Merge pull request #2312 from kohya-ss/dev

Merge dev to main

README-ja.md

@@ -50,6 +50,12 @@ 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
 
+- 次のリリースに含まれる予定の主な変更点は以下の通りです。リリース前の変更点は予告なく変更される可能性があります。
+    - Intel GPUの互換性を向上しました。[PR #2307](https://github.com/kohya-ss/sd-scripts/pull/2307) WhitePr氏に感謝します。
+
+- **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 +47,12 @@ If you find this project helpful, please consider supporting its development via
 
 ### Change History
 
+- The following are the main changes planned for the next release. Please note that these changes may be subject to change without notice before the release.
+    - Improved compatibility with Intel GPUs. Thanks to WhitePr for [PR #2307](https://github.com/kohya-ss/sd-scripts/pull/2307).
+
+- **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.

gen_img.py

@@ -542,20 +542,10 @@ class PipelineLike:
                 uncond_embeddings = torch.cat([uncond_embeddings, tes_uncond_embs[i]], dim=2)  # n,77,2048
 
         if do_classifier_free_guidance:
-            lcm = uncond_embeddings.shape[1] * text_embeddings.shape[1] // math.gcd(uncond_embeddings.shape[1], text_embeddings.shape[1])
-            
             if negative_scale is None:
-                text_embeddings = torch.cat([
-                    uncond_embeddings.repeat(1, lcm // uncond_embeddings.shape[1], 1), 
-                    text_embeddings.repeat(1, lcm // text_embeddings.shape[1], 1),
-                ])
+                text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
             else:
-                lcm = real_uncond_embeddings.shape[1] * text_embeddings.shape[1] // math.gcd(real_uncond_embeddings.shape[1], text_embeddings.shape[1])
-                text_embeddings = torch.cat([
-                    uncond_embeddings.repeat(1, lcm // uncond_embeddings.shape[1], 1), 
-                    text_embeddings.repeat(1, lcm // text_embeddings.shape[1], 1), 
-                    real_uncond_embeddings.repeat(1, lcm // real_uncond_embeddings.shape[1], 1)
-                ])
+                text_embeddings = torch.cat([uncond_embeddings, text_embeddings, real_uncond_embeddings])
 
         if self.control_net_lllites or (self.control_nets and self.is_sdxl):
             # ControlNetのhintにguide imageを流用する。ControlNetの場合はControlNet側で行う
@@ -1115,17 +1105,22 @@ def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, pad, no_boseos
     """
     max_embeddings_multiples = (max_length - 2) // (chunk_length - 2)
     weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length
-    lcm = chunk_length
     for i in range(len(tokens)):
-        target_length = ((len(tokens[i]) + 2) // chunk_length + 1) * chunk_length
-        lcm = target_length * lcm // math.gcd(target_length, lcm)
-        tokens[i] = [bos] + tokens[i] + [eos] + [pad] * (target_length - 2 - len(tokens[i]))
-        weights[i] = [1.0] + weights[i] + [1.0] * (target_length - 1 - len(weights[i]))
-            
-    for i in range(len(tokens)):
-        tokens[i] = tokens[i] * (lcm // len(tokens[i]))
-        weights[i] = weights[i] * (lcm // len(weights[i]))
-    
+        tokens[i] = [bos] + tokens[i] + [eos] + [pad] * (max_length - 2 - len(tokens[i]))
+        if no_boseos_middle:
+            weights[i] = [1.0] + weights[i] + [1.0] * (max_length - 1 - len(weights[i]))
+        else:
+            w = []
+            if len(weights[i]) == 0:
+                w = [1.0] * weights_length
+            else:
+                for j in range(max_embeddings_multiples):
+                    w.append(1.0)  # weight for starting token in this chunk
+                    w += weights[i][j * (chunk_length - 2) : min(len(weights[i]), (j + 1) * (chunk_length - 2))]
+                    w.append(1.0)  # weight for ending token in this chunk
+                w += [1.0] * (weights_length - len(w))
+            weights[i] = w[:]
+
     return tokens, weights
 
 
@@ -1143,21 +1138,56 @@ def get_unweighted_text_embeddings(
     When the length of tokens is a multiple of the capacity of the text encoder,
     it should be split into chunks and sent to the text encoder individually.
     """
-    pool = None
-    text_embeddings = []
-    
-    for chunk in text_input.chunk(text_input.shape[1] // chunk_length, dim=1):
-        enc_out = text_encoder(chunk, output_hidden_states=True, return_dict=True)
-        text_embedding = enc_out["hidden_states"][-clip_skip]
+    max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2)
+    if max_embeddings_multiples > 1:
+        text_embeddings = []
+        pool = None
+        for i in range(max_embeddings_multiples):
+            # extract the i-th chunk
+            text_input_chunk = text_input[:, i * (chunk_length - 2) : (i + 1) * (chunk_length - 2) + 2].clone()
+
+            # cover the head and the tail by the starting and the ending tokens
+            text_input_chunk[:, 0] = text_input[0, 0]
+            if pad == eos:  # v1
+                text_input_chunk[:, -1] = text_input[0, -1]
+            else:  # v2
+                for j in range(len(text_input_chunk)):
+                    if text_input_chunk[j, -1] != eos and text_input_chunk[j, -1] != pad:  # 最後に普通の文字がある
+                        text_input_chunk[j, -1] = eos
+                    if text_input_chunk[j, 1] == pad:  # BOSだけであとはPAD
+                        text_input_chunk[j, 1] = eos
+
+            # in sdxl, value of clip_skip is same for Text Encoder 1 and 2
+            enc_out = text_encoder(text_input_chunk, output_hidden_states=True, return_dict=True)
+            text_embedding = enc_out["hidden_states"][-clip_skip]
+            if not is_sdxl:  # SD 1.5 requires final_layer_norm
+                text_embedding = text_encoder.text_model.final_layer_norm(text_embedding)
+            if pool is None:
+                pool = enc_out.get("text_embeds", None)  # use 1st chunk, if provided
+                if pool is not None:
+                    pool = train_util.pool_workaround(text_encoder, enc_out["last_hidden_state"], text_input_chunk, eos)
+
+            if no_boseos_middle:
+                if i == 0:
+                    # discard the ending token
+                    text_embedding = text_embedding[:, :-1]
+                elif i == max_embeddings_multiples - 1:
+                    # discard the starting token
+                    text_embedding = text_embedding[:, 1:]
+                else:
+                    # discard both starting and ending tokens
+                    text_embedding = text_embedding[:, 1:-1]
+
+            text_embeddings.append(text_embedding)
+        text_embeddings = torch.concat(text_embeddings, axis=1)
+    else:
+        enc_out = text_encoder(text_input, output_hidden_states=True, return_dict=True)
+        text_embeddings = enc_out["hidden_states"][-clip_skip]
         if not is_sdxl:  # SD 1.5 requires final_layer_norm
-            text_embedding = text_encoder.text_model.final_layer_norm(text_embedding)
-        text_embeddings.append(text_embedding)
-        if pool is None:
-            pool = enc_out.get("text_embeds", None)  # text encoder 1 doesn't return this
-            if pool is not None:
-                pool = train_util.pool_workaround(text_encoder, enc_out["last_hidden_state"], text_input, eos)
-        
-    text_embeddings = torch.cat(text_embeddings, dim=1)
+            text_embeddings = text_encoder.text_model.final_layer_norm(text_embeddings)
+        pool = enc_out.get("text_embeds", None)  # text encoder 1 doesn't return this
+        if pool is not None:
+            pool = train_util.pool_workaround(text_encoder, enc_out["last_hidden_state"], text_input, eos)
     return text_embeddings, pool
 
 

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/ipex/__init__.py

@@ -1,6 +1,7 @@
 import os
 import sys
 import torch
+from packaging import version
 try:
     import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
     has_ipex = True
@@ -8,7 +9,7 @@ except Exception:
     has_ipex = False
 from .hijacks import ipex_hijacks
 
-torch_version = float(torch.__version__[:3])
+torch_version = version.parse(torch.__version__)
 
 # pylint: disable=protected-access, missing-function-docstring, line-too-long
 
@@ -56,7 +57,6 @@ def ipex_init(): # pylint: disable=too-many-statements
             torch.cuda.__path__ = torch.xpu.__path__
             torch.cuda.set_stream = torch.xpu.set_stream
             torch.cuda.torch = torch.xpu.torch
-            torch.cuda.Union = torch.xpu.Union
             torch.cuda.__annotations__ = torch.xpu.__annotations__
             torch.cuda.__package__ = torch.xpu.__package__
             torch.cuda.__builtins__ = torch.xpu.__builtins__
@@ -64,14 +64,12 @@ def ipex_init(): # pylint: disable=too-many-statements
             torch.cuda.StreamContext = torch.xpu.StreamContext
             torch.cuda._lazy_call = torch.xpu._lazy_call
             torch.cuda.random = torch.xpu.random
-            torch.cuda._device = torch.xpu._device
             torch.cuda.__name__ = torch.xpu.__name__
-            torch.cuda._device_t = torch.xpu._device_t
             torch.cuda.__spec__ = torch.xpu.__spec__
             torch.cuda.__file__ = torch.xpu.__file__
             # torch.cuda.is_current_stream_capturing = torch.xpu.is_current_stream_capturing
 
-            if torch_version < 2.3:
+            if torch_version < version.parse("2.3"):
                 torch.cuda._initialization_lock = torch.xpu.lazy_init._initialization_lock
                 torch.cuda._initialized = torch.xpu.lazy_init._initialized
                 torch.cuda._is_in_bad_fork = torch.xpu.lazy_init._is_in_bad_fork
@@ -114,17 +112,22 @@ def ipex_init(): # pylint: disable=too-many-statements
                 torch.cuda.threading = torch.xpu.threading
                 torch.cuda.traceback = torch.xpu.traceback
 
-            if torch_version < 2.5:
+            if torch_version < version.parse("2.5"):
                 torch.cuda.os = torch.xpu.os
                 torch.cuda.Device = torch.xpu.Device
                 torch.cuda.warnings = torch.xpu.warnings
                 torch.cuda.classproperty = torch.xpu.classproperty
                 torch.UntypedStorage.cuda = torch.UntypedStorage.xpu
 
-            if torch_version < 2.7:
+            if torch_version < version.parse("2.7"):
                 torch.cuda.Tuple = torch.xpu.Tuple
                 torch.cuda.List = torch.xpu.List
 
+            if torch_version < version.parse("2.11"):
+                torch.cuda._device_t = torch.xpu._device_t
+                torch.cuda._device = torch.xpu._device
+                torch.cuda.Union = torch.xpu.Union
+
 
             # Memory:
             if 'linux' in sys.platform and "WSL2" in os.popen("uname -a").read():
@@ -160,7 +163,7 @@ def ipex_init(): # pylint: disable=too-many-statements
             torch.cuda.initial_seed = torch.xpu.initial_seed
 
             # C
-            if torch_version < 2.3:
+            if torch_version < version.parse("2.3"):
                 torch._C._cuda_getCurrentRawStream = ipex._C._getCurrentRawStream
                 ipex._C._DeviceProperties.multi_processor_count = ipex._C._DeviceProperties.gpu_subslice_count
                 ipex._C._DeviceProperties.major = 12