Merge 68877e789b 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
 

sdxl_train.py

@@ -197,10 +197,6 @@ def train(args):
         )
         return
 
-    if cache_latents:
-        assert (
-            train_dataset_group.is_latent_cacheable()
-        ), "when caching latents, either color_aug or random_crop cannot be used / latentをキャッシュするときはcolor_augとrandom_cropは使えません"
 
     if args.cache_text_encoder_outputs:
         assert (
@@ -225,7 +221,32 @@ def train(args):
         logit_scale,
         ckpt_info,
     ) = sdxl_train_util.load_target_model(args, accelerator, "sdxl", weight_dtype)
-    # logit_scale = logit_scale.to(accelerator.device, dtype=weight_dtype)
+
+    if cache_latents:
+        assert (
+            train_dataset_group.is_latent_cacheable()
+        ), "when caching latents, either color_aug or random_crop cannot be used / latentをキャッシュするときはcolor_augとrandom_cropは使えません"
+
+        # Force FP16 for caching even if training uses FP32
+        temp_vae_dtype = torch.float16 if not args.no_half_vae else torch.float32
+        vae = vae.to(accelerator.device, dtype=temp_vae_dtype)
+        
+        # Optimize VAE performance
+        vae = vae.to(memory_format=torch.channels_last)
+        # if not isinstance(vae, torch._dynamo.eval_frame.OptimizedModule):
+        #     vae = torch.compile(vae, mode="reduce-overhead")
+        
+        vae.requires_grad_(False)
+        vae.eval()
+        with torch.no_grad():
+            train_dataset_group.cache_latents(
+                vae,
+                args.vae_batch_size,
+                args.cache_latents_to_disk,
+                accelerator.is_main_process
+            )
+        vae.to("cpu")
+        clean_memory_on_device(accelerator.device)
 
     # verify load/save model formats
     if load_stable_diffusion_format: