update cache_latents/text_encoder_outputs

tools/cache_text_encoder_outputs.py

@@ -9,55 +9,68 @@ from accelerate.utils import set_seed
 import torch
 from tqdm import tqdm
 
-from library import config_util
+from library import (
+    config_util,
+    flux_train_utils,
+    flux_utils,
+    sdxl_model_util,
+    strategy_base,
+    strategy_flux,
+    strategy_sd,
+    strategy_sdxl,
+)
 from library import train_util
 from library import sdxl_train_util
+from library import utils
 from library.config_util import (
     ConfigSanitizer,
     BlueprintGenerator,
 )
 from library.utils import setup_logging, add_logging_arguments
+from tools import cache_latents
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
+
 def cache_to_disk(args: argparse.Namespace) -> None:
     setup_logging(args, reset=True)
     train_util.prepare_dataset_args(args, True)
 
-    # check cache arg
-    assert (
-        args.cache_text_encoder_outputs_to_disk
-    ), "cache_text_encoder_outputs_to_disk must be True / cache_text_encoder_outputs_to_diskはTrueである必要があります"
-
-    # できるだけ準備はしておくが今のところSDXLのみしか動かない
-    assert (
-        args.sdxl
-    ), "cache_text_encoder_outputs_to_disk is only available for SDXL / cache_text_encoder_outputs_to_diskはSDXLのみ利用可能です"
+    args.cache_text_encoder_outputs = True
+    args.cache_text_encoder_outputs_to_disk = True
 
     use_dreambooth_method = args.in_json is None
 
     if args.seed is not None:
         set_seed(args.seed)  # 乱数系列を初期化する
 
-    # tokenizerを準備する：datasetを動かすために必要
-    if args.sdxl:
-        tokenizer1, tokenizer2 = sdxl_train_util.load_tokenizers(args)
-        tokenizers = [tokenizer1, tokenizer2]
-    else:
-        tokenizer = train_util.load_tokenizer(args)
-        tokenizers = [tokenizer]
+    is_sd = not args.sdxl and not args.flux
+    is_sdxl = args.sdxl
+    is_flux = args.flux
+
+    assert (
+        is_sdxl or is_flux
+    ), "Cache text encoder outputs to disk is only supported for SDXL and FLUX models / テキストエンコーダ出力のディスクキャッシュはSDXLまたはFLUXでのみ有効です"
+    assert (
+        is_sdxl or args.weighted_captions is None
+    ), "Weighted captions are only supported for SDXL models / 重み付きキャプションはSDXLモデルでのみ有効です"
+    
+    cache_latents.set_tokenize_strategy(is_sd, is_sdxl, is_flux, args)
 
     # データセットを準備する
+    use_user_config = args.dataset_config is not None
     if args.dataset_class is None:
-        blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, False, True))
-        if args.dataset_config is not None:
-            logger.info(f"Load dataset config from {args.dataset_config}")
+        blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, args.masked_loss, True))
+        if use_user_config:
+            logger.info(f"Loading dataset config from {args.dataset_config}")
             user_config = config_util.load_user_config(args.dataset_config)
-            ignored = ["train_data_dir", "in_json"]
+            ignored = ["train_data_dir", "reg_data_dir", "in_json"]
             if any(getattr(args, attr) is not None for attr in ignored):
                 logger.warning(
-                    "ignore following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format(
+                    "ignoring the following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format(
                         ", ".join(ignored)
                     )
                 )
@@ -88,15 +101,11 @@ def cache_to_disk(args: argparse.Namespace) -> None:
                     ]
                 }
 
-        blueprint = blueprint_generator.generate(user_config, args, tokenizer=tokenizers)
+        blueprint = blueprint_generator.generate(user_config, args)
         train_dataset_group = config_util.generate_dataset_group_by_blueprint(blueprint.dataset_group)
     else:
-        train_dataset_group = train_util.load_arbitrary_dataset(args, tokenizers)
-
-    current_epoch = Value("i", 0)
-    current_step = Value("i", 0)
-    ds_for_collator = train_dataset_group if args.max_data_loader_n_workers == 0 else None
-    collator = train_util.collator_class(current_epoch, current_step, ds_for_collator)
+        # use arbitrary dataset class
+        train_dataset_group = train_util.load_arbitrary_dataset(args)
 
     # acceleratorを準備する
     logger.info("prepare accelerator")
@@ -105,66 +114,68 @@ def cache_to_disk(args: argparse.Namespace) -> None:
 
     # mixed precisionに対応した型を用意しておき適宜castする
     weight_dtype, _ = train_util.prepare_dtype(args)
+    t5xxl_dtype = utils.str_to_dtype(args.t5xxl_dtype, weight_dtype)
 
     # モデルを読み込む
     logger.info("load model")
-    if args.sdxl:
-        (_, text_encoder1, text_encoder2, _, _, _, _) = sdxl_train_util.load_target_model(args, accelerator, "sdxl", weight_dtype)
+    if is_sdxl:
+        _, text_encoder1, text_encoder2, _, _, _, _ = sdxl_train_util.load_target_model(
+            args, accelerator, sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0, weight_dtype
+        )
+        text_encoder1.to(accelerator.device, weight_dtype)
+        text_encoder2.to(accelerator.device, weight_dtype)
         text_encoders = [text_encoder1, text_encoder2]
     else:
-        text_encoder1, _, _, _ = train_util.load_target_model(args, weight_dtype, accelerator)
-        text_encoders = [text_encoder1]
+        clip_l = flux_utils.load_clip_l(
+            args.clip_l, weight_dtype, accelerator.device, disable_mmap=args.disable_mmap_load_safetensors
+        )
+
+        t5xxl = flux_utils.load_t5xxl(args.t5xxl, None, accelerator.device, disable_mmap=args.disable_mmap_load_safetensors)
+
+        if t5xxl.dtype == torch.float8_e4m3fnuz or t5xxl.dtype == torch.float8_e5m2 or t5xxl.dtype == torch.float8_e5m2fnuz:
+            raise ValueError(f"Unsupported fp8 model dtype: {t5xxl.dtype}")
+        elif t5xxl.dtype == torch.float8_e4m3fn:
+            logger.info("Loaded fp8 T5XXL model")
+
+        if t5xxl_dtype != t5xxl_dtype:
+            if t5xxl.dtype == torch.float8_e4m3fn and t5xxl_dtype.itemsize() >= 2:
+                logger.warning(
+                    "The loaded model is fp8, but the specified T5XXL dtype is larger than fp8.  This may cause a performance drop."
+                    " / ロードされたモデルはfp8ですが、指定されたT5XXLのdtypeがfp8より高精度です。精度低下が発生する可能性があります。"
+                )
+            logger.info(f"Casting T5XXL model to {t5xxl_dtype}")
+            t5xxl.to(t5xxl_dtype)
+
+        text_encoders = [clip_l, t5xxl]
 
     for text_encoder in text_encoders:
-        text_encoder.to(accelerator.device, dtype=weight_dtype)
         text_encoder.requires_grad_(False)
         text_encoder.eval()
 
-    # dataloaderを準備する
-    train_dataset_group.set_caching_mode("text")
+    # build text encoder outputs caching strategy
+    if is_sdxl:
+        text_encoder_outputs_caching_strategy = strategy_sdxl.SdxlTextEncoderOutputsCachingStrategy(
+            args.cache_text_encoder_outputs_to_disk, None, args.skip_cache_check, is_weighted=args.weighted_captions
+        )
+    else:
+        text_encoder_outputs_caching_strategy = strategy_flux.FluxTextEncoderOutputsCachingStrategy(
+            args.cache_text_encoder_outputs_to_disk,
+            args.text_encoder_batch_size,
+            args.skip_cache_check,
+            is_partial=False,
+            apply_t5_attn_mask=args.apply_t5_attn_mask,
+        )
+    strategy_base.TextEncoderOutputsCachingStrategy.set_strategy(text_encoder_outputs_caching_strategy)
 
-    # DataLoaderのプロセス数：0 は persistent_workers が使えないので注意
-    n_workers = min(args.max_data_loader_n_workers, os.cpu_count())  # cpu_count or max_data_loader_n_workers
+    # build text encoding strategy
+    if is_sdxl:
+        text_encoding_strategy = strategy_sdxl.SdxlTextEncodingStrategy()
+    else:
+        text_encoding_strategy = strategy_flux.FluxTextEncodingStrategy(args.apply_t5_attn_mask)
+    strategy_base.TextEncodingStrategy.set_strategy(text_encoding_strategy)
 
-    train_dataloader = torch.utils.data.DataLoader(
-        train_dataset_group,
-        batch_size=1,
-        shuffle=True,
-        collate_fn=collator,
-        num_workers=n_workers,
-        persistent_workers=args.persistent_data_loader_workers,
-    )
-
-    # acceleratorを使ってモデルを準備する：マルチGPUで使えるようになるはず
-    train_dataloader = accelerator.prepare(train_dataloader)
-
-    # データ取得のためのループ
-    for batch in tqdm(train_dataloader):
-        absolute_paths = batch["absolute_paths"]
-        input_ids1_list = batch["input_ids1_list"]
-        input_ids2_list = batch["input_ids2_list"]
-
-        image_infos = []
-        for absolute_path, input_ids1, input_ids2 in zip(absolute_paths, input_ids1_list, input_ids2_list):
-            image_info = train_util.ImageInfo(absolute_path, 1, "dummy", False, absolute_path)
-            image_info.text_encoder_outputs_npz = os.path.splitext(absolute_path)[0] + train_util.TEXT_ENCODER_OUTPUTS_CACHE_SUFFIX
-            image_info
-
-            if args.skip_existing:
-                if os.path.exists(image_info.text_encoder_outputs_npz):
-                    logger.warning(f"Skipping {image_info.text_encoder_outputs_npz} because it already exists.")
-                    continue
-                
-            image_info.input_ids1 = input_ids1
-            image_info.input_ids2 = input_ids2
-            image_infos.append(image_info)
-
-        if len(image_infos) > 0:
-            b_input_ids1 = torch.stack([image_info.input_ids1 for image_info in image_infos])
-            b_input_ids2 = torch.stack([image_info.input_ids2 for image_info in image_infos])
-            train_util.cache_batch_text_encoder_outputs(
-                image_infos, tokenizers, text_encoders, args.max_token_length, True, b_input_ids1, b_input_ids2, weight_dtype
-            )
+    # cache text encoder outputs
+    train_dataset_group.new_cache_text_encoder_outputs(text_encoders, accelerator)
 
     accelerator.wait_for_everyone()
     accelerator.print(f"Finished caching latents for {len(train_dataset_group)} batches.")
@@ -179,11 +190,20 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_dataset_arguments(parser, True, True, True)
     config_util.add_config_arguments(parser)
     sdxl_train_util.add_sdxl_training_arguments(parser)
+    flux_train_utils.add_flux_train_arguments(parser)
     parser.add_argument("--sdxl", action="store_true", help="Use SDXL model / SDXLモデルを使用する")
+    parser.add_argument("--flux", action="store_true", help="Use FLUX model / FLUXモデルを使用する")
+    parser.add_argument(
+        "--t5xxl_dtype",
+        type=str,
+        default=None,
+        help="T5XXL model dtype, default: None (use mixed precision dtype) / T5XXLモデルのdtype, デフォルト: None (mixed precisionのdtypeを使用)",
+    )
     parser.add_argument(
         "--skip_existing",
         action="store_true",
-        help="skip images if npz already exists (both normal and flipped exists if flip_aug is enabled) / npzが既に存在する画像をスキップする（flip_aug有効時は通常、反転の両方が存在する画像をスキップ）",
+        help="[Deprecated] This option does not work. Existing .npz files are always checked. Use `--skip_cache_check` to skip the check."
+        " / [非推奨] このオプションは機能しません。既存の .npz は常に検証されます。`--skip_cache_check` で検証をスキップできます。",
     )
     return parser