import glob import gradio as gr import os import toml from musubi_tuner.gui.config_manager import ConfigManager from musubi_tuner.gui.i18n_data import I18N_DATA config_manager = ConfigManager() i18n = gr.I18n(en=I18N_DATA["en"], ja=I18N_DATA["ja"]) def construct_ui(): # I18N doesn't work for gr.Blocks title # with gr.Blocks(title=i18n("app_title")) as demo: with gr.Blocks(title="Musubi Tuner GUI") as demo: gr.Markdown(i18n("app_header")) gr.Markdown(i18n("app_desc")) with gr.Accordion(i18n("acc_project"), open=True): gr.Markdown(i18n("desc_project")) with gr.Row(): project_dir = gr.Textbox(label=i18n("lbl_proj_dir"), placeholder=i18n("ph_proj_dir"), max_lines=1) # Placeholder for project initialization or loading init_btn = gr.Button(i18n("btn_init_project")) project_status = gr.Markdown("") with gr.Accordion(i18n("acc_model"), open=False): gr.Markdown(i18n("desc_model")) with gr.Row(): model_arch = gr.Dropdown( label=i18n("lbl_model_arch"), choices=[ "Qwen-Image", "Z-Image-Turbo", ], value="Qwen-Image", ) vram_size = gr.Dropdown(label=i18n("lbl_vram"), choices=["12", "16", "24", "32", ">32"], value="24") with gr.Row(): comfy_models_dir = gr.Textbox(label=i18n("lbl_comfy_dir"), placeholder=i18n("ph_comfy_dir"), max_lines=1) # Validation for ComfyUI models directory models_status = gr.Markdown("") validate_models_btn = gr.Button(i18n("btn_validate_models")) # Placeholder for Dataset Settings (Step 3) gr.Markdown(i18n("header_dataset")) gr.Markdown(i18n("desc_dataset")) with gr.Row(): set_rec_settings_btn = gr.Button(i18n("btn_rec_res_batch")) with gr.Row(): resolution_w = gr.Number(label=i18n("lbl_res_w"), value=1024, precision=0) resolution_h = gr.Number(label=i18n("lbl_res_h"), value=1024, precision=0) batch_size = gr.Number(label=i18n("lbl_batch_size"), value=1, precision=0) gen_toml_btn = gr.Button(i18n("btn_gen_config")) dataset_status = gr.Markdown("") toml_preview = gr.Code(label=i18n("lbl_toml_preview"), interactive=False) def load_project_settings(project_path): settings = {} try: settings_path = os.path.join(project_path, "musubi_project.toml") if os.path.exists(settings_path): with open(settings_path, "r", encoding="utf-8") as f: settings = toml.load(f) except Exception as e: print(f"Error loading project settings: {e}") return settings def load_dataset_config_content(project_path): content = "" try: config_path = os.path.join(project_path, "dataset_config.toml") if os.path.exists(config_path): with open(config_path, "r", encoding="utf-8") as f: content = f.read() except Exception as e: print(f"Error reading dataset config: {e}") return content def save_project_settings(project_path, **kwargs): try: # Load existing settings to support partial updates settings = load_project_settings(project_path) # Update with new values settings.update(kwargs) settings_path = os.path.join(project_path, "musubi_project.toml") with open(settings_path, "w", encoding="utf-8") as f: toml.dump(settings, f) except Exception as e: print(f"Error saving project settings: {e}") def init_project(path): if not path: return ( "Please enter a project directory path.", gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), ) try: os.makedirs(os.path.join(path, "training"), exist_ok=True) # Load settings if available settings = load_project_settings(path) new_model = settings.get("model_arch", "Qwen-Image") new_vram = settings.get("vram_size", "16") new_comfy = settings.get("comfy_models_dir", "") new_w = settings.get("resolution_w", 1328) new_h = settings.get("resolution_h", 1328) new_batch = settings.get("batch_size", 1) new_vae = settings.get("vae_path", "") new_te1 = settings.get("text_encoder1_path", "") new_te2 = settings.get("text_encoder2_path", "") # Training params new_dit = settings.get("dit_path", "") new_out_nm = settings.get("output_name", "my_lora") new_dim = settings.get("network_dim", 4) new_lr = settings.get("learning_rate", 1e-4) new_epochs = settings.get("num_epochs", 16) new_save_n = settings.get("save_every_n_epochs", 1) new_flow = settings.get("discrete_flow_shift", 2.0) new_swap = settings.get("block_swap", 0) new_use_pinned_memory_for_block_swap = settings.get("use_pinned_memory_for_block_swap", False) new_prec = settings.get("mixed_precision", "bf16") new_grad_cp = settings.get("gradient_checkpointing", True) new_fp8_s = settings.get("fp8_scaled", True) new_fp8_l = settings.get("fp8_llm", True) new_add_args = settings.get("additional_args", "") # Sample image params new_sample_enable = settings.get("sample_images", False) new_sample_every_n = settings.get("sample_every_n_epochs", 1) new_sample_prompt = settings.get("sample_prompt", "") new_sample_negative = settings.get("sample_negative_prompt", "") new_sample_w = settings.get("sample_w", new_w) new_sample_h = settings.get("sample_h", new_h) # Post-processing params new_in_lora = settings.get("input_lora_path", "") new_out_comfy = settings.get("output_comfy_lora_path", "") # Load dataset config content preview_content = load_dataset_config_content(path) msg = f"Project initialized at {path}. " if settings: msg += " Settings loaded." msg += " 'training' folder ready. Configure the dataset in the 'training' folder. Images and caption files (same name as image, extension is '.txt') should be placed in the 'training' folder." msg += "\n\nプロジェクトが初期化されました。" if settings: msg += "設定が読み込まれました。" msg += "'training' フォルダが準備されました。画像とキャプションファイル（画像と同じファイル名で拡張子は '.txt'）を配置してください。" return ( msg, new_model, new_vram, new_comfy, new_w, new_h, new_batch, preview_content, new_vae, new_te1, new_te2, new_dit, new_out_nm, new_dim, new_lr, new_epochs, new_save_n, new_flow, new_swap, new_use_pinned_memory_for_block_swap, new_prec, new_grad_cp, new_fp8_s, new_fp8_l, new_add_args, new_sample_enable, new_sample_every_n, new_sample_prompt, new_sample_negative, new_sample_w, new_sample_h, new_in_lora, new_out_comfy, ) except Exception as e: return ( f"Error initializing project: {str(e)}", gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), gr.update(), ) def generate_config(project_path, w, h, batch, model_val, vram_val, comfy_val, vae_val, te1_val, te2_val): if not project_path: return "Error: Project directory not specified.\nエラー: プロジェクトディレクトリが指定されていません。", "" # Save project settings first save_project_settings( project_path, model_arch=model_val, vram_size=vram_val, comfy_models_dir=comfy_val, resolution_w=w, resolution_h=h, batch_size=batch, vae_path=vae_val, text_encoder1_path=te1_val, text_encoder2_path=te2_val, ) # Normalize paths project_path = os.path.abspath(project_path) image_dir = os.path.join(project_path, "training").replace("\\", "/") cache_dir = os.path.join(project_path, "cache").replace("\\", "/") toml_content = f"""# Auto-generated by Musubi Tuner GUI [general] resolution = [{int(w)}, {int(h)}] caption_extension = ".txt" batch_size = {int(batch)} enable_bucket = true bucket_no_upscale = false [[datasets]] image_directory = "{image_dir}" cache_directory = "{cache_dir}" num_repeats = 1 """ try: config_path = os.path.join(project_path, "dataset_config.toml") with open(config_path, "w", encoding="utf-8") as f: f.write(toml_content) return f"Successfully generated config at / 設定ファイルが作成されました: {config_path}", toml_content except Exception as e: return f"Error generating config / 設定ファイルの生成に失敗しました: {str(e)}", "" with gr.Accordion(i18n("acc_preprocessing"), open=False): gr.Markdown(i18n("desc_preprocessing")) with gr.Row(): set_preprocessing_defaults_btn = gr.Button(i18n("btn_set_paths")) with gr.Row(): vae_path = gr.Textbox(label=i18n("lbl_vae_path"), placeholder=i18n("ph_vae_path"), max_lines=1) text_encoder1_path = gr.Textbox(label=i18n("lbl_te1_path"), placeholder=i18n("ph_te1_path"), max_lines=1) text_encoder2_path = gr.Textbox(label=i18n("lbl_te2_path"), placeholder=i18n("ph_te2_path"), max_lines=1) with gr.Row(): cache_latents_btn = gr.Button(i18n("btn_cache_latents")) cache_text_btn = gr.Button(i18n("btn_cache_text")) # Simple output area for caching logs caching_output = gr.Textbox(label=i18n("lbl_cache_log"), lines=10, interactive=False) def validate_models_dir(path): if not path: return "Please enter a ComfyUI models directory. / ComfyUIのmodelsディレクトリを入力してください。" required_subdirs = ["diffusion_models", "vae", "text_encoders"] missing = [] for d in required_subdirs: if not os.path.exists(os.path.join(path, d)): missing.append(d) if missing: return f"Error: Missing subdirectories in models folder / modelsフォルダに以下のサブディレクトリが見つかりません: {', '.join(missing)}" return "Valid ComfyUI models directory structure found / 有効なComfyUI modelsディレクトリ構造が見つかりました。" def set_recommended_settings(project_path, model_arch, vram_val): w, h = config_manager.get_resolution(model_arch) recommended_batch_size = config_manager.get_batch_size(model_arch, vram_val) if project_path: save_project_settings(project_path, resolution_w=w, resolution_h=h, batch_size=recommended_batch_size) return w, h, recommended_batch_size def set_preprocessing_defaults(project_path, comfy_models_dir, model_arch): if not comfy_models_dir: return gr.update(), gr.update(), gr.update() vae_default, te1_default, te2_default = config_manager.get_preprocessing_paths(model_arch, comfy_models_dir) if not te2_default: te2_default = "" # Ensure empty string for text input if project_path: save_project_settings( project_path, vae_path=vae_default, text_encoder1_path=te1_default, text_encoder2_path=te2_default ) return vae_default, te1_default, te2_default def set_training_defaults(project_path, comfy_models_dir, model_arch, vram_val): # Get number of images from project_path to adjust num_epochs later cache_dir = os.path.join(project_path, "cache") pattern = "*" + ("_qi" if model_arch == "Qwen-Image" else "_zi") + ".safetensors" num_images = len(glob.glob(os.path.join(cache_dir, pattern))) if os.path.exists(cache_dir) else 0 # Get training defaults from config manager defaults = config_manager.get_training_defaults(model_arch, vram_val, comfy_models_dir) # Adjust num_epochs based on number of images (simple heuristic) default_num_steps = defaults.get("default_num_steps", 1000) if num_images > 0: adjusted_epochs = max(1, int((default_num_steps / num_images))) else: adjusted_epochs = 16 # Fallback default sample_every_n_epochs = (adjusted_epochs // 4) if adjusted_epochs >= 4 else 1 dit_default = defaults.get("dit_path", "") dim = defaults.get("network_dim", 4) lr = defaults.get("learning_rate", 1e-4) epochs = adjusted_epochs save_n = defaults.get("save_every_n_epochs", 1) flow = defaults.get("discrete_flow_shift", 2.0) swap = defaults.get("block_swap", 0) use_pinned_memory_for_block_swap = defaults.get("use_pinned_memory_for_block_swap", False) prec = defaults.get("mixed_precision", "bf16") grad_cp = defaults.get("gradient_checkpointing", True) fp8_s = defaults.get("fp8_scaled", True) fp8_l = defaults.get("fp8_llm", True) sample_w_default, sample_h_default = config_manager.get_resolution(model_arch) if project_path: save_project_settings( project_path, dit_path=dit_default, network_dim=dim, learning_rate=lr, num_epochs=epochs, save_every_n_epochs=save_n, discrete_flow_shift=flow, block_swap=swap, use_pinned_memory_for_block_swap=use_pinned_memory_for_block_swap, mixed_precision=prec, gradient_checkpointing=grad_cp, fp8_scaled=fp8_s, fp8_llm=fp8_l, vram_size=vram_val, # Ensure VRAM size is saved sample_every_n_epochs=sample_every_n_epochs, sample_w=sample_w_default, sample_h=sample_h_default, ) return ( dit_default, dim, lr, epochs, save_n, flow, swap, use_pinned_memory_for_block_swap, prec, grad_cp, fp8_s, fp8_l, sample_every_n_epochs, sample_w_default, sample_h_default, ) def set_post_processing_defaults(project_path, output_nm): if not project_path or not output_nm: return gr.update(), gr.update() models_dir = os.path.join(project_path, "models") in_lora = os.path.join(models_dir, f"{output_nm}.safetensors") out_lora = os.path.join(models_dir, f"{output_nm}_comfy.safetensors") save_project_settings(project_path, input_lora_path=in_lora, output_comfy_lora_path=out_lora) return in_lora, out_lora import subprocess import sys def run_command(command): try: process = subprocess.Popen( command, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, shell=True, text=True, encoding="utf-8", creationflags=subprocess.CREATE_NO_WINDOW if os.name == "nt" else 0, ) output_log = command + "\n\n" for line in process.stdout: output_log += line yield output_log process.wait() if process.returncode != 0: output_log += ( f"\nError: Process exited with code / プロセスが次のコードでエラー終了しました: {process.returncode}" ) yield output_log else: output_log += "\nProcess completed successfully / プロセスが正常に完了しました" yield output_log except Exception as e: yield f"Error executing command / コマンドの実行中にエラーが発生しました: {str(e)}" def cache_latents(project_path, vae_path_val, te1, te2, model, comfy, w, h, batch, vram_val): if not project_path: yield "Error: Project directory not set. / プロジェクトディレクトリが設定されていません。" return # Save settings first save_project_settings( project_path, model_arch=model, comfy_models_dir=comfy, resolution_w=w, resolution_h=h, batch_size=batch, vae_path=vae_path_val, text_encoder1_path=te1, text_encoder2_path=te2, ) if not vae_path_val: yield "Error: VAE path not set. / VAEのパスが設定されていません。" return if not os.path.exists(vae_path_val): yield f"Error: VAE model not found at / 指定されたパスにVAEモデルが見つかりません: {vae_path_val}" return config_path = os.path.join(project_path, "dataset_config.toml") if not os.path.exists(config_path): yield f"Error: dataset_config.toml not found in {project_path}. Please generate it first. / dataset_config.tomlが {project_path} に見つかりません。先に設定ファイルを生成してください。" return script_name = "zimage_cache_latents.py" if model == "Qwen-Image": script_name = "qwen_image_cache_latents.py" script_path = os.path.join("src", "musubi_tuner", script_name) cmd = [sys.executable, script_path, "--dataset_config", config_path, "--vae", vae_path_val] # Placeholder for argument modification if model == "Z-Image-Turbo": pass elif model == "Qwen-Image": pass command_str = " ".join(cmd) yield f"Starting Latent Caching. Please wait for the first log to appear. / Latentのキャッシュを開始します。最初のログが表示されるまでにしばらくかかります。\nCommand: {command_str}\n\n" yield from run_command(command_str) def cache_text_encoder(project_path, te1_path_val, te2_path_val, vae, model, comfy, w, h, batch, vram_val): if not project_path: yield "Error: Project directory not set. / プロジェクトディレクトリが設定されていません。" return # Save settings first save_project_settings( project_path, model_arch=model, comfy_models_dir=comfy, resolution_w=w, resolution_h=h, batch_size=batch, vae_path=vae, text_encoder1_path=te1_path_val, text_encoder2_path=te2_path_val, ) if not te1_path_val: yield "Error: Text Encoder 1 path not set. / Text Encoder 1のパスが設定されていません。" return if not os.path.exists(te1_path_val): yield f"Error: Text Encoder 1 model not found at / 指定されたパスにText Encoder 1モデルが見つかりません: {te1_path_val}" return # Z-Image only uses te1 for now, but keeping te2 in signature if needed later or for other models config_path = os.path.join(project_path, "dataset_config.toml") if not os.path.exists(config_path): yield f"Error: dataset_config.toml not found in {project_path}. Please generate it first. / dataset_config.tomlが {project_path} に見つかりません。先に設定ファイルを生成してください。" return script_name = "zimage_cache_text_encoder_outputs.py" if model == "Qwen-Image": script_name = "qwen_image_cache_text_encoder_outputs.py" script_path = os.path.join("src", "musubi_tuner", script_name) cmd = [ sys.executable, script_path, "--dataset_config", config_path, "--text_encoder", te1_path_val, "--batch_size", "1", # Conservative default ] # Model-specific argument modification if model == "Z-Image-Turbo": pass elif model == "Qwen-Image": # Add --fp8_vl for low VRAM (16GB or less) if vram_val in ["12", "16"]: cmd.append("--fp8_vl") command_str = " ".join(cmd) yield f"Starting Text Encoder Caching. Please wait for the first log to appear. / Text Encoderのキャッシュを開始します。最初のログが表示されるまでにしばらくかかります。\nCommand: {command_str}\n\n" yield from run_command(command_str) with gr.Accordion(i18n("acc_training"), open=False): gr.Markdown(i18n("desc_training_basic")) training_model_info = gr.Markdown(i18n("desc_training_zimage")) with gr.Row(): set_training_defaults_btn = gr.Button(i18n("btn_rec_params")) with gr.Row(): dit_path = gr.Textbox(label=i18n("lbl_dit_path"), placeholder=i18n("ph_dit_path"), max_lines=1) with gr.Row(): output_name = gr.Textbox(label=i18n("lbl_output_name"), value="my_lora", max_lines=1) with gr.Group(): gr.Markdown(i18n("header_basic_params")) with gr.Row(): network_dim = gr.Number(label=i18n("lbl_dim"), value=4) learning_rate = gr.Number(label=i18n("lbl_lr"), value=1e-4) num_epochs = gr.Number(label=i18n("lbl_epochs"), value=16) save_every_n_epochs = gr.Number(label=i18n("lbl_save_every"), value=1) with gr.Group(): with gr.Row(): discrete_flow_shift = gr.Number(label=i18n("lbl_flow_shift"), value=2.0) block_swap = gr.Slider(label=i18n("lbl_block_swap"), minimum=0, maximum=60, step=1, value=0) use_pinned_memory_for_block_swap = gr.Checkbox( label=i18n("lbl_use_pinned_memory_for_block_swap"), value=False, ) with gr.Accordion(i18n("accordion_advanced"), open=False): gr.Markdown(i18n("desc_training_detailed")) with gr.Row(): mixed_precision = gr.Dropdown(label=i18n("lbl_mixed_precision"), choices=["bf16", "fp16", "no"], value="bf16") gradient_checkpointing = gr.Checkbox(label=i18n("lbl_grad_cp"), value=True) with gr.Row(): fp8_scaled = gr.Checkbox(label=i18n("lbl_fp8_scaled"), value=True) fp8_llm = gr.Checkbox(label=i18n("lbl_fp8_llm"), value=True) with gr.Group(): gr.Markdown(i18n("header_sample_images")) sample_images = gr.Checkbox(label=i18n("lbl_enable_sample"), value=False) with gr.Row(): sample_prompt = gr.Textbox(label=i18n("lbl_sample_prompt"), placeholder=i18n("ph_sample_prompt")) with gr.Row(): sample_negative_prompt = gr.Textbox( label=i18n("lbl_sample_negative_prompt"), placeholder=i18n("ph_sample_negative_prompt"), ) with gr.Row(): sample_w = gr.Number(label=i18n("lbl_sample_w"), value=1024, precision=0) sample_h = gr.Number(label=i18n("lbl_sample_h"), value=1024, precision=0) sample_every_n = gr.Number(label=i18n("lbl_sample_every_n"), value=1, precision=0) with gr.Accordion(i18n("accordion_additional"), open=False): gr.Markdown(i18n("desc_additional_args")) additional_args = gr.Textbox(label=i18n("lbl_additional_args"), placeholder=i18n("ph_additional_args")) training_status = gr.Markdown("") start_training_btn = gr.Button(i18n("btn_start_training"), variant="primary") with gr.Accordion(i18n("acc_post_processing"), open=False): gr.Markdown(i18n("desc_post_proc")) with gr.Row(): set_post_proc_defaults_btn = gr.Button(i18n("btn_set_paths")) with gr.Row(): input_lora = gr.Textbox(label=i18n("lbl_input_lora"), placeholder=i18n("ph_input_lora"), max_lines=1) output_comfy_lora = gr.Textbox(label=i18n("lbl_output_comfy"), placeholder=i18n("ph_output_comfy"), max_lines=1) convert_btn = gr.Button(i18n("btn_convert")) conversion_log = gr.Textbox(label=i18n("lbl_conversion_log"), lines=5, interactive=False) def convert_lora_to_comfy(project_path, input_path, output_path, model, comfy, w, h, batch, vae, te1, te2): if not project_path: yield "Error: Project directory not set. / プロジェクトディレクトリが設定されていません。" return # Save settings save_project_settings( project_path, model_arch=model, comfy_models_dir=comfy, resolution_w=w, resolution_h=h, batch_size=batch, vae_path=vae, text_encoder1_path=te1, text_encoder2_path=te2, input_lora_path=input_path, output_comfy_lora_path=output_path, ) if not input_path or not output_path: yield "Error: Input and Output paths must be specified. / 入力・出力パスを指定してください。" return if not os.path.exists(input_path): yield f"Error: Input file not found at {input_path} / 入力ファイルが見つかりません: {input_path}" return # Script path script_path = os.path.join("src", "musubi_tuner", "networks", "convert_z_image_lora_to_comfy.py") if not os.path.exists(script_path): yield f"Error: Conversion script not found at {script_path} / 変換スクリプトが見つかりません: {script_path}" return cmd = [sys.executable, script_path, input_path, output_path] command_str = " ".join(cmd) yield f"Starting Conversion. / 変換を開始します。\nCommand: {command_str}\n\n" yield from run_command(command_str) def start_training( project_path, model, dit, vae, te1, output_nm, dim, lr, epochs, save_n, flow_shift, swap, use_pinned_memory_for_block_swap, prec, grad_cp, fp8_s, fp8_l, add_args, should_sample_images, sample_every_n, sample_prompt_val, sample_negative_prompt_val, sample_w_val, sample_h_val, ): import shlex if not project_path: return "Error: Project directory not set. / プロジェクトディレクトリが設定されていません。" if not dit: return "Error: Base Model / DiT Path not set. / Base Model / DiTのパスが設定されていません。" if not os.path.exists(dit): return f"Error: Base Model / DiT file not found at {dit} / Base Model / DiTファイルが見つかりません: {dit}" if not vae: return "Error: VAE Path not set (configure in Preprocessing). / VAEのパスが設定されていません (Preprocessingで設定してください)。" if not te1: return "Error: Text Encoder 1 Path not set (configure in Preprocessing). / Text Encoder 1のパスが設定されていません (Preprocessingで設定してください)。" dataset_config = os.path.join(project_path, "dataset_config.toml") if not os.path.exists(dataset_config): return "Error: dataset_config.toml not found. Please generate it. / dataset_config.toml が見つかりません。生成してください。" output_dir = os.path.join(project_path, "models") logging_dir = os.path.join(project_path, "logs") # Save settings save_project_settings( project_path, dit_path=dit, output_name=output_nm, network_dim=dim, learning_rate=lr, num_epochs=epochs, save_every_n_epochs=save_n, discrete_flow_shift=flow_shift, block_swap=swap, use_pinned_memory_for_block_swap=use_pinned_memory_for_block_swap, mixed_precision=prec, gradient_checkpointing=grad_cp, fp8_scaled=fp8_s, fp8_llm=fp8_l, vae_path=vae, text_encoder1_path=te1, additional_args=add_args, sample_images=should_sample_images, sample_every_n_epochs=sample_every_n, sample_prompt=sample_prompt_val, sample_negative_prompt=sample_negative_prompt_val, sample_w=sample_w_val, sample_h=sample_h_val, ) # Model specific command modification if model == "Z-Image-Turbo": arch_name = "zimage" elif model == "Qwen-Image": arch_name = "qwen_image" # Construct command for cmd /c to run and then pause # We assume 'accelerate' is in the PATH. script_path = os.path.join("src", "musubi_tuner", f"{arch_name}_train_network.py") # Inner command list - arguments for accelerate launch inner_cmd = [ "accelerate", "launch", # accelerate args: we don't configure default_config.yaml, so we need to specify all here "--num_cpu_threads_per_process", "1", "--mixed_precision", prec, "--dynamo_backend=no", "--gpu_ids", "all", "--machine_rank", "0", "--main_training_function", "main", "--num_machines", "1", "--num_processes", "1", # script and its args script_path, "--dit", dit, "--vae", vae, "--text_encoder", te1, "--dataset_config", dataset_config, "--output_dir", output_dir, "--output_name", output_nm, "--network_module", f"networks.lora_{arch_name}", "--network_dim", str(int(dim)), "--optimizer_type", "adamw8bit", "--learning_rate", str(lr), "--max_train_epochs", str(int(epochs)), "--save_every_n_epochs", str(int(save_n)), "--timestep_sampling", "shift", "--weighting_scheme", "none", "--discrete_flow_shift", str(flow_shift), "--max_data_loader_n_workers", "2", "--persistent_data_loader_workers", "--seed", "42", "--logging_dir", logging_dir, "--log_with", "tensorboard", ] # Sample image generation options if should_sample_images: sample_prompt_path = os.path.join(project_path, "sample_prompt.txt") templates = { # prompt, negative prompt, width, height, flow shift, steps, CFG scale, seed "Qwen-Image": "{prompt} --n {neg} --w {w} --h {h} --fs 2.2 --s 20 --l 4.0 --d 1234", "Z-Image-Turbo": "{prompt} --n {neg} --w {w} --h {h} --fs 3.0 --s 20 --l 5.0 --d 1234", } template = templates.get(model, templates["Z-Image-Turbo"]) prompt_str = (sample_prompt_val or "").replace("\n", " ").strip() neg_str = (sample_negative_prompt_val or "").replace("\n", " ").strip() try: w_int = int(sample_w_val) h_int = int(sample_h_val) except Exception: return "Error: Sample width/height must be integers. / サンプル画像の幅と高さは整数で指定してください。" line = template.format(prompt=prompt_str, neg=neg_str, w=w_int, h=h_int) try: with open(sample_prompt_path, "w", encoding="utf-8") as f: f.write(line + "\n") except Exception as e: return f"Error writing sample_prompt.txt / sample_prompt.txt の作成に失敗しました: {str(e)}" inner_cmd.extend( [ "--sample_prompts", sample_prompt_path, "--sample_at_first", "--sample_every_n_epochs", str(int(sample_every_n)), ] ) if prec != "no": inner_cmd.extend(["--mixed_precision", prec]) if grad_cp: inner_cmd.append("--gradient_checkpointing") if fp8_s: inner_cmd.append("--fp8_base") inner_cmd.append("--fp8_scaled") if fp8_l: if model == "Z-Image-Turbo": inner_cmd.append("--fp8_llm") elif model == "Qwen-Image": inner_cmd.append("--fp8_vl") if swap > 0: inner_cmd.extend(["--blocks_to_swap", str(int(swap))]) if use_pinned_memory_for_block_swap: inner_cmd.append("--use_pinned_memory_for_block_swap") inner_cmd.append("--sdpa") inner_cmd.append("--split_attn") # Model specific command modification if model == "Z-Image-Turbo": pass elif model == "Qwen-Image": pass # Parse and append additional args if add_args: try: split_args = shlex.split(add_args) inner_cmd.extend(split_args) except Exception as e: return f"Error parsing additional arguments / 追加引数の解析に失敗しました: {str(e)}" # Construct the full command string for cmd /c # list2cmdline will quote arguments as needed for Windows inner_cmd_str = subprocess.list2cmdline(inner_cmd) # Chain commands: Run training -> echo message -> pause >nul (hides default message) final_cmd_str = f"{inner_cmd_str} & echo. & echo Training finished. Press any key to close this window... 学習が完了しました。このウィンドウを閉じるには任意のキーを押してください。 & pause >nul" try: # Open in new console window flags = subprocess.CREATE_NEW_CONSOLE if os.name == "nt" else 0 # Pass explicit 'cmd', '/c', string to ensure proper execution subprocess.Popen(["cmd", "/c", final_cmd_str], creationflags=flags, shell=False) return f"Training started in a new window! / 新しいウィンドウで学習が開始されました！\nCommand: {inner_cmd_str}" except Exception as e: return f"Error starting training / 学習の開始に失敗しました: {str(e)}" def update_model_info(model): if model == "Z-Image-Turbo": return i18n("desc_training_zimage") elif model == "Qwen-Image": return i18n("desc_qwen_notes") return "" # Event wiring moved to end to prevent UnboundLocalError init_btn.click( fn=init_project, inputs=[project_dir], outputs=[ project_status, model_arch, vram_size, comfy_models_dir, resolution_w, resolution_h, batch_size, toml_preview, vae_path, text_encoder1_path, text_encoder2_path, dit_path, output_name, network_dim, learning_rate, num_epochs, save_every_n_epochs, discrete_flow_shift, block_swap, use_pinned_memory_for_block_swap, mixed_precision, gradient_checkpointing, fp8_scaled, fp8_llm, additional_args, sample_images, sample_every_n, sample_prompt, sample_negative_prompt, sample_w, sample_h, input_lora, output_comfy_lora, ], ) model_arch.change(fn=update_model_info, inputs=[model_arch], outputs=[training_model_info]) gen_toml_btn.click( fn=generate_config, inputs=[ project_dir, resolution_w, resolution_h, batch_size, model_arch, vram_size, comfy_models_dir, vae_path, text_encoder1_path, text_encoder2_path, ], outputs=[dataset_status, toml_preview], ) validate_models_btn.click(fn=validate_models_dir, inputs=[comfy_models_dir], outputs=[models_status]) set_rec_settings_btn.click( fn=set_recommended_settings, inputs=[project_dir, model_arch, vram_size], outputs=[resolution_w, resolution_h, batch_size], ) set_preprocessing_defaults_btn.click( fn=set_preprocessing_defaults, inputs=[project_dir, comfy_models_dir, model_arch], outputs=[vae_path, text_encoder1_path, text_encoder2_path], ) set_post_proc_defaults_btn.click( fn=set_post_processing_defaults, inputs=[project_dir, output_name], outputs=[input_lora, output_comfy_lora] ) set_training_defaults_btn.click( fn=set_training_defaults, inputs=[project_dir, comfy_models_dir, model_arch, vram_size], outputs=[ dit_path, network_dim, learning_rate, num_epochs, save_every_n_epochs, discrete_flow_shift, block_swap, use_pinned_memory_for_block_swap, mixed_precision, gradient_checkpointing, fp8_scaled, fp8_llm, sample_every_n, sample_w, sample_h, ], ) cache_latents_btn.click( fn=cache_latents, inputs=[ project_dir, vae_path, text_encoder1_path, text_encoder2_path, model_arch, comfy_models_dir, resolution_w, resolution_h, batch_size, vram_size, ], outputs=[caching_output], ) cache_text_btn.click( fn=cache_text_encoder, inputs=[ project_dir, text_encoder1_path, text_encoder2_path, vae_path, model_arch, comfy_models_dir, resolution_w, resolution_h, batch_size, vram_size, ], outputs=[caching_output], ) start_training_btn.click( fn=start_training, inputs=[ project_dir, model_arch, dit_path, vae_path, text_encoder1_path, output_name, network_dim, learning_rate, num_epochs, save_every_n_epochs, discrete_flow_shift, block_swap, use_pinned_memory_for_block_swap, mixed_precision, gradient_checkpointing, fp8_scaled, fp8_llm, additional_args, sample_images, sample_every_n, sample_prompt, sample_negative_prompt, sample_w, sample_h, ], outputs=[training_status], ) convert_btn.click( fn=convert_lora_to_comfy, inputs=[ project_dir, input_lora, output_comfy_lora, model_arch, comfy_models_dir, resolution_w, resolution_h, batch_size, vae_path, text_encoder1_path, text_encoder2_path, ], outputs=[conversion_log], ) return demo if __name__ == "__main__": demo = construct_ui() demo.launch(i18n=i18n)