invalid weight format %s\n",
-                        arg);
+                LOG_ERROR("error: invalid weight format %s",
+                          arg);
                 return -1;
             }
             return 1;
@@ -524,8 +636,8 @@ struct SDContextParams {
             const char* arg = argv[index];
             rng_type        = str_to_rng_type(arg);
             if (rng_type == RNG_TYPE_COUNT) {
-                fprintf(stderr, "error: invalid rng type %s\n",
-                        arg);
+                LOG_ERROR("error: invalid rng type %s",
+                          arg);
                 return -1;
             }
             return 1;
@@ -538,8 +650,8 @@ struct SDContextParams {
             const char* arg  = argv[index];
             sampler_rng_type = str_to_rng_type(arg);
             if (sampler_rng_type == RNG_TYPE_COUNT) {
-                fprintf(stderr, "error: invalid sampler rng type %s\n",
-                        arg);
+                LOG_ERROR("error: invalid sampler rng type %s",
+                          arg);
                 return -1;
             }
             return 1;
@@ -552,8 +664,8 @@ struct SDContextParams {
             const char* arg = argv[index];
             prediction      = str_to_prediction(arg);
             if (prediction == PREDICTION_COUNT) {
-                fprintf(stderr, "error: invalid prediction type %s\n",
-                        arg);
+                LOG_ERROR("error: invalid prediction type %s",
+                          arg);
                 return -1;
             }
             return 1;
@@ -566,8 +678,8 @@ struct SDContextParams {
             const char* arg = argv[index];
             lora_apply_mode = str_to_lora_apply_mode(arg);
             if (lora_apply_mode == LORA_APPLY_MODE_COUNT) {
-                fprintf(stderr, "error: invalid lora apply model %s\n",
-                        arg);
+                LOG_ERROR("error: invalid lora apply model %s",
+                          arg);
                 return -1;
             }
             return 1;
@@ -691,13 +803,13 @@ struct SDContextParams {
 
     bool process_and_check(SDMode mode) {
         if (mode != UPSCALE && model_path.length() == 0 && diffusion_model_path.length() == 0) {
-            fprintf(stderr, "error: the following arguments are required: model_path/diffusion_model\n");
+            LOG_ERROR("error: the following arguments are required: model_path/diffusion_model\n");
             return false;
         }
 
         if (mode == UPSCALE) {
             if (esrgan_path.length() == 0) {
-                fprintf(stderr, "error: upscale mode needs an upscaler model (--upscale-model)\n");
+                LOG_ERROR("error: upscale mode needs an upscaler model (--upscale-model)\n");
                 return false;
             }
         }
@@ -862,7 +974,7 @@ static bool is_absolute_path(const std::string& p) {
 
 struct SDGenerationParams {
     std::string prompt;
-    std::string prompt_with_lora; // for metadata record only
+    std::string prompt_with_lora;  // for metadata record only
     std::string negative_prompt;
     int clip_skip   = -1;  // <= 0 represents unspecified
     int width       = 512;
@@ -1114,8 +1226,8 @@ struct SDGenerationParams {
             const char* arg             = argv[index];
             sample_params.sample_method = str_to_sample_method(arg);
             if (sample_params.sample_method == SAMPLE_METHOD_COUNT) {
-                fprintf(stderr, "error: invalid sample method %s\n",
-                        arg);
+                LOG_ERROR("error: invalid sample method %s",
+                          arg);
                 return -1;
             }
             return 1;
@@ -1128,8 +1240,8 @@ struct SDGenerationParams {
             const char* arg                        = argv[index];
             high_noise_sample_params.sample_method = str_to_sample_method(arg);
             if (high_noise_sample_params.sample_method == SAMPLE_METHOD_COUNT) {
-                fprintf(stderr, "error: invalid high noise sample method %s\n",
-                        arg);
+                LOG_ERROR("error: invalid high noise sample method %s",
+                          arg);
                 return -1;
             }
             return 1;
@@ -1142,8 +1254,8 @@ struct SDGenerationParams {
             const char* arg         = argv[index];
             sample_params.scheduler = str_to_scheduler(arg);
             if (sample_params.scheduler == SCHEDULER_COUNT) {
-                fprintf(stderr, "error: invalid scheduler %s\n",
-                        arg);
+                LOG_ERROR("error: invalid scheduler %s",
+                          arg);
                 return -1;
             }
             return 1;
@@ -1224,17 +1336,17 @@ struct SDGenerationParams {
                     try {
                         custom_sigmas.push_back(std::stof(item));
                     } catch (const std::invalid_argument& e) {
-                        fprintf(stderr, "error: invalid float value '%s' in --sigmas\n", item.c_str());
+                        LOG_ERROR("error: invalid float value '%s' in --sigmas", item.c_str());
                         return -1;
                     } catch (const std::out_of_range& e) {
-                        fprintf(stderr, "error: float value '%s' out of range in --sigmas\n", item.c_str());
+                        LOG_ERROR("error: float value '%s' out of range in --sigmas", item.c_str());
                         return -1;
                     }
                 }
             }
 
             if (custom_sigmas.empty() && !sigmas_str.empty()) {
-                fprintf(stderr, "error: could not parse any sigma values from '%s'\n", argv[index]);
+                LOG_ERROR("error: could not parse any sigma values from '%s'", argv[index]);
                 return -1;
             }
             return 1;
@@ -1330,7 +1442,7 @@ struct SDGenerationParams {
         try {
             j = json::parse(json_str);
         } catch (...) {
-            fprintf(stderr, "json parse failed %s\n", json_str.c_str());
+            LOG_ERROR("json parse failed %s", json_str.c_str());
             return false;
         }
 
@@ -1439,7 +1551,7 @@ struct SDGenerationParams {
                     }
                 }
                 if (!found) {
-                    printf("can not found lora %s\n", final_path.lexically_normal().string().c_str());
+                    LOG_WARN("can not found lora %s", final_path.lexically_normal().string().c_str());
                     tmp    = m.suffix().str();
                     prompt = std::regex_replace(prompt, re, "", std::regex_constants::format_first_only);
                     continue;
@@ -1478,17 +1590,17 @@ struct SDGenerationParams {
     bool process_and_check(SDMode mode, const std::string& lora_model_dir) {
         prompt_with_lora = prompt;
         if (width <= 0) {
-            fprintf(stderr, "error: the width must be greater than 0\n");
+            LOG_ERROR("error: the width must be greater than 0\n");
             return false;
         }
 
         if (height <= 0) {
-            fprintf(stderr, "error: the height must be greater than 0\n");
+            LOG_ERROR("error: the height must be greater than 0\n");
             return false;
         }
 
         if (sample_params.sample_steps <= 0) {
-            fprintf(stderr, "error: the sample_steps must be greater than 0\n");
+            LOG_ERROR("error: the sample_steps must be greater than 0\n");
             return false;
         }
 
@@ -1497,7 +1609,7 @@ struct SDGenerationParams {
         }
 
         if (strength < 0.f || strength > 1.f) {
-            fprintf(stderr, "error: can only work with strength in [0.0, 1.0]\n");
+            LOG_ERROR("error: can only work with strength in [0.0, 1.0]\n");
             return false;
         }
 
@@ -1519,31 +1631,31 @@ struct SDGenerationParams {
                 };
                 trim(token);
                 if (token.empty()) {
-                    fprintf(stderr, "error: invalid easycache option '%s'\n", easycache_option.c_str());
+                    LOG_ERROR("error: invalid easycache option '%s'", easycache_option.c_str());
                     return false;
                 }
                 if (idx >= 3) {
-                    fprintf(stderr, "error: easycache expects exactly 3 comma-separated values (threshold,start,end)\n");
+                    LOG_ERROR("error: easycache expects exactly 3 comma-separated values (threshold,start,end)\n");
                     return false;
                 }
                 try {
                     values[idx] = std::stof(token);
                 } catch (const std::exception&) {
-                    fprintf(stderr, "error: invalid easycache value '%s'\n", token.c_str());
+                    LOG_ERROR("error: invalid easycache value '%s'", token.c_str());
                     return false;
                 }
                 idx++;
             }
             if (idx != 3) {
-                fprintf(stderr, "error: easycache expects exactly 3 comma-separated values (threshold,start,end)\n");
+                LOG_ERROR("error: easycache expects exactly 3 comma-separated values (threshold,start,end)\n");
                 return false;
             }
             if (values[0] < 0.0f) {
-                fprintf(stderr, "error: easycache threshold must be non-negative\n");
+                LOG_ERROR("error: easycache threshold must be non-negative\n");
                 return false;
             }
             if (values[1] < 0.0f || values[1] >= 1.0f || values[2] <= 0.0f || values[2] > 1.0f || values[1] >= values[2]) {
-                fprintf(stderr, "error: easycache start/end percents must satisfy 0.0 <= start < end <= 1.0\n");
+                LOG_ERROR("error: easycache start/end percents must satisfy 0.0 <= start < end <= 1.0\n");
                 return false;
             }
             easycache_params.enabled         = true;
@@ -1583,7 +1695,7 @@ struct SDGenerationParams {
 
         if (mode == UPSCALE) {
             if (init_image_path.length() == 0) {
-                fprintf(stderr, "error: upscale mode needs an init image (--init-img)\n");
+                LOG_ERROR("error: upscale mode needs an init image (--init-img)\n");
                 return false;
             }
         }
@@ -1698,13 +1810,13 @@ uint8_t* load_image_common(bool from_memory,
         image_buffer = (uint8_t*)stbi_load(image_path_or_bytes, &width, &height, &c, expected_channel);
     }
     if (image_buffer == nullptr) {
-        fprintf(stderr, "load image from '%s' failed\n", image_path);
+        LOG_ERROR("load image from '%s' failed", image_path);
         return nullptr;
     }
     if (c < expected_channel) {
         fprintf(stderr,
                 "the number of channels for the input image must be >= %d,"
-                "but got %d channels, image_path = %s\n",
+                "but got %d channels, image_path = %s",
                 expected_channel,
                 c,
                 image_path);
@@ -1712,12 +1824,12 @@ uint8_t* load_image_common(bool from_memory,
         return nullptr;
     }
     if (width <= 0) {
-        fprintf(stderr, "error: the width of image must be greater than 0, image_path = %s\n", image_path);
+        LOG_ERROR("error: the width of image must be greater than 0, image_path = %s", image_path);
         free(image_buffer);
         return nullptr;
     }
     if (height <= 0) {
-        fprintf(stderr, "error: the height of image must be greater than 0, image_path = %s\n", image_path);
+        LOG_ERROR("error: the height of image must be greater than 0, image_path = %s", image_path);
         free(image_buffer);
         return nullptr;
     }
@@ -1739,10 +1851,10 @@ uint8_t* load_image_common(bool from_memory,
         }
 
         if (crop_x != 0 || crop_y != 0) {
-            printf("crop input image from %dx%d to %dx%d, image_path = %s\n", width, height, crop_w, crop_h, image_path);
+            LOG_INFO("crop input image from %dx%d to %dx%d, image_path = %s", width, height, crop_w, crop_h, image_path);
             uint8_t* cropped_image_buffer = (uint8_t*)malloc(crop_w * crop_h * expected_channel);
             if (cropped_image_buffer == nullptr) {
-                fprintf(stderr, "error: allocate memory for crop\n");
+                LOG_ERROR("error: allocate memory for crop\n");
                 free(image_buffer);
                 return nullptr;
             }
@@ -1758,13 +1870,13 @@ uint8_t* load_image_common(bool from_memory,
             image_buffer = cropped_image_buffer;
         }
 
-        printf("resize input image from %dx%d to %dx%d\n", width, height, expected_width, expected_height);
+        LOG_INFO("resize input image from %dx%d to %dx%d", width, height, expected_width, expected_height);
         int resized_height = expected_height;
         int resized_width  = expected_width;
 
         uint8_t* resized_image_buffer = (uint8_t*)malloc(resized_height * resized_width * expected_channel);
         if (resized_image_buffer == nullptr) {
-            fprintf(stderr, "error: allocate memory for resize input image\n");
+            LOG_ERROR("error: allocate memory for resize input image\n");
             free(image_buffer);
             return nullptr;
         }

examples/server/README.md

@@ -24,6 +24,7 @@ Context Options:
   --high-noise-diffusion-model <string>    path to the standalone high noise diffusion model
   --vae <string>                           path to standalone vae model
   --taesd <string>                         path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
+  --tae <string>                           alias of --taesd
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory

examples/server/main.cpp

@@ -151,12 +151,12 @@ struct SDSvrParams {
 
     bool process_and_check() {
         if (listen_ip.empty()) {
-            fprintf(stderr, "error: the following arguments are required: listen_ip\n");
+            LOG_ERROR("error: the following arguments are required: listen_ip");
             return false;
         }
 
         if (listen_port < 0 || listen_port > 65535) {
-            fprintf(stderr, "error: listen_port should be in the range [0, 65535]\n");
+            LOG_ERROR("error: listen_port should be in the range [0, 65535]");
             return false;
         }
         return true;
@@ -256,69 +256,36 @@ std::vector<uint8_t> write_image_to_vector(
     return buffer;
 }
 
-/* Enables Printing the log level tag in color using ANSI escape codes */
 void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
     SDSvrParams* svr_params = (SDSvrParams*)data;
-    int tag_color;
-    const char* level_str;
-    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
-
-    if (!log || (!svr_params->verbose && level <= SD_LOG_DEBUG)) {
-        return;
-    }
-
-    switch (level) {
-        case SD_LOG_DEBUG:
-            tag_color = 37;
-            level_str = "DEBUG";
-            break;
-        case SD_LOG_INFO:
-            tag_color = 34;
-            level_str = "INFO";
-            break;
-        case SD_LOG_WARN:
-            tag_color = 35;
-            level_str = "WARN";
-            break;
-        case SD_LOG_ERROR:
-            tag_color = 31;
-            level_str = "ERROR";
-            break;
-        default: /* Potential future-proofing */
-            tag_color = 33;
-            level_str = "?????";
-            break;
-    }
-
-    if (svr_params->color == true) {
-        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
-    } else {
-        fprintf(out_stream, "[%-5s] ", level_str);
-    }
-    fputs(log, out_stream);
-    fflush(out_stream);
+    log_print(level, log, svr_params->verbose, svr_params->color);
 }
 
 int main(int argc, const char** argv) {
+    if (argc > 1 && std::string(argv[1]) == "--version") {
+        std::cout << version_string() << "\n";
+        return EXIT_SUCCESS;
+    }
     SDSvrParams svr_params;
     SDContextParams ctx_params;
     SDGenerationParams default_gen_params;
     parse_args(argc, argv, svr_params, ctx_params, default_gen_params);
 
     sd_set_log_callback(sd_log_cb, (void*)&svr_params);
+    log_verbose = svr_params.verbose;
+    log_color   = svr_params.color;
 
-    if (svr_params.verbose) {
-        printf("%s", sd_get_system_info());
-        printf("%s\n", svr_params.to_string().c_str());
-        printf("%s\n", ctx_params.to_string().c_str());
-        printf("%s\n", default_gen_params.to_string().c_str());
-    }
+    LOG_DEBUG("version: %s", version_string().c_str());
+    LOG_DEBUG("%s", sd_get_system_info());
+    LOG_DEBUG("%s", svr_params.to_string().c_str());
+    LOG_DEBUG("%s", ctx_params.to_string().c_str());
+    LOG_DEBUG("%s", default_gen_params.to_string().c_str());
 
     sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(false, false, false);
     sd_ctx_t* sd_ctx              = new_sd_ctx(&sd_ctx_params);
 
     if (sd_ctx == nullptr) {
-        printf("new_sd_ctx_t failed\n");
+        LOG_ERROR("new_sd_ctx_t failed");
         return 1;
     }
 
@@ -431,9 +398,7 @@ int main(int argc, const char** argv) {
                 return;
             }
 
-            if (svr_params.verbose) {
-                printf("%s\n", gen_params.to_string().c_str());
-            }
+            LOG_DEBUG("%s\n", gen_params.to_string().c_str());
 
             sd_image_t init_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
             sd_image_t control_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
@@ -490,7 +455,7 @@ int main(int argc, const char** argv) {
                                                          results[i].channel,
                                                          output_compression);
                 if (image_bytes.empty()) {
-                    printf("write image to mem failed\n");
+                    LOG_ERROR("write image to mem failed");
                     continue;
                 }
 
@@ -611,9 +576,7 @@ int main(int argc, const char** argv) {
                 return;
             }
 
-            if (svr_params.verbose) {
-                printf("%s\n", gen_params.to_string().c_str());
-            }
+            LOG_DEBUG("%s\n", gen_params.to_string().c_str());
 
             sd_image_t init_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
             sd_image_t control_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
@@ -735,7 +698,7 @@ int main(int argc, const char** argv) {
         }
     });
 
-    printf("listening on: %s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
+    LOG_INFO("listening on: %s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
     svr.listen(svr_params.listen_ip, svr_params.listen_port);
 
     // cleanup

model.cpp

@@ -1520,6 +1520,11 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                         i64_to_i32_vec((int64_t*)read_buf, (int32_t*)target_buf, tensor_storage.nelements());
                     }
                     if (tensor_storage.type != dst_tensor->type) {
+                        if (convert_buf == nullptr) {
+                            LOG_ERROR("read tensor data failed: too less memory for conversion");
+                            failed = true;
+                            return;
+                        }
                         convert_tensor((void*)target_buf,
                                        tensor_storage.type,
                                        convert_buf,

stable-diffusion.cpp

@@ -537,6 +537,9 @@ public:
             if (sd_ctx_params->diffusion_flash_attn) {
                 LOG_INFO("Using flash attention in the diffusion model");
                 diffusion_model->set_flash_attn_enabled(true);
+                if (high_noise_diffusion_model) {
+                    high_noise_diffusion_model->set_flash_attn_enabled(true);
+                }
             }
 
             cond_stage_model->alloc_params_buffer();
@@ -562,14 +565,27 @@ public:
             }
 
             if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
-                first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
-                                                                        offload_params_to_cpu,
-                                                                        tensor_storage_map,
-                                                                        "first_stage_model",
-                                                                        vae_decode_only,
-                                                                        version);
-                first_stage_model->alloc_params_buffer();
-                first_stage_model->get_param_tensors(tensors, "first_stage_model");
+                if (!use_tiny_autoencoder) {
+                    first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
+                                                                            offload_params_to_cpu,
+                                                                            tensor_storage_map,
+                                                                            "first_stage_model",
+                                                                            vae_decode_only,
+                                                                            version);
+                    first_stage_model->alloc_params_buffer();
+                    first_stage_model->get_param_tensors(tensors, "first_stage_model");
+                } else {
+                    tae_first_stage = std::make_shared<TinyVideoAutoEncoder>(vae_backend,
+                                                                             offload_params_to_cpu,
+                                                                             tensor_storage_map,
+                                                                             "decoder",
+                                                                             vae_decode_only,
+                                                                             version);
+                    if (sd_ctx_params->vae_conv_direct) {
+                        LOG_INFO("Using Conv2d direct in the tae model");
+                        tae_first_stage->set_conv2d_direct_enabled(true);
+                    }
+                }
             } else if (version == VERSION_CHROMA_RADIANCE) {
                 first_stage_model = std::make_shared<FakeVAE>(vae_backend,
                                                               offload_params_to_cpu);
@@ -596,14 +612,13 @@ public:
                 }
                 first_stage_model->alloc_params_buffer();
                 first_stage_model->get_param_tensors(tensors, "first_stage_model");
-            }
-            if (use_tiny_autoencoder) {
-                tae_first_stage = std::make_shared<TinyAutoEncoder>(vae_backend,
-                                                                    offload_params_to_cpu,
-                                                                    tensor_storage_map,
-                                                                    "decoder.layers",
-                                                                    vae_decode_only,
-                                                                    version);
+            } else if (use_tiny_autoencoder) {
+                tae_first_stage = std::make_shared<TinyImageAutoEncoder>(vae_backend,
+                                                                         offload_params_to_cpu,
+                                                                         tensor_storage_map,
+                                                                         "decoder.layers",
+                                                                         vae_decode_only,
+                                                                         version);
                 if (sd_ctx_params->vae_conv_direct) {
                     LOG_INFO("Using Conv2d direct in the tae model");
                     tae_first_stage->set_conv2d_direct_enabled(true);
@@ -3614,7 +3629,8 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
         denoise_mask = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, init_latent->ne[0], init_latent->ne[1], init_latent->ne[2], 1);
         ggml_set_f32(denoise_mask, 1.f);
 
-        sd_ctx->sd->process_latent_out(init_latent);
+        if (!sd_ctx->sd->use_tiny_autoencoder)
+            sd_ctx->sd->process_latent_out(init_latent);
 
         ggml_ext_tensor_iter(init_image_latent, [&](ggml_tensor* t, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
             float value = ggml_ext_tensor_get_f32(t, i0, i1, i2, i3);
@@ -3624,7 +3640,8 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
             }
         });
 
-        sd_ctx->sd->process_latent_in(init_latent);
+        if (!sd_ctx->sd->use_tiny_autoencoder)
+            sd_ctx->sd->process_latent_in(init_latent);
 
         int64_t t2 = ggml_time_ms();
         LOG_INFO("encode_first_stage completed, taking %" PRId64 " ms", t2 - t1);
@@ -3709,6 +3726,9 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
         init_latent = sd_ctx->sd->generate_init_latent(work_ctx, width, height, frames, true);
     }
 
+    print_ggml_tensor(init_latent, true);
+    print_ggml_tensor(concat_latent, true);
+
     // Get learned condition
     ConditionerParams condition_params;
     condition_params.clip_skip       = sd_vid_gen_params->clip_skip;
@@ -3847,7 +3867,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
     struct ggml_tensor* vid = sd_ctx->sd->decode_first_stage(work_ctx, final_latent, true);
     int64_t t5              = ggml_time_ms();
     LOG_INFO("decode_first_stage completed, taking %.2fs", (t5 - t4) * 1.0f / 1000);
-    if (sd_ctx->sd->free_params_immediately) {
+    if (sd_ctx->sd->free_params_immediately && !sd_ctx->sd->use_tiny_autoencoder) {
         sd_ctx->sd->first_stage_model->free_params_buffer();
     }
 

tae.hpp

@@ -162,6 +162,311 @@ public:
     }
 };
 
+class TPool : public UnaryBlock {
+    int stride;
+
+public:
+    TPool(int channels, int stride)
+        : stride(stride) {
+        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * stride, channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+    }
+
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+        auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks["conv"]);
+        auto h    = x;
+        if (stride != 1) {
+            h = ggml_reshape_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2] * stride, h->ne[3] / stride);
+        }
+        h = conv->forward(ctx, h);
+        return h;
+    }
+};
+
+class TGrow : public UnaryBlock {
+    int stride;
+
+public:
+    TGrow(int channels, int stride)
+        : stride(stride) {
+        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels * stride, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+    }
+
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+        auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks["conv"]);
+        auto h    = conv->forward(ctx, x);
+        if (stride != 1) {
+            h = ggml_reshape_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2] / stride, h->ne[3] * stride);
+        }
+        return h;
+    }
+};
+
+class MemBlock : public GGMLBlock {
+    bool has_skip_conv = false;
+
+public:
+    MemBlock(int channels, int out_channels)
+        : has_skip_conv(channels != out_channels) {
+        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * 2, out_channels, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+        if (has_skip_conv) {
+            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+        }
+    }
+
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, struct ggml_tensor* past) {
+        // x: [n, channels, h, w]
+        auto conv0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.2"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
+
+        auto h = ggml_concat(ctx->ggml_ctx, x, past, 2);
+        h      = conv0->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv1->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv2->forward(ctx, h);
+
+        auto skip = x;
+        if (has_skip_conv) {
+            auto skip_conv = std::dynamic_pointer_cast<Conv2d>(blocks["skip"]);
+            skip           = skip_conv->forward(ctx, x);
+        }
+        h = ggml_add_inplace(ctx->ggml_ctx, h, skip);
+        h = ggml_relu_inplace(ctx->ggml_ctx, h);
+        return h;
+    }
+};
+
+struct ggml_tensor* patchify(struct ggml_context* ctx,
+                             struct ggml_tensor* x,
+                             int64_t patch_size,
+                             int64_t b = 1) {
+    // x: [f, b*c, h*q, w*r]
+    // return: [f, b*c*r*q, h, w]
+    if (patch_size == 1) {
+        return x;
+    }
+    int64_t r = patch_size;
+    int64_t q = patch_size;
+
+    int64_t W = x->ne[0];
+    int64_t H = x->ne[1];
+    int64_t C = x->ne[2];
+    int64_t f = x->ne[3];
+
+    int64_t w = W / r;
+    int64_t h = H / q;
+
+    x = ggml_reshape_4d(ctx, x, W, q, h, C * f);                         // [W, q, h, C*f]
+    x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));  // [W, h, q, C*f]
+    x = ggml_reshape_4d(ctx, x, r, w, h, q * C * f);                     // [r, w, h, q*C*f]
+    x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3));  // [w, h, r, q*C*f]
+    x = ggml_reshape_4d(ctx, x, w, h, r * q * C, f);                     // [f, b*c*r*q, h, w]
+
+    return x;
+}
+
+struct ggml_tensor* unpatchify(struct ggml_context* ctx,
+                               struct ggml_tensor* x,
+                               int64_t patch_size,
+                               int64_t b = 1) {
+    // x: [f, b*c*r*q, h, w]
+    // return: [f, b*c, h*q, w*r]
+    if (patch_size == 1) {
+        return x;
+    }
+    int64_t r = patch_size;
+    int64_t q = patch_size;
+    int64_t c = x->ne[2] / b / q / r;
+    int64_t f = x->ne[3];
+    int64_t h = x->ne[1];
+    int64_t w = x->ne[0];
+
+    x = ggml_reshape_4d(ctx, x, w, h, r, q * c * b * f);                 // [q*c*b*f, r, h, w]
+    x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3));  // [r, w, h, q*c*b*f]
+    x = ggml_reshape_4d(ctx, x, r * w, h, q, c * b * f);                 // [c*b*f, q, h, r*w]
+    x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 2, 1, 3));  // [r*w, q, h, c*b*f]
+    x = ggml_reshape_4d(ctx, x, r * w, q * h, c * b, f);
+
+    return x;
+}
+
+class TinyVideoEncoder : public UnaryBlock {
+    int in_channels = 3;
+    int hidden      = 64;
+    int z_channels  = 4;
+    int num_blocks  = 3;
+    int num_layers  = 3;
+    int patch_size  = 1;
+
+public:
+    TinyVideoEncoder(int z_channels = 4, int patch_size = 1)
+        : z_channels(z_channels), patch_size(patch_size) {
+        int index                       = 0;
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels * patch_size * patch_size, hidden, {3, 3}, {1, 1}, {1, 1}));
+        index++;  // nn.ReLU()
+        for (int i = 0; i < num_layers; i++) {
+            int stride                      = i == num_layers - 1 ? 1 : 2;
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TPool(hidden, stride));
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, hidden, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+            for (int j = 0; j < num_blocks; j++) {
+                blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new MemBlock(hidden, hidden));
+            }
+        }
+        blocks[std::to_string(index)] = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, z_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) override {
+        auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["0"]);
+
+        if (patch_size > 1) {
+            z = patchify(ctx->ggml_ctx, z, patch_size, 1);
+        }
+
+        auto h = first_conv->forward(ctx, z);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+
+        int index = 2;
+        for (int i = 0; i < num_layers; i++) {
+            auto pool = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(index++)]);
+            auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(index++)]);
+
+            h = pool->forward(ctx, h);
+            h = conv->forward(ctx, h);
+            for (int j = 0; j < num_blocks; j++) {
+                auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
+                auto mem   = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                mem        = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
+                h          = block->forward(ctx, h, mem);
+            }
+        }
+        auto last_conv = std::dynamic_pointer_cast<Conv2d>(blocks[std::to_string(index)]);
+        h              = last_conv->forward(ctx, h);
+        return h;
+    }
+};
+
+class TinyVideoDecoder : public UnaryBlock {
+    int z_channels               = 4;
+    int out_channels             = 3;
+    int num_blocks               = 3;
+    static const int num_layers  = 3;
+    int channels[num_layers + 1] = {256, 128, 64, 64};
+    int patch_size               = 1;
+
+public:
+    TinyVideoDecoder(int z_channels = 4, int patch_size = 1)
+        : z_channels(z_channels), patch_size(patch_size) {
+        int index                       = 1;  // Clamp()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels[0], {3, 3}, {1, 1}, {1, 1}));
+        index++;  // nn.ReLU()
+        for (int i = 0; i < num_layers; i++) {
+            int stride = i == 0 ? 1 : 2;
+            for (int j = 0; j < num_blocks; j++) {
+                blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new MemBlock(channels[i], channels[i]));
+            }
+            index++;  // nn.Upsample()
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TGrow(channels[i], stride));
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels[i], channels[i + 1], {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+        }
+        index++;  // nn.ReLU()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels[num_layers], out_channels * patch_size * patch_size, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) override {
+        auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["1"]);
+
+        // Clamp()
+        auto h = ggml_scale_inplace(ctx->ggml_ctx,
+                                    ggml_tanh_inplace(ctx->ggml_ctx,
+                                                      ggml_scale(ctx->ggml_ctx, z, 1.0f / 3.0f)),
+                                    3.0f);
+
+        h         = first_conv->forward(ctx, h);
+        h         = ggml_relu_inplace(ctx->ggml_ctx, h);
+        int index = 3;
+        for (int i = 0; i < num_layers; i++) {
+            for (int j = 0; j < num_blocks; j++) {
+                auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
+                auto mem   = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                mem        = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
+                h          = block->forward(ctx, h, mem);
+            }
+            // upsample
+            index++;
+            h          = ggml_upscale(ctx->ggml_ctx, h, 2, GGML_SCALE_MODE_NEAREST);
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(index++)]);
+            h          = block->forward(ctx, h);
+            block      = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(index++)]);
+            h          = block->forward(ctx, h);
+        }
+        h = ggml_relu_inplace(ctx->ggml_ctx, h);
+
+        auto last_conv = std::dynamic_pointer_cast<Conv2d>(blocks[std::to_string(++index)]);
+        h              = last_conv->forward(ctx, h);
+        if (patch_size > 1) {
+            h = unpatchify(ctx->ggml_ctx, h, patch_size, 1);
+        }
+        // shape(W, H, 3, 3 + T) => shape(W, H, 3, T)
+        h = ggml_view_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2], h->ne[3] - 3, h->nb[1], h->nb[2], h->nb[3], 3 * h->nb[3]);
+        return h;
+    }
+};
+
+class TAEHV : public GGMLBlock {
+protected:
+    bool decode_only;
+    SDVersion version;
+
+public:
+    TAEHV(bool decode_only = true, SDVersion version = VERSION_WAN2)
+        : decode_only(decode_only), version(version) {
+        int z_channels = 16;
+        int patch      = 1;
+        if (version == VERSION_WAN2_2_TI2V) {
+            z_channels = 48;
+            patch      = 2;
+        }
+        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoDecoder(z_channels, patch));
+        if (!decode_only) {
+            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoEncoder(z_channels, patch));
+        }
+    }
+
+    struct ggml_tensor* decode(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
+        auto decoder = std::dynamic_pointer_cast<TinyVideoDecoder>(blocks["decoder"]);
+        if (sd_version_is_wan(version)) {
+            // (W, H, C, T) -> (W, H, T, C)
+            z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 0, 1, 3, 2));
+        }
+        auto result = decoder->forward(ctx, z);
+        if (sd_version_is_wan(version)) {
+            // (W, H, C, T) -> (W, H, T, C)
+            result = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, result, 0, 1, 3, 2));
+        }
+        return result;
+    }
+
+    struct ggml_tensor* encode(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        auto encoder = std::dynamic_pointer_cast<TinyVideoEncoder>(blocks["encoder"]);
+        // (W, H, T, C) -> (W, H, C, T)
+        x                  = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        int64_t num_frames = x->ne[3];
+        if (num_frames % 4) {
+            // pad to multiple of 4 at the end
+            auto last_frame = ggml_view_4d(ctx->ggml_ctx, x, x->ne[0], x->ne[1], x->ne[2], 1, x->nb[1], x->nb[2], x->nb[3], (num_frames - 1) * x->nb[3]);
+            for (int i = 0; i < 4 - num_frames % 4; i++) {
+                x = ggml_concat(ctx->ggml_ctx, x, last_frame, 3);
+            }
+        }
+        x = encoder->forward(ctx, x);
+        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        return x;
+    }
+};
+
 class TAESD : public GGMLBlock {
 protected:
     bool decode_only;
@@ -192,18 +497,30 @@ public:
 };
 
 struct TinyAutoEncoder : public GGMLRunner {
+    TinyAutoEncoder(ggml_backend_t backend, bool offload_params_to_cpu)
+        : GGMLRunner(backend, offload_params_to_cpu) {}
+    virtual bool compute(const int n_threads,
+                         struct ggml_tensor* z,
+                         bool decode_graph,
+                         struct ggml_tensor** output,
+                         struct ggml_context* output_ctx = nullptr) = 0;
+
+    virtual bool load_from_file(const std::string& file_path, int n_threads) = 0;
+};
+
+struct TinyImageAutoEncoder : public TinyAutoEncoder {
     TAESD taesd;
     bool decode_only = false;
 
-    TinyAutoEncoder(ggml_backend_t backend,
-                    bool offload_params_to_cpu,
-                    const String2TensorStorage& tensor_storage_map,
-                    const std::string prefix,
-                    bool decoder_only = true,
-                    SDVersion version = VERSION_SD1)
+    TinyImageAutoEncoder(ggml_backend_t backend,
+                         bool offload_params_to_cpu,
+                         const String2TensorStorage& tensor_storage_map,
+                         const std::string prefix,
+                         bool decoder_only = true,
+                         SDVersion version = VERSION_SD1)
         : decode_only(decoder_only),
           taesd(decoder_only, version),
-          GGMLRunner(backend, offload_params_to_cpu) {
+          TinyAutoEncoder(backend, offload_params_to_cpu) {
         taesd.init(params_ctx, tensor_storage_map, prefix);
     }
 
@@ -260,4 +577,73 @@ struct TinyAutoEncoder : public GGMLRunner {
     }
 };
 
+struct TinyVideoAutoEncoder : public TinyAutoEncoder {
+    TAEHV taehv;
+    bool decode_only = false;
+
+    TinyVideoAutoEncoder(ggml_backend_t backend,
+                         bool offload_params_to_cpu,
+                         const String2TensorStorage& tensor_storage_map,
+                         const std::string prefix,
+                         bool decoder_only = true,
+                         SDVersion version = VERSION_WAN2)
+        : decode_only(decoder_only),
+          taehv(decoder_only, version),
+          TinyAutoEncoder(backend, offload_params_to_cpu) {
+        taehv.init(params_ctx, tensor_storage_map, prefix);
+    }
+
+    std::string get_desc() override {
+        return "taehv";
+    }
+
+    bool load_from_file(const std::string& file_path, int n_threads) {
+        LOG_INFO("loading taehv from '%s', decode_only = %s", file_path.c_str(), decode_only ? "true" : "false");
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> taehv_tensors;
+        taehv.get_param_tensors(taehv_tensors);
+        std::set<std::string> ignore_tensors;
+        if (decode_only) {
+            ignore_tensors.insert("encoder.");
+        }
+
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(file_path)) {
+            LOG_ERROR("init taehv model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool success = model_loader.load_tensors(taehv_tensors, ignore_tensors, n_threads);
+
+        if (!success) {
+            LOG_ERROR("load tae tensors from model loader failed");
+            return false;
+        }
+
+        LOG_INFO("taehv model loaded");
+        return success;
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
+        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        z                       = to_backend(z);
+        auto runner_ctx         = get_context();
+        struct ggml_tensor* out = decode_graph ? taehv.decode(&runner_ctx, z) : taehv.encode(&runner_ctx, z);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    bool compute(const int n_threads,
+                 struct ggml_tensor* z,
+                 bool decode_graph,
+                 struct ggml_tensor** output,
+                 struct ggml_context* output_ctx = nullptr) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(z, decode_graph);
+        };
+
+        return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+};
+
 #endif  // __TAE_HPP__