Merge remote-tracking branch 'upstream/sd3' into sd3-upstream

Author: 67372a

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

@@ -736,9 +736,9 @@ def forward(
         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
@@ -861,11 +861,11 @@ def _forward(
         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()
@@ -957,6 +957,7 @@ def forward(
         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,
@@ -970,6 +971,7 @@ def forward(
                     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,
@@ -991,6 +993,7 @@ def custom_forward(*inputs):
                     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,
@@ -1004,6 +1007,7 @@ def custom_forward(*inputs):
                     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,
@@ -1015,6 +1019,7 @@ def custom_forward(*inputs):
                 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,
@@ -1335,16 +1340,19 @@ def forward_mini_train_dit(
 
         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
 

networks/resize_lora.py

@@ -212,7 +212,6 @@ def rank_resize(S, rank, dynamic_method, dynamic_param, scale=1):
 def resize_lora_model(lora_sd, new_rank, new_conv_rank, save_dtype, device, dynamic_method, dynamic_param, verbose, svd_lowrank_niter=2):
     max_old_rank = None
     new_alpha = None
-    verbose_str = "\n"
     fro_list = []
 
     if dynamic_method:
@@ -285,15 +284,13 @@ def resize_lora_model(lora_sd, new_rank, new_conv_rank, save_dtype, device, dyna
                     if not np.isnan(fro_retained):
                         fro_list.append(float(fro_retained))
 
-                    verbose_str += f"{block_down_name:75} | "
+                    verbose_str = f"{block_down_name:75} | "
                     verbose_str += (
                         f"sum(S) retained: {sum_retained:.1%}, fro retained: {fro_retained:.1%}, max(S) ratio: {max_ratio:0.1f}"
                     )
-
-                if verbose and dynamic_method:
-                    verbose_str += f", dynamic | dim: {param_dict['new_rank']}, alpha: {param_dict['new_alpha']}\n"
-                else:
-                    verbose_str += "\n"
+                    if dynamic_method:
+                        verbose_str += f", dynamic | dim: {param_dict['new_rank']}, alpha: {param_dict['new_alpha']}"
+                    tqdm.write(verbose_str)
 
                 new_alpha = param_dict["new_alpha"]
                 o_lora_sd[block_down_name + lora_down_name + weight_name] = param_dict["lora_down"].to(save_dtype).contiguous()
@@ -308,7 +305,6 @@ def resize_lora_model(lora_sd, new_rank, new_conv_rank, save_dtype, device, dyna
                 del param_dict
 
     if verbose:
-        print(verbose_str)
         print(f"Average Frobenius norm retention: {np.mean(fro_list):.2%} | std: {np.std(fro_list):0.3f}")
     logger.info("resizing complete")
     return o_lora_sd, max_old_rank, new_alpha