Fix fp16 mixed precision, model is in bf16 without full_bf16

README.md

@@ -4,21 +4,28 @@ This repository contains training, generation and utility scripts for Stable Dif
 
 SD3 training is done with `sd3_train.py`. 
 
+__Jun 29, 2024__: Fixed mixed precision training with fp16 is not working. Fixed the model is in bf16 dtype even without `--full_bf16` option (this could worsen the training result).
+
+`fp16` and `bf16` are available for mixed precision training. We are not sure which is better.
+
 `optimizer_type = "adafactor"` is recommended for 24GB VRAM GPUs. `cache_text_encoder_outputs_to_disk` and `cache_latents_to_disk` are necessary currently. 
 
 `clip_l`, `clip_g` and `t5xxl` can be specified if the checkpoint does not include them.  
 
-t5xxl doesn't seem to work with `fp16`, so use`bf16` or `fp32`. 
+t5xxl doesn't seem to work with `fp16`, so 1) use`bf16` for mixed precision, or 2) use `bf16` or `float32` for `t5xxl_dtype`. 
 
 There are `t5xxl_device` and `t5xxl_dtype` options for `t5xxl` device and dtype. 
 
+`text_encoder_batch_size` is added experimentally for caching faster.
+
 ```toml
-learning_rate = 1e-5 # seems to be too high
+learning_rate = 1e-6 # seems to depend on the batch size
 optimizer_type = "adafactor"
 optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False" ]
 cache_text_encoder_outputs = true
 cache_text_encoder_outputs_to_disk = true
 vae_batch_size = 1
+text_encoder_batch_size = 4
 cache_latents = true
 cache_latents_to_disk = true
 ```

library/sd3_train_utils.py

@@ -28,14 +28,14 @@ logger = logging.getLogger(__name__)
 from .sdxl_train_util import match_mixed_precision
 
 
-def load_target_model(args, accelerator, attn_mode, weight_dtype, t5xxl_device, t5xxl_dtype) -> Tuple[
+def load_target_model(args, accelerator, attn_mode, weight_dtype, clip_dtype, t5xxl_device, t5xxl_dtype, vae_dtype) -> Tuple[
     sd3_models.MMDiT,
     Optional[sd3_models.SDClipModel],
     Optional[sd3_models.SDXLClipG],
     Optional[sd3_models.T5XXLModel],
     sd3_models.SDVAE,
 ]:
-    model_dtype = match_mixed_precision(args, weight_dtype)  # prepare fp16/bf16
+    model_dtype = match_mixed_precision(args, weight_dtype)  # prepare fp16/bf16, None or fp16/bf16
 
     for pi in range(accelerator.state.num_processes):
         if pi == accelerator.state.local_process_index:
@@ -49,13 +49,15 @@ def load_target_model(args, accelerator, attn_mode, weight_dtype, t5xxl_device,
                 args.vae,
                 attn_mode,
                 accelerator.device if args.lowram else "cpu",
-                weight_dtype,
+                model_dtype,
                 args.disable_mmap_load_safetensors,
+                clip_dtype,
                 t5xxl_device,
                 t5xxl_dtype,
+                vae_dtype,
             )
 
-            # work on low-ram device
+            # work on low-ram device: models are already loaded on accelerator.device, but we ensure they are on device
             if args.lowram:
                 if clip_l is not None:
                     clip_l.to(accelerator.device)

library/sd3_utils.py

@@ -28,11 +28,41 @@ def load_models(
     vae_path: str,
     attn_mode: str,
     device: Union[str, torch.device],
-    weight_dtype: torch.dtype,
+    default_dtype: Optional[Union[str, torch.dtype]] = None,
     disable_mmap: bool = False,
-    t5xxl_device: Optional[str] = None,
+    clip_dtype: Optional[Union[str, torch.dtype]] = None,
-    t5xxl_dtype: Optional[str] = None,
+    t5xxl_device: Optional[Union[str, torch.device]] = None,
+    t5xxl_dtype: Optional[Union[str, torch.dtype]] = None,
+    vae_dtype: Optional[Union[str, torch.dtype]] = None,
 ):
+    """
+    Load SD3 models from checkpoint files.
+
+    Args:
+        ckpt_path: Path to the SD3 checkpoint file.
+        clip_l_path: Path to the clip_l checkpoint file.
+        clip_g_path: Path to the clip_g checkpoint file.
+        t5xxl_path: Path to the t5xxl checkpoint file.
+        vae_path: Path to the VAE checkpoint file.
+        attn_mode: Attention mode for MMDiT model.
+        device: Device for MMDiT model.
+        default_dtype: Default dtype for each model. In training, it's usually None. None means using float32.
+        disable_mmap: Disable memory mapping when loading state dict.
+        clip_dtype: Dtype for Clip models, or None to use default dtype.
+        t5xxl_device: Device for T5XXL model to load T5XXL in another device (eg. gpu). Default is None to use device.
+        t5xxl_dtype: Dtype for T5XXL model, or None to use default dtype.
+        vae_dtype: Dtype for VAE model, or None to use default dtype.
+
+    Returns:
+        Tuple of MMDiT, ClipL, ClipG, T5XXL, and VAE models.
+    """
+
+    # In SD1/2 and SDXL, the model is created with empty weights and then loaded with state dict.
+    # However, in SD3, Clip and T5XXL models are created with dtype, so we need to set dtype before loading state dict.
+    # Therefore, we need clip_dtype and t5xxl_dtype.
+
+    # default_dtype is used for full_fp16/full_bf16 training.
+
     def load_state_dict(path: str, dvc: Union[str, torch.device] = device):
         if disable_mmap:
             return safetensors.torch.load(open(path, "rb").read())
@@ -43,6 +73,9 @@ def load_models(
                 return load_file(path)  # prevent device invalid Error
 
     t5xxl_device = t5xxl_device or device
+    clip_dtype = clip_dtype or default_dtype or torch.float32
+    t5xxl_dtype = t5xxl_dtype or default_dtype or torch.float32
+    vae_dtype = vae_dtype or default_dtype or torch.float32
 
     logger.info(f"Loading SD3 models from {ckpt_path}...")
     state_dict = load_state_dict(ckpt_path)
@@ -124,7 +157,7 @@ def load_models(
         mmdit = sd3_models.create_mmdit_sd3_medium_configs(attn_mode)
 
     logger.info("Loading state dict...")
-    info = sdxl_model_util._load_state_dict_on_device(mmdit, state_dict, device, weight_dtype)
+    info = sdxl_model_util._load_state_dict_on_device(mmdit, state_dict, device, default_dtype)
     logger.info(f"Loaded MMDiT: {info}")
 
     # load ClipG and ClipL
@@ -132,7 +165,7 @@ def load_models(
         clip_l = None
     else:
         logger.info("Building ClipL")
-        clip_l = sd3_models.create_clip_l(device, weight_dtype, clip_l_sd)
+        clip_l = sd3_models.create_clip_l(device, clip_dtype, clip_l_sd)
         logger.info("Loading state dict...")
         info = clip_l.load_state_dict(clip_l_sd)
         logger.info(f"Loaded ClipL: {info}")
@@ -142,7 +175,7 @@ def load_models(
         clip_g = None
     else:
         logger.info("Building ClipG")
-        clip_g = sd3_models.create_clip_g(device, weight_dtype, clip_g_sd)
+        clip_g = sd3_models.create_clip_g(device, clip_dtype, clip_g_sd)
         logger.info("Loading state dict...")
         info = clip_g.load_state_dict(clip_g_sd)
         logger.info(f"Loaded ClipG: {info}")
@@ -165,6 +198,7 @@ def load_models(
     logger.info("Loading state dict...")
     info = vae.load_state_dict(vae_sd)
     logger.info(f"Loaded VAE: {info}")
+    vae.to(device=device, dtype=vae_dtype)
 
     return mmdit, clip_l, clip_g, t5xxl, vae
 

sd3_train.py

@@ -182,6 +182,8 @@ def train(args):
             raise ValueError(f"unexpected t5xxl_dtype: {args.t5xxl_dtype}")
     t5xxl_device = accelerator.device if args.t5xxl_device is None else args.t5xxl_device
 
+    clip_dtype = weight_dtype # if not args.train_text_encoder else None
+
     # モデルを読み込む
     attn_mode = "xformers" if args.xformers else "torch"
 
@@ -189,8 +191,9 @@ def train(args):
         attn_mode == "torch"
     ), f"attn_mode {attn_mode} is not supported. Please use `--sdpa` instead of `--xformers`. / attn_mode {attn_mode} はサポートされていません。`--xformers`の代わりに`--sdpa`を使ってください。"
 
+    # models are usually loaded on CPU and moved to GPU later. This is to avoid OOM on GPU0.
     mmdit, clip_l, clip_g, t5xxl, vae = sd3_train_utils.load_target_model(
-        args, accelerator, attn_mode, weight_dtype, t5xxl_device, t5xxl_dtype
+        args, accelerator, attn_mode, None, clip_dtype, t5xxl_device, t5xxl_dtype, vae_dtype
     )
     assert clip_l is not None, "clip_l is required / clip_lは必須です"
     assert clip_g is not None, "clip_g is required / clip_gは必須です"
@@ -868,8 +871,9 @@ def setup_parser() -> argparse.ArgumentParser:
     custom_train_functions.add_custom_train_arguments(parser)
     sd3_train_utils.add_sd3_training_arguments(parser)
 
-    # TE training is disabled temporarily
+    # parser.add_argument("--train_text_encoder", action="store_true", help="train text encoder / text encoderも学習する")
 
+    # TE training is disabled temporarily
     # parser.add_argument(
     #     "--learning_rate_te1",
     #     type=float,
@@ -886,7 +890,6 @@ def setup_parser() -> argparse.ArgumentParser:
     # parser.add_argument(
     #     "--diffusers_xformers", action="store_true", help="use xformers by diffusers / Diffusersでxformersを使用する"
     # )
-    # parser.add_argument("--train_text_encoder", action="store_true", help="train text encoder / text encoderも学習する")
     # parser.add_argument(
     #     "--no_half_vae",
     #     action="store_true",