Merge fad68161c9 into 51435f1718

fix: improve numerical stability by conditionally using float32 in Anima with fp16 training

README-ja.md

@@ -50,6 +50,9 @@ Stable Diffusion等の画像生成モデルの学習、モデルによる画像
 
 ### 更新履歴
 
+- **Version 0.10.3 (2026-04-02):**
+    - Animaでfp16で学習する際の安定性をさらに改善しました。[PR #2302](https://github.com/kohya-ss/sd-scripts/pull/2302) 問題をご報告いただいた方々に深く感謝します。
+
 - **Version 0.10.2 (2026-03-30):**
     - SD/SDXLのLECO学習に対応しました。[PR #2285](https://github.com/kohya-ss/sd-scripts/pull/2285) および [PR #2294](https://github.com/kohya-ss/sd-scripts/pull/2294) umisetokikaze氏に深く感謝します。
         - 詳細は[ドキュメント](./docs/train_leco.md)をご覧ください。

README.md

@@ -47,6 +47,9 @@ If you find this project helpful, please consider supporting its development via
 
 ### Change History
 
+- **Version 0.10.3 (2026-04-02):**
+    - Stability when training with fp16 on Anima has been further improved. See [PR #2302](https://github.com/kohya-ss/sd-scripts/pull/2302) for details. We deeply appreciate those who reported the issue.
+
 - **Version 0.10.2 (2026-03-30):**
     - LECO training for SD/SDXL is now supported. Many thanks to umisetokikaze for [PR #2285](https://github.com/kohya-ss/sd-scripts/pull/2285) and [PR #2294](https://github.com/kohya-ss/sd-scripts/pull/2294).
         - Please refer to the [documentation](./docs/train_leco.md) for details.

library/anima_models.py

@@ -738,9 +738,9 @@ class FinalLayer(nn.Module):
         x_B_T_H_W_D: torch.Tensor,
         emb_B_T_D: torch.Tensor,
         adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
+        use_fp32: bool = False,
     ):
         # Compute AdaLN modulation parameters (in float32 when fp16 to avoid overflow in Linear layers)
-        use_fp32 = x_B_T_H_W_D.dtype == torch.float16
         with torch.autocast(device_type=x_B_T_H_W_D.device.type, dtype=torch.float32, enabled=use_fp32):
             if self.use_adaln_lora:
                 assert adaln_lora_B_T_3D is not None
@@ -863,11 +863,11 @@ class Block(nn.Module):
         emb_B_T_D: torch.Tensor,
         crossattn_emb: torch.Tensor,
         attn_params: attention.AttentionParams,
+        use_fp32: bool = False,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
         adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
         extra_per_block_pos_emb: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
-        use_fp32 = x_B_T_H_W_D.dtype == torch.float16
         if use_fp32:
             # Cast to float32 for better numerical stability in residual connections. Each module will cast back to float16 by enclosing autocast context.
             x_B_T_H_W_D = x_B_T_H_W_D.float()
@@ -959,6 +959,7 @@ class Block(nn.Module):
         emb_B_T_D: torch.Tensor,
         crossattn_emb: torch.Tensor,
         attn_params: attention.AttentionParams,
+        use_fp32: bool = False,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
         adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
         extra_per_block_pos_emb: Optional[torch.Tensor] = None,
@@ -972,6 +973,7 @@ class Block(nn.Module):
                     emb_B_T_D,
                     crossattn_emb,
                     attn_params,
+                    use_fp32,
                     rope_emb_L_1_1_D,
                     adaln_lora_B_T_3D,
                     extra_per_block_pos_emb,
@@ -994,6 +996,7 @@ class Block(nn.Module):
                     emb_B_T_D,
                     crossattn_emb,
                     attn_params,
+                    use_fp32,
                     rope_emb_L_1_1_D,
                     adaln_lora_B_T_3D,
                     extra_per_block_pos_emb,
@@ -1007,6 +1010,7 @@ class Block(nn.Module):
                     emb_B_T_D,
                     crossattn_emb,
                     attn_params,
+                    use_fp32,
                     rope_emb_L_1_1_D,
                     adaln_lora_B_T_3D,
                     extra_per_block_pos_emb,
@@ -1018,6 +1022,7 @@ class Block(nn.Module):
                 emb_B_T_D,
                 crossattn_emb,
                 attn_params,
+                use_fp32,
                 rope_emb_L_1_1_D,
                 adaln_lora_B_T_3D,
                 extra_per_block_pos_emb,
@@ -1338,16 +1343,19 @@ class Anima(nn.Module):
 
         attn_params = attention.AttentionParams.create_attention_params(self.attn_mode, self.split_attn)
 
+        # Determine whether to use float32 for block computations based on input dtype (use float32 for better stability when input is float16)
+        use_fp32 = x_B_T_H_W_D.dtype == torch.float16
+
         for block_idx, block in enumerate(self.blocks):
             if self.blocks_to_swap:
                 self.offloader.wait_for_block(block_idx)
 
-            x_B_T_H_W_D = block(x_B_T_H_W_D, t_embedding_B_T_D, crossattn_emb, attn_params, **block_kwargs)
+            x_B_T_H_W_D = block(x_B_T_H_W_D, t_embedding_B_T_D, crossattn_emb, attn_params, use_fp32, **block_kwargs)
 
             if self.blocks_to_swap:
                 self.offloader.submit_move_blocks(self.blocks, block_idx)
 
-        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
+        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D, use_fp32=use_fp32)
         x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)
         return x_B_C_Tt_Hp_Wp
 

library/model_util.py

@@ -11,7 +11,7 @@ init_ipex()
 
 import diffusers
 from transformers import CLIPTextModel, CLIPTokenizer, CLIPTextConfig, logging
-from diffusers import AutoencoderKL, DDIMScheduler, StableDiffusionPipeline  # , UNet2DConditionModel
+from diffusers import AutoencoderKL, DDIMScheduler, StableDiffusionPipeline, StableUnCLIPImg2ImgPipeline  # , UNet2DConditionModel
 from safetensors.torch import load_file, save_file
 from library.original_unet import UNet2DConditionModel
 from library.utils import setup_logging
@@ -658,6 +658,77 @@ def convert_ldm_clip_checkpoint_v2(checkpoint, max_length):
     return new_sd
 
 
+def convert_ldm_clip_checkpoint_v2_fix(checkpoint, max_length):
+    # 嫌になるくらい違うぞ！
+    def convert_key(key):
+        if not key.startswith("cond_stage_model"):
+            return None
+
+        # common conversion
+        key = key.replace("cond_stage_model.model.transformer.", "text_model.encoder.")
+        key = key.replace("cond_stage_model.model.", "text_model.")
+
+        if "resblocks" in key:
+            # resblocks conversion
+            key = key.replace(".resblocks.", ".layers.")
+            if ".ln_" in key:
+                key = key.replace(".ln_", ".layer_norm")
+            elif ".mlp." in key:
+                key = key.replace(".c_fc.", ".fc1.")
+                key = key.replace(".c_proj.", ".fc2.")
+            elif ".attn.out_proj" in key:
+                key = key.replace(".attn.out_proj.", ".self_attn.out_proj.")
+            elif ".attn.in_proj" in key:
+                key = None  # 特殊なので後で処理する
+            else:
+                raise ValueError(f"unexpected key in SD: {key}")
+        elif ".positional_embedding" in key:
+            key = key.replace(".positional_embedding", ".embeddings.position_embedding.weight")
+        elif ".text_projection" in key:
+            key = None  # 使われない???
+        elif ".logit_scale" in key:
+            key = None  # 使われない???
+        elif ".token_embedding" in key:
+            key = key.replace(".token_embedding.weight", ".embeddings.token_embedding.weight")
+        elif ".ln_final" in key:
+            key = key.replace(".ln_final", ".final_layer_norm")
+        return key
+
+    keys = list(checkpoint.keys())
+    new_sd = {}
+    for key in keys:
+        # remove resblocks 23
+        if ".resblocks.23." in key:
+            continue
+        if 'embedder.model' in key:
+            continue
+        new_key = convert_key(key)
+        if new_key is None:
+            continue
+        new_sd[new_key] = checkpoint[key]
+
+    # attnの変換
+    for key in keys:
+        if ".resblocks.23." in key:
+            continue
+        if 'embedder.model' in key:
+            continue
+        if ".resblocks" in key and ".attn.in_proj_" in key:
+            # 三つに分割
+            values = torch.chunk(checkpoint[key], 3)
+
+            key_suffix = ".weight" if "weight" in key else ".bias"
+            key_pfx = key.replace("cond_stage_model.model.transformer.resblocks.", "text_model.encoder.layers.")
+            key_pfx = key_pfx.replace("_weight", "")
+            key_pfx = key_pfx.replace("_bias", "")
+            key_pfx = key_pfx.replace(".attn.in_proj", ".self_attn.")
+            new_sd[key_pfx + "q_proj" + key_suffix] = values[0]
+            new_sd[key_pfx + "k_proj" + key_suffix] = values[1]
+            new_sd[key_pfx + "v_proj" + key_suffix] = values[2]
+
+    return new_sd
+
+
 # endregion
 
 
@@ -1017,33 +1088,58 @@ def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dt
     vae = AutoencoderKL(**vae_config).to(device)
     info = vae.load_state_dict(converted_vae_checkpoint)
     logger.info(f"loading vae: {info}")
-
-    # convert text_model
+    
     if v2:
-        converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v2(state_dict, 77)
-        cfg = CLIPTextConfig(
-            vocab_size=49408,
-            hidden_size=1024,
-            intermediate_size=4096,
-            num_hidden_layers=23,
-            num_attention_heads=16,
-            max_position_embeddings=77,
-            hidden_act="gelu",
-            layer_norm_eps=1e-05,
-            dropout=0.0,
-            attention_dropout=0.0,
-            initializer_range=0.02,
-            initializer_factor=1.0,
-            pad_token_id=1,
-            bos_token_id=0,
-            eos_token_id=2,
-            model_type="clip_text_model",
-            projection_dim=512,
-            torch_dtype="float32",
-            transformers_version="4.25.0.dev0",
-        )
-        text_model = CLIPTextModel._from_config(cfg)
-        info = text_model.load_state_dict(converted_text_encoder_checkpoint)
+        try:
+            converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v2_fix(state_dict, 77)
+            cfg = CLIPTextConfig(
+                attention_dropout = 0.0,
+                bos_token_id = 0,
+                dropout = 0.0,
+                eos_token_id = 2,
+                hidden_act = "gelu",
+                hidden_size = 1024,
+                initializer_factor = 1.0,
+                initializer_range = 0.02,
+                intermediate_size = 4096,
+                layer_norm_eps = 1e-05,
+                max_position_embeddings = 77,
+                model_type = "clip_text_model",
+                num_attention_heads = 16,
+                num_hidden_layers = 23,
+                pad_token_id = 1,
+                projection_dim = 512,
+                torch_dtype = "float16",
+                transformers_version = "4.28.0.dev0",
+                vocab_size = 49408
+            )
+            text_model = CLIPTextModel._from_config(cfg)
+            info = text_model.load_state_dict(converted_text_encoder_checkpoint)
+        except Exception as e:
+            converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v2(state_dict, 77)
+            cfg = CLIPTextConfig(
+                vocab_size=49408,
+                hidden_size=1024,
+                intermediate_size=4096,
+                num_hidden_layers=23,
+                num_attention_heads=16,
+                max_position_embeddings=77,
+                hidden_act="gelu",
+                layer_norm_eps=1e-05,
+                dropout=0.0,
+                attention_dropout=0.0,
+                initializer_range=0.02,
+                initializer_factor=1.0,
+                pad_token_id=1,
+                bos_token_id=0,
+                eos_token_id=2,
+                model_type="clip_text_model",
+                projection_dim=512,
+                torch_dtype="float32",
+                transformers_version="4.25.0.dev0",
+            )
+            text_model = CLIPTextModel._from_config(cfg)
+            info = text_model.load_state_dict(converted_text_encoder_checkpoint)     
     else:
         converted_text_encoder_checkpoint = convert_ldm_clip_checkpoint_v1(state_dict)
 
@@ -1077,6 +1173,25 @@ def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dt
 
     return text_model, vae, unet
 
+# def load_models_from_stable_diffusion_checkpoint(v2, ckpt_path, device="cpu", dtype=torch.float32):
+#     pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(ckpt_path, torch_dtype=torch.float32).to(device)
+    
+#     # Load the UNet model
+#     unet = pipe.unet.to(device)
+
+#     # Load the VAE model
+#     vae = pipe.vae.to(device)
+
+#     # Load the text model
+#     text_encoder = pipe.text_encoder.to(device)
+
+#     # Log information
+#     logger.info(f"Loaded UNet: {unet}")
+#     logger.info(f"Loaded VAE: {vae}")
+#     logger.info(f"Loaded Text Encoder: {text_encoder}")
+
+#     return text_encoder, vae, unet
+
 
 def get_model_version_str_for_sd1_sd2(v2, v_parameterization):
     # only for reference