Merge b4b35c34bd into fa53f71ec0

* fix: improve numerical stability by conditionally using float32 in block computations

* doc: update README for improvement stability for fp16 training on Anima in version 0.10.3

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/strategy_sd.py

@@ -31,81 +31,171 @@ class SdTokenizeStrategy(TokenizeStrategy):
             )
         else:
             self.tokenizer = self._load_tokenizer(CLIPTokenizer, TOKENIZER_ID, tokenizer_cache_dir=tokenizer_cache_dir)
-
         if max_length is None:
             self.max_length = self.tokenizer.model_max_length
         else:
             self.max_length = max_length + 2
-
+        
+        self.break_separator = "BREAK"
+    
+    def _split_on_break(self, text: str) -> List[str]:
+        """Split text on BREAK separator (case-sensitive), filtering empty segments."""
+        segments = text.split(self.break_separator)
+        # Filter out empty or whitespace-only segments
+        filtered = [seg.strip() for seg in segments if seg.strip()]
+        # Return at least one segment to maintain consistency
+        return filtered if filtered else [""]
+    
+    def _tokenize_segments(self, segments: List[str], weighted: bool = False) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        """Tokenize multiple segments and concatenate them."""
+        if len(segments) == 1:
+            # No BREAK present, use existing logic
+            if weighted:
+                return self._get_input_ids(self.tokenizer, segments[0], self.max_length, weighted=True)
+            else:
+                tokens = self._get_input_ids(self.tokenizer, segments[0], self.max_length)
+                return tokens, None
+        
+        # Multiple segments - tokenize each separately
+        all_tokens = []
+        all_weights = [] if weighted else None
+        
+        for segment in segments:
+            if weighted:
+                seg_tokens, seg_weights = self._get_input_ids(self.tokenizer, segment, self.max_length, weighted=True)
+                all_tokens.append(seg_tokens)
+                all_weights.append(seg_weights)
+            else:
+                seg_tokens = self._get_input_ids(self.tokenizer, segment, self.max_length)
+                all_tokens.append(seg_tokens)
+        
+        # Concatenate along the sequence dimension (dim=1 for tokens that are [batch, seq_len] or [n_chunks, seq_len])
+        combined_tokens = torch.cat(all_tokens, dim=1) if all_tokens[0].dim() == 2 else torch.cat(all_tokens, dim=0)
+        combined_weights = None
+        if weighted:
+            combined_weights = torch.cat(all_weights, dim=1) if all_weights[0].dim() == 2 else torch.cat(all_weights, dim=0)
+        
+        return combined_tokens, combined_weights
+    
     def tokenize(self, text: Union[str, List[str]]) -> List[torch.Tensor]:
         text = [text] if isinstance(text, str) else text
-        return [torch.stack([self._get_input_ids(self.tokenizer, t, self.max_length) for t in text], dim=0)]
-
+        
+        tokens_list = []
+        for t in text:
+            segments = self._split_on_break(t)
+            tokens, _ = self._tokenize_segments(segments, weighted=False)
+            tokens_list.append(tokens)
+        
+        # Pad tokens to same length for stacking
+        max_length = max(t.shape[-1] for t in tokens_list)
+        padded_tokens = []
+        for tokens in tokens_list:
+            if tokens.shape[-1] < max_length:
+                # Pad with pad_token_id
+                pad_size = max_length - tokens.shape[-1]
+                if tokens.dim() == 2:
+                    padding = torch.full((tokens.shape[0], pad_size), self.tokenizer.pad_token_id, dtype=tokens.dtype)
+                    tokens = torch.cat([tokens, padding], dim=1)
+                else:
+                    padding = torch.full((pad_size,), self.tokenizer.pad_token_id, dtype=tokens.dtype)
+                    tokens = torch.cat([tokens, padding], dim=0)
+            padded_tokens.append(tokens)
+        
+        return [torch.stack(padded_tokens, dim=0)]
+    
     def tokenize_with_weights(self, text: str | List[str]) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
         text = [text] if isinstance(text, str) else text
+        
         tokens_list = []
         weights_list = []
         for t in text:
-            tokens, weights = self._get_input_ids(self.tokenizer, t, self.max_length, weighted=True)
+            segments = self._split_on_break(t)
+            tokens, weights = self._tokenize_segments(segments, weighted=True)
             tokens_list.append(tokens)
             weights_list.append(weights)
+        
         return [torch.stack(tokens_list, dim=0)], [torch.stack(weights_list, dim=0)]
 
 
 class SdTextEncodingStrategy(TextEncodingStrategy):
     def __init__(self, clip_skip: Optional[int] = None) -> None:
         self.clip_skip = clip_skip
-
+    
+    def _encode_with_clip_skip(self, text_encoder: Any, tokens: torch.Tensor) -> torch.Tensor:
+        """Encode tokens with optional CLIP skip."""
+        if self.clip_skip is None:
+            return text_encoder(tokens)[0]
+        
+        enc_out = text_encoder(tokens, output_hidden_states=True, return_dict=True)
+        hidden_states = enc_out["hidden_states"][-self.clip_skip]
+        return text_encoder.text_model.final_layer_norm(hidden_states)
+    
+    def _reconstruct_embeddings(self, encoder_hidden_states: torch.Tensor, tokens: torch.Tensor,
+                                max_token_length: int, model_max_length: int, 
+                                tokenizer: Any) -> torch.Tensor:
+        """Reconstruct embeddings from chunked encoding."""
+        v1 = tokenizer.pad_token_id == tokenizer.eos_token_id
+        states_list = [encoder_hidden_states[:, 0].unsqueeze(1)]  # <BOS>
+        
+        if not v1:
+            # v2: <BOS>...<EOS> <PAD> ... の三連を <BOS>...<EOS> <PAD> ... へ戻す
+            for i in range(1, max_token_length, model_max_length):
+                chunk = encoder_hidden_states[:, i : i + model_max_length - 2]
+                if i > 0:
+                    for j in range(len(chunk)):
+                        if tokens[j, 1] == tokenizer.eos_token:
+                            chunk[j, 0] = chunk[j, 1]
+                states_list.append(chunk)
+            states_list.append(encoder_hidden_states[:, -1].unsqueeze(1))
+        else:
+            # v1: <BOS>...<EOS> の三連を <BOS>...<EOS> へ戻す
+            for i in range(1, max_token_length, model_max_length):
+                states_list.append(encoder_hidden_states[:, i : i + model_max_length - 2])
+            states_list.append(encoder_hidden_states[:, -1].unsqueeze(1))
+        
+        return torch.cat(states_list, dim=1)
+    
+    def _apply_weights_single_chunk(self, encoder_hidden_states: torch.Tensor, 
+                                     weights: torch.Tensor) -> torch.Tensor:
+        """Apply weights for single chunk case (no max_token_length)."""
+        return encoder_hidden_states * weights.squeeze(1).unsqueeze(2)
+    
+    def _apply_weights_multi_chunk(self, encoder_hidden_states: torch.Tensor, 
+                                    weights: torch.Tensor) -> torch.Tensor:
+        """Apply weights for multi-chunk case (with max_token_length)."""
+        for i in range(weights.shape[1]):
+            start_idx = i * 75 + 1
+            end_idx = i * 75 + 76
+            encoder_hidden_states[:, start_idx:end_idx] = (
+                encoder_hidden_states[:, start_idx:end_idx] * weights[:, i, 1:-1].unsqueeze(-1)
+            )
+        return encoder_hidden_states
+    
     def encode_tokens(
         self, tokenize_strategy: TokenizeStrategy, models: List[Any], tokens: List[torch.Tensor]
     ) -> List[torch.Tensor]:
         text_encoder = models[0]
         tokens = tokens[0]
         sd_tokenize_strategy = tokenize_strategy  # type: SdTokenizeStrategy
-
-        # tokens: b,n,77
+        
         b_size = tokens.size()[0]
         max_token_length = tokens.size()[1] * tokens.size()[2]
         model_max_length = sd_tokenize_strategy.tokenizer.model_max_length
-        tokens = tokens.reshape((-1, model_max_length))  # batch_size*3, 77
-
+        
+        tokens = tokens.reshape((-1, model_max_length))
         tokens = tokens.to(text_encoder.device)
-
-        if self.clip_skip is None:
-            encoder_hidden_states = text_encoder(tokens)[0]
-        else:
-            enc_out = text_encoder(tokens, output_hidden_states=True, return_dict=True)
-            encoder_hidden_states = enc_out["hidden_states"][-self.clip_skip]
-            encoder_hidden_states = text_encoder.text_model.final_layer_norm(encoder_hidden_states)
-
-        # bs*3, 77, 768 or 1024
+        
+        encoder_hidden_states = self._encode_with_clip_skip(text_encoder, tokens)
         encoder_hidden_states = encoder_hidden_states.reshape((b_size, -1, encoder_hidden_states.shape[-1]))
-
+        
         if max_token_length != model_max_length:
-            v1 = sd_tokenize_strategy.tokenizer.pad_token_id == sd_tokenize_strategy.tokenizer.eos_token_id
-            if not v1:
-                # v2: <BOS>...<EOS> <PAD> ... の三連を <BOS>...<EOS> <PAD> ... へ戻す　正直この実装でいいのかわからん
-                states_list = [encoder_hidden_states[:, 0].unsqueeze(1)]  # <BOS>
-                for i in range(1, max_token_length, model_max_length):
-                    chunk = encoder_hidden_states[:, i : i + model_max_length - 2]  # <BOS> の後から 最後の前まで
-                    if i > 0:
-                        for j in range(len(chunk)):
-                            if tokens[j, 1] == sd_tokenize_strategy.tokenizer.eos_token:
-                                # 空、つまり <BOS> <EOS> <PAD> ...のパターン
-                                chunk[j, 0] = chunk[j, 1]  # 次の <PAD> の値をコピーする
-                    states_list.append(chunk)  # <BOS> の後から <EOS> の前まで
-                states_list.append(encoder_hidden_states[:, -1].unsqueeze(1))  # <EOS> か <PAD> のどちらか
-                encoder_hidden_states = torch.cat(states_list, dim=1)
-            else:
-                # v1: <BOS>...<EOS> の三連を <BOS>...<EOS> へ戻す
-                states_list = [encoder_hidden_states[:, 0].unsqueeze(1)]  # <BOS>
-                for i in range(1, max_token_length, model_max_length):
-                    states_list.append(encoder_hidden_states[:, i : i + model_max_length - 2])  # <BOS> の後から <EOS> の前まで
-                states_list.append(encoder_hidden_states[:, -1].unsqueeze(1))  # <EOS>
-                encoder_hidden_states = torch.cat(states_list, dim=1)
-
+            encoder_hidden_states = self._reconstruct_embeddings(
+                encoder_hidden_states, tokens, max_token_length, 
+                model_max_length, sd_tokenize_strategy.tokenizer
+            )
+        
         return [encoder_hidden_states]
-
+    
     def encode_tokens_with_weights(
         self,
         tokenize_strategy: TokenizeStrategy,
@@ -114,23 +204,15 @@ class SdTextEncodingStrategy(TextEncodingStrategy):
         weights_list: List[torch.Tensor],
     ) -> List[torch.Tensor]:
         encoder_hidden_states = self.encode_tokens(tokenize_strategy, models, tokens_list)[0]
-
         weights = weights_list[0].to(encoder_hidden_states.device)
-
-        # apply weights
-        if weights.shape[1] == 1:  # no max_token_length
-            # weights: ((b, 1, 77), (b, 1, 77)), hidden_states: (b, 77, 768), (b, 77, 768)
-            encoder_hidden_states = encoder_hidden_states * weights.squeeze(1).unsqueeze(2)
+        
+        if weights.shape[1] == 1:
+            encoder_hidden_states = self._apply_weights_single_chunk(encoder_hidden_states, weights)
         else:
-            # weights: ((b, n, 77), (b, n, 77)), hidden_states: (b, n*75+2, 768), (b, n*75+2, 768)
-            for i in range(weights.shape[1]):
-                encoder_hidden_states[:, i * 75 + 1 : i * 75 + 76] = encoder_hidden_states[:, i * 75 + 1 : i * 75 + 76] * weights[
-                    :, i, 1:-1
-                ].unsqueeze(-1)
-
+            encoder_hidden_states = self._apply_weights_multi_chunk(encoder_hidden_states, weights)
+        
         return [encoder_hidden_states]
 
-
 class SdSdxlLatentsCachingStrategy(LatentsCachingStrategy):
     # sd and sdxl share the same strategy. we can make them separate, but the difference is only the suffix.
     # and we keep the old npz for the backward compatibility.

tests/library/test_strategy_sd_text_encoding.py

@@ -0,0 +1,140 @@
+import pytest
+import torch
+from unittest.mock import Mock
+
+from library.strategy_sd import SdTextEncodingStrategy
+
+
+class TestSdTextEncodingStrategy:
+    @pytest.fixture
+    def strategy(self):
+        """Create strategy instance with default settings."""
+        return SdTextEncodingStrategy(clip_skip=None)
+
+    @pytest.fixture
+    def strategy_with_clip_skip(self):
+        """Create strategy instance with CLIP skip enabled."""
+        return SdTextEncodingStrategy(clip_skip=2)
+
+    @pytest.fixture
+    def mock_tokenizer(self):
+        """Create a mock tokenizer."""
+        tokenizer = Mock()
+        tokenizer.model_max_length = 77
+        tokenizer.pad_token_id = 0
+        tokenizer.eos_token = 2
+        tokenizer.eos_token_id = 2
+        return tokenizer
+
+    @pytest.fixture
+    def mock_text_encoder(self):
+        """Create a mock text encoder."""
+        encoder = Mock()
+        encoder.device = torch.device("cpu")
+
+        def encode_side_effect(tokens, output_hidden_states=False, return_dict=False):
+            batch_size = tokens.shape[0]
+            seq_len = tokens.shape[1]
+            hidden_size = 768
+
+            # Create deterministic hidden states
+            hidden_state = torch.ones(batch_size, seq_len, hidden_size) * 0.5
+
+            if return_dict:
+                result = {
+                    "hidden_states": [
+                        hidden_state * 0.8,
+                        hidden_state * 0.9,
+                        hidden_state * 1.0,
+                    ]
+                }
+                return result
+            else:
+                return [hidden_state]
+
+        encoder.side_effect = encode_side_effect
+        encoder.text_model = Mock()
+        encoder.text_model.final_layer_norm = lambda x: x
+
+        return encoder
+
+    @pytest.fixture
+    def mock_tokenize_strategy(self, mock_tokenizer):
+        """Create a mock tokenize strategy."""
+        strategy = Mock()
+        strategy.tokenizer = mock_tokenizer
+        return strategy
+
+    # Test _encode_with_clip_skip
+    def test_encode_without_clip_skip(self, strategy, mock_text_encoder):
+        """Test encoding without CLIP skip."""
+        tokens = torch.arange(154).reshape(2, 77)
+        result = strategy._encode_with_clip_skip(mock_text_encoder, tokens)
+        assert result.shape == (2, 77, 768)
+        # Verify deterministic output
+        assert torch.allclose(result[0, 0, 0], torch.tensor(0.5))
+
+    def test_encode_with_clip_skip(self, strategy_with_clip_skip, mock_text_encoder):
+        """Test encoding with CLIP skip."""
+        tokens = torch.arange(154).reshape(2, 77)
+        result = strategy_with_clip_skip._encode_with_clip_skip(mock_text_encoder, tokens)
+        assert result.shape == (2, 77, 768)
+        # With clip_skip=2, should use second-to-last hidden state (0.5 * 0.9 = 0.45)
+        assert torch.allclose(result[0, 0, 0], torch.tensor(0.45))
+
+    # Test _apply_weights_single_chunk
+    def test_apply_weights_single_chunk(self, strategy):
+        """Test applying weights for single chunk case."""
+        encoder_hidden_states = torch.ones(2, 77, 768)
+        weights = torch.ones(2, 1, 77) * 0.5
+        result = strategy._apply_weights_single_chunk(encoder_hidden_states, weights)
+        assert result.shape == (2, 77, 768)
+        # Verify weights were applied: 1.0 * 0.5 = 0.5
+        assert torch.allclose(result[0, 0, 0], torch.tensor(0.5))
+
+    # Test _apply_weights_multi_chunk
+    def test_apply_weights_multi_chunk(self, strategy):
+        """Test applying weights for multi-chunk case."""
+        # Simulating 2 chunks: 2*75+2 = 152 tokens
+        encoder_hidden_states = torch.ones(2, 152, 768)
+        weights = torch.ones(2, 2, 77) * 0.5
+        result = strategy._apply_weights_multi_chunk(encoder_hidden_states, weights)
+        assert result.shape == (2, 152, 768)
+        # Check that weights were applied to middle sections
+        assert torch.allclose(result[0, 1, 0], torch.tensor(0.5))
+        assert torch.allclose(result[0, 76, 0], torch.tensor(0.5))
+
+    # Integration tests
+    def test_encode_tokens_basic(self, strategy, mock_tokenize_strategy, mock_text_encoder):
+        """Test basic token encoding flow."""
+        tokens = torch.arange(154).reshape(2, 1, 77)
+        models = [mock_text_encoder]
+        tokens_list = [tokens]
+
+        result = strategy.encode_tokens(mock_tokenize_strategy, models, tokens_list)
+
+        assert len(result) == 1
+        assert result[0].shape[0] == 2  # batch size
+        assert result[0].shape[2] == 768  # hidden size
+        # Verify deterministic output
+        assert torch.allclose(result[0][0, 0, 0], torch.tensor(0.5))
+
+    def test_encode_tokens_with_weights_single_chunk(self, strategy, mock_tokenize_strategy, mock_text_encoder):
+        """Test weighted encoding with single chunk."""
+        tokens = torch.arange(154).reshape(2, 1, 77)
+        weights = torch.ones(2, 1, 77) * 0.5
+        models = [mock_text_encoder]
+        tokens_list = [tokens]
+        weights_list = [weights]
+
+        result = strategy.encode_tokens_with_weights(mock_tokenize_strategy, models, tokens_list, weights_list)
+
+        assert len(result) == 1
+        assert result[0].shape[0] == 2
+        assert result[0].shape[2] == 768
+        # Verify weights were applied: 0.5 (encoder output) * 0.5 (weight) = 0.25
+        assert torch.allclose(result[0][0, 0, 0], torch.tensor(0.25))
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])

tests/library/test_strategy_sd_tokenize.py

@@ -0,0 +1,378 @@
+import pytest
+import torch
+from unittest.mock import Mock, patch
+
+from library.strategy_sd import SdTokenizeStrategy
+
+
+class TestSdTokenizeStrategy:
+    @pytest.fixture
+    def mock_tokenizer(self):
+        """Create a mock CLIP tokenizer."""
+        tokenizer = Mock()
+        tokenizer.model_max_length = 77
+        tokenizer.bos_token_id = 49406
+        tokenizer.eos_token_id = 49407
+        tokenizer.pad_token_id = 49407
+
+        def tokenize_side_effect(text, **kwargs):
+            # Simple mock: return incrementing IDs based on text length
+            # Real tokenizer would split into subwords
+            num_tokens = min(len(text.split()), 75)
+            input_ids = torch.arange(1, num_tokens + 1)
+
+            if kwargs.get("return_tensors") == "pt":
+                max_length = kwargs.get("max_length", 77)
+                padded = torch.cat(
+                    [
+                        torch.tensor([tokenizer.bos_token_id]),
+                        input_ids,
+                        torch.tensor([tokenizer.eos_token_id]),
+                        torch.full((max_length - num_tokens - 2,), tokenizer.pad_token_id),
+                    ]
+                )
+                return Mock(input_ids=padded.unsqueeze(0))
+            else:
+                return Mock(
+                    input_ids=torch.cat([torch.tensor([tokenizer.bos_token_id]), input_ids, torch.tensor([tokenizer.eos_token_id])])
+                )
+
+        tokenizer.side_effect = tokenize_side_effect
+        return tokenizer
+
+    @pytest.fixture
+    def strategy_v1(self, mock_tokenizer):
+        """Create a v1 strategy instance with mocked tokenizer."""
+        with patch.object(SdTokenizeStrategy, "_load_tokenizer", return_value=mock_tokenizer):
+            strategy = SdTokenizeStrategy(v2=False, max_length=75, tokenizer_cache_dir=None)
+            return strategy
+
+    @pytest.fixture
+    def strategy_v2(self, mock_tokenizer):
+        """Create a v2 strategy instance with mocked tokenizer."""
+        mock_tokenizer.pad_token_id = 0  # v2 has different pad token
+        with patch.object(SdTokenizeStrategy, "_load_tokenizer", return_value=mock_tokenizer):
+            strategy = SdTokenizeStrategy(v2=True, max_length=75, tokenizer_cache_dir=None)
+            return strategy
+
+    # Test _split_on_break
+    def test_split_on_break_no_break(self, strategy_v1):
+        """Test splitting when no BREAK is present."""
+        text = "a cat and a dog"
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 1
+        assert result[0] == "a cat and a dog"
+
+    def test_split_on_break_single_break(self, strategy_v1):
+        """Test splitting with single BREAK."""
+        text = "a cat BREAK a dog"
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 2
+        assert result[0] == "a cat"
+        assert result[1] == "a dog"
+
+    def test_split_on_break_multiple_breaks(self, strategy_v1):
+        """Test splitting with multiple BREAKs."""
+        text = "a cat BREAK a dog BREAK a bird"
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 3
+        assert result[0] == "a cat"
+        assert result[1] == "a dog"
+        assert result[2] == "a bird"
+
+    def test_split_on_break_case_sensitive(self, strategy_v1):
+        """Test that BREAK splitting is case-sensitive."""
+        text = "a cat break a dog"  # lowercase 'break' should not split
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 1
+        assert result[0] == "a cat break a dog"
+
+        text = "a cat Break a dog"  # mixed case should not split
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 1
+
+    def test_split_on_break_with_whitespace(self, strategy_v1):
+        """Test splitting with extra whitespace around BREAK."""
+        text = "a cat  BREAK  a dog"
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 2
+        assert result[0] == "a cat"
+        assert result[1] == "a dog"
+
+    def test_split_on_break_empty_segments(self, strategy_v1):
+        """Test splitting filters out empty segments."""
+        text = "BREAK a cat BREAK BREAK a dog BREAK"
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 2
+        assert result[0] == "a cat"
+        assert result[1] == "a dog"
+
+    def test_split_on_break_only_break(self, strategy_v1):
+        """Test splitting with only BREAK returns empty string."""
+        text = "BREAK"
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 1
+        assert result[0] == ""
+
+    def test_split_on_break_empty_string(self, strategy_v1):
+        """Test splitting empty string."""
+        text = ""
+        result = strategy_v1._split_on_break(text)
+        assert len(result) == 1
+        assert result[0] == ""
+
+    # Test tokenize without BREAK
+    def test_tokenize_single_text_no_break(self, strategy_v1):
+        """Test tokenizing single text without BREAK."""
+        text = "a cat"
+        result = strategy_v1.tokenize(text)
+        assert len(result) == 1
+        assert isinstance(result[0], torch.Tensor)
+        assert result[0].dim() == 3  # [batch, n_chunks, seq_len]
+
+    def test_tokenize_list_no_break(self, strategy_v1):
+        """Test tokenizing list of texts without BREAK."""
+        texts = ["a cat", "a dog"]
+        result = strategy_v1.tokenize(texts)
+        assert len(result) == 1
+        assert result[0].shape[0] == 2  # batch size
+
+    # Test tokenize with BREAK
+    def test_tokenize_single_break(self, strategy_v1):
+        """Test tokenizing text with single BREAK."""
+        text = "a cat BREAK a dog"
+        result = strategy_v1.tokenize(text)
+        assert len(result) == 1
+        assert isinstance(result[0], torch.Tensor)
+        # Should have concatenated tokens from both segments
+
+    def test_tokenize_multiple_breaks(self, strategy_v1):
+        """Test tokenizing text with multiple BREAKs."""
+        text = "a cat BREAK a dog BREAK a bird"
+        result = strategy_v1.tokenize(text)
+        assert len(result) == 1
+        assert isinstance(result[0], torch.Tensor)
+
+    def test_tokenize_list_with_breaks(self, strategy_v1):
+        """Test tokenizing list where some texts have BREAKs."""
+        texts = ["a cat BREAK a dog", "a bird"]
+        result = strategy_v1.tokenize(texts)
+        assert len(result) == 1
+        assert result[0].shape[0] == 2  # batch size
+
+    # Test tokenize_with_weights without BREAK
+    def test_tokenize_with_weights_no_break(self, strategy_v1):
+        """Test weighted tokenization without BREAK."""
+        text = "a cat"
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(text)
+        assert len(tokens_list) == 1
+        assert len(weights_list) == 1
+        assert isinstance(tokens_list[0], torch.Tensor)
+        assert isinstance(weights_list[0], torch.Tensor)
+        assert tokens_list[0].shape == weights_list[0].shape
+
+    def test_tokenize_with_weights_list_no_break(self, strategy_v1):
+        """Test weighted tokenization of list without BREAK."""
+        texts = ["a cat", "a dog"]
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(texts)
+        assert len(tokens_list) == 1
+        assert len(weights_list) == 1
+        assert tokens_list[0].shape[0] == 2  # batch size
+        assert tokens_list[0].shape == weights_list[0].shape
+
+    # Test tokenize_with_weights with BREAK
+    def test_tokenize_with_weights_single_break(self, strategy_v1):
+        """Test weighted tokenization with single BREAK."""
+        text = "a cat BREAK a dog"
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(text)
+        assert len(tokens_list) == 1
+        assert len(weights_list) == 1
+        assert isinstance(tokens_list[0], torch.Tensor)
+        assert isinstance(weights_list[0], torch.Tensor)
+        assert tokens_list[0].shape == weights_list[0].shape
+
+    def test_tokenize_with_weights_multiple_breaks(self, strategy_v1):
+        """Test weighted tokenization with multiple BREAKs."""
+        text = "a cat BREAK a dog BREAK a bird"
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(text)
+        assert len(tokens_list) == 1
+        assert len(weights_list) == 1
+        assert tokens_list[0].shape == weights_list[0].shape
+
+    def test_tokenize_with_weights_list_with_breaks(self, strategy_v1):
+        """Test weighted tokenization of list with BREAKs."""
+        texts = ["a cat BREAK a dog", "a bird BREAK a fish"]
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(texts)
+        assert len(tokens_list) == 1
+        assert len(weights_list) == 1
+        assert tokens_list[0].shape[0] == 2  # batch size
+        assert tokens_list[0].shape == weights_list[0].shape
+
+    # Test weighted prompts (with attention syntax)
+    def test_tokenize_with_weights_attention_syntax(self, strategy_v1):
+        """Test weighted tokenization with attention syntax like (word:1.5)."""
+        text = "a (cat:1.5) and a dog"
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(text)
+        assert len(tokens_list) == 1
+        assert len(weights_list) == 1
+        # Weights should differ from 1.0 for the emphasized word
+
+    def test_tokenize_with_weights_attention_and_break(self, strategy_v1):
+        """Test weighted tokenization with both attention syntax and BREAK."""
+        text = "a (cat:1.5) BREAK a [dog:0.8]"
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(text)
+        assert len(tokens_list) == 1
+        assert len(weights_list) == 1
+        assert tokens_list[0].shape == weights_list[0].shape
+
+    def test_break_splits_long_prompts_into_chunks(self, strategy_v1):
+        """Test that BREAK causes long prompts to split into expected number of chunks."""
+        # Create a prompt with 80 tokens before BREAK and 80 after
+        # Each "word" typically becomes 1-2 tokens, so ~40-80 words for 80 tokens
+        long_segment = " ".join([f"word{i}" for i in range(40)])  # ~80 tokens
+        text = f"{long_segment} BREAK {long_segment}"
+        
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(text)
+        
+        # With model_max_length=77, we expect:
+        # - First segment: 80 tokens -> needs 2 chunks (77 + remainder)
+        # - Second segment: 80 tokens -> needs 2 chunks (77 + remainder)
+        # Total: 4 chunks (2 per segment)
+        
+        assert len(tokens_list) == 1
+        assert len(weights_list) == 1
+        
+        # Check that we got multiple chunks by looking at the shape
+        # The concatenated result should be longer than a single chunk (77 tokens)
+        tokens = tokens_list[0]
+        weights = weights_list[0]
+        
+        # Should have significantly more than 77 tokens due to concatenation
+        assert tokens.shape[-1] > 77, f"Expected >77 tokens but got {tokens.shape[-1]}"
+        
+        # With 2 segments of ~80 tokens each, we expect ~160 total tokens after concatenation
+        # (exact number depends on tokenizer behavior, but should be in this range)
+        assert tokens.shape[-1] >= 150, f"Expected >=150 tokens for 2 long segments but got {tokens.shape[-1]}"
+    
+    def test_break_splits_result_in_proper_chunks(self, strategy_v1):
+        """Test that BREAK splitting results in proper chunk structure."""
+        # Segment 1: ~40 tokens, Segment 2: ~40 tokens
+        segment1 = " ".join([f"word{i}" for i in range(20)])
+        segment2 = " ".join([f"word{i}" for i in range(20, 40)])
+        text = f"{segment1} BREAK {segment2}"
+        
+        tokens_list, weights_list = strategy_v1.tokenize_with_weights(text)
+        
+        tokens = tokens_list[0]
+        weights = weights_list[0]
+        
+        # Should be concatenated from 2 segments
+        # Each segment fits in one chunk (< 77 tokens), so total should be ~80 tokens
+        assert tokens.shape == weights.shape
+        assert tokens.shape[-1] > 40, "Should have tokens from both segments"
+
+    # Test v1 vs v2
+    def test_v1_vs_v2_initialization(self, mock_tokenizer):
+        """Test that v1 and v2 are initialized differently."""
+        with patch.object(SdTokenizeStrategy, "_load_tokenizer", return_value=mock_tokenizer):
+            strategy_v1 = SdTokenizeStrategy(v2=False, max_length=75)
+            strategy_v2 = SdTokenizeStrategy(v2=True, max_length=75)
+
+            assert strategy_v1.tokenizer is not None
+            assert strategy_v2.tokenizer is not None
+            assert strategy_v1.max_length == 77  # 75 + 2 for BOS/EOS
+            assert strategy_v2.max_length == 77
+
+    # Test max_length handling
+    def test_max_length_none(self, mock_tokenizer):
+        """Test that None max_length uses tokenizer's model_max_length."""
+        with patch.object(SdTokenizeStrategy, "_load_tokenizer", return_value=mock_tokenizer):
+            strategy = SdTokenizeStrategy(v2=False, max_length=None)
+            assert strategy.max_length == mock_tokenizer.model_max_length
+
+    def test_max_length_custom(self, mock_tokenizer):
+        """Test custom max_length."""
+        with patch.object(SdTokenizeStrategy, "_load_tokenizer", return_value=mock_tokenizer):
+            strategy = SdTokenizeStrategy(v2=False, max_length=150)
+            assert strategy.max_length == 152  # 150 + 2 for BOS/EOS
+
+
+class TestEdgeCases:
+    """Test edge cases for tokenization."""
+
+    @pytest.fixture
+    def mock_tokenizer(self):
+        """Create a mock CLIP tokenizer."""
+        tokenizer = Mock()
+        tokenizer.model_max_length = 77
+        tokenizer.bos_token_id = 49406
+        tokenizer.eos_token_id = 49407
+        tokenizer.pad_token_id = 49407
+
+        def tokenize_side_effect(text, **kwargs):
+            num_tokens = min(len(text.split()), 75)
+            input_ids = torch.arange(1, num_tokens + 1)
+
+            if kwargs.get("return_tensors") == "pt":
+                max_length = kwargs.get("max_length", 77)
+                padded = torch.cat(
+                    [
+                        torch.tensor([tokenizer.bos_token_id]),
+                        input_ids,
+                        torch.tensor([tokenizer.eos_token_id]),
+                        torch.full((max_length - num_tokens - 2,), tokenizer.pad_token_id),
+                    ]
+                )
+                return Mock(input_ids=padded.unsqueeze(0))
+            else:
+                return Mock(
+                    input_ids=torch.cat([torch.tensor([tokenizer.bos_token_id]), input_ids, torch.tensor([tokenizer.eos_token_id])])
+                )
+
+        tokenizer.side_effect = tokenize_side_effect
+        return tokenizer
+
+    def test_very_long_text_with_breaks(self, mock_tokenizer):
+        """Test very long text with multiple BREAKs."""
+        with patch.object(SdTokenizeStrategy, "_load_tokenizer", return_value=mock_tokenizer):
+            strategy = SdTokenizeStrategy(v2=False, max_length=75)
+            # Create long text segments
+            long_text = " ".join([f"word{i}" for i in range(50)])
+            text = f"{long_text} BREAK {long_text} BREAK {long_text}"
+
+            result = strategy.tokenize(text)
+            assert len(result) == 1
+            assert isinstance(result[0], torch.Tensor)
+
+    def test_break_at_boundaries(self, mock_tokenizer):
+        """Test BREAK at start and end of text."""
+        with patch.object(SdTokenizeStrategy, "_load_tokenizer", return_value=mock_tokenizer):
+            strategy = SdTokenizeStrategy(v2=False, max_length=75)
+
+            # BREAK at start
+            text = "BREAK a cat"
+            result = strategy.tokenize(text)
+            assert len(result) == 1
+
+            # BREAK at end
+            text = "a cat BREAK"
+            result = strategy.tokenize(text)
+            assert len(result) == 1
+
+            # BREAK at both ends
+            text = "BREAK a cat BREAK"
+            result = strategy.tokenize(text)
+            assert len(result) == 1
+
+    def test_consecutive_breaks(self, mock_tokenizer):
+        """Test multiple consecutive BREAKs."""
+        with patch.object(SdTokenizeStrategy, "_load_tokenizer", return_value=mock_tokenizer):
+            strategy = SdTokenizeStrategy(v2=False, max_length=75)
+            text = "a cat BREAK BREAK BREAK a dog"
+            result = strategy.tokenize(text)
+            assert len(result) == 1
+            # Should only create 2 segments (consecutive BREAKs create empty segments that are filtered)
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])