Diffusion Templates framework

.gitignore

@@ -2,6 +2,7 @@
 /models
 /scripts
 /diffusers
+/.vscode
 *.pkl
 *.safetensors
 *.pth

diffsynth/diffusion/base_pipeline.py

@@ -9,6 +9,7 @@ from ..utils.lora import GeneralLoRALoader
 from ..models.model_loader import ModelPool
 from ..utils.controlnet import ControlNetInput
 from ..core.device import get_device_name, IS_NPU_AVAILABLE
+from .template import load_template_model, load_template_data_processor
 
 
 class PipelineUnit:
@@ -319,14 +320,21 @@ class BasePipeline(torch.nn.Module):
     
     
     def cfg_guided_model_fn(self, model_fn, cfg_scale, inputs_shared, inputs_posi, inputs_nega, **inputs_others):
+        # Positive side forward
         if inputs_shared.get("positive_only_lora", None) is not None:
-            self.clear_lora(verbose=0)
             self.load_lora(self.dit, state_dict=inputs_shared["positive_only_lora"], verbose=0)
         noise_pred_posi = model_fn(**inputs_posi, **inputs_shared, **inputs_others)
+        if inputs_shared.get("positive_only_lora", None) is not None:
+            self.clear_lora(verbose=0)
+            
         if cfg_scale != 1.0:
-            if inputs_shared.get("positive_only_lora", None) is not None:
-                self.clear_lora(verbose=0)
+            # Negative side forward
+            if inputs_shared.get("negative_only_lora", None) is not None:
+                self.load_lora(self.dit, state_dict=inputs_shared["negative_only_lora"], verbose=0)
             noise_pred_nega = model_fn(**inputs_nega, **inputs_shared, **inputs_others)
+            if inputs_shared.get("negative_only_lora", None) is not None:
+                self.clear_lora(verbose=0)
+            
             if isinstance(noise_pred_posi, tuple):
                 # Separately handling different output types of latents, eg. video and audio latents.
                 noise_pred = tuple(
@@ -340,6 +348,14 @@ class BasePipeline(torch.nn.Module):
         return noise_pred
     
 
+    def load_training_template_model(self, model_config: ModelConfig = None):
+        if model_config is not None:
+            model_config.download_if_necessary()
+            self.template_model = load_template_model(model_config.path, torch_dtype=self.torch_dtype, device=self.device)
+            self.template_data_processor = load_template_data_processor(model_config.path)()
+
+
+
 class PipelineUnitGraph:
     def __init__(self):
         pass

diffsynth/diffusion/parsers.py

@@ -60,6 +60,10 @@ def add_gradient_config(parser: argparse.ArgumentParser):
     parser.add_argument("--gradient_accumulation_steps", type=int, default=1, help="Gradient accumulation steps.")
     return parser
 
+def add_template_model_config(parser: argparse.ArgumentParser):
+    parser.add_argument("--template_model_id_or_path", type=str, default=None, help="Model ID of path of template models.")
+    return parser
+
 def add_general_config(parser: argparse.ArgumentParser):
     parser = add_dataset_base_config(parser)
     parser = add_model_config(parser)
@@ -67,4 +71,5 @@ def add_general_config(parser: argparse.ArgumentParser):
     parser = add_output_config(parser)
     parser = add_lora_config(parser)
     parser = add_gradient_config(parser)
+    parser = add_template_model_config(parser)
     return parser

diffsynth/diffusion/template.py

@@ -0,0 +1,176 @@
+import torch, os, importlib, warnings, json, inspect
+from typing import Dict, List, Tuple, Union
+from ..core import ModelConfig, load_model
+from ..core.device.npu_compatible_device import get_device_type
+
+
+KVCache = Dict[str, Tuple[torch.Tensor, torch.Tensor]]
+
+
+class TemplateModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+    
+    @torch.no_grad()
+    def process_inputs(self, **kwargs):
+        return {}
+    
+    def forward(self, **kwargs):
+        raise NotImplementedError()
+
+
+def check_template_model_format(model):
+    if not hasattr(model, "process_inputs"):
+        raise NotImplementedError("`process_inputs` is not implemented in the Template model.")
+    if "kwargs" not in inspect.signature(model.process_inputs).parameters:
+        raise NotImplementedError("`**kwargs` is not included in `process_inputs`.")
+    if not hasattr(model, "forward"):
+        raise NotImplementedError("`forward` is not implemented in the Template model.")
+    if "kwargs" not in inspect.signature(model.forward).parameters:
+        raise NotImplementedError("`**kwargs` is not included in `forward`.")
+
+
+def load_template_model(path, torch_dtype=torch.bfloat16, device="cuda", verbose=1):
+    spec = importlib.util.spec_from_file_location("template_model", os.path.join(path, "model.py"))
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    template_model_path = getattr(module, 'TEMPLATE_MODEL_PATH') if hasattr(module, 'TEMPLATE_MODEL_PATH') else None
+    if template_model_path is not None:
+        # With `TEMPLATE_MODEL_PATH`, a pretrained model will be loaded.
+        model = load_model(
+            model_class=getattr(module, 'TEMPLATE_MODEL'),
+            config=getattr(module, 'TEMPLATE_MODEL_CONFIG') if hasattr(module, 'TEMPLATE_MODEL_CONFIG') else None,
+            path=os.path.join(path, getattr(module, 'TEMPLATE_MODEL_PATH')),
+            torch_dtype=torch_dtype,
+            device=device,
+        )
+    else:
+        # Without `TEMPLATE_MODEL_PATH`, a randomly initialized model or a non-model module will be loaded.
+        model = module.TEMPLATE_MODEL()
+        if hasattr(model, "to"):
+            model = model.to(dtype=torch_dtype, device=device)
+        if hasattr(model, "eval"):
+            model = model.eval()
+    check_template_model_format(model)
+    if verbose > 0:
+        metadata = {
+            "model_architecture": getattr(module, 'TEMPLATE_MODEL').__name__,
+            "code_path": os.path.join(path, "model.py"),
+            "weight_path": template_model_path,
+        }
+        print(f"Template model loaded: {json.dumps(metadata, indent=4)}")
+    return model
+
+
+def load_template_data_processor(path):
+    spec = importlib.util.spec_from_file_location("template_model", os.path.join(path, "model.py"))
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    if hasattr(module, 'TEMPLATE_DATA_PROCESSOR'):
+        processor = getattr(module, 'TEMPLATE_DATA_PROCESSOR')
+        return processor
+    else:
+        return None
+
+
+class TemplatePipeline(torch.nn.Module):
+    def __init__(self, models: List[TemplateModel]):
+        super().__init__()
+        self.models = torch.nn.ModuleList(models)
+
+    def merge_kv_cache(self, kv_cache_list: List[KVCache]) -> KVCache:
+        names = {}
+        for kv_cache in kv_cache_list:
+            for name in kv_cache:
+                names[name] = None
+        kv_cache_merged = {}
+        for name in names:
+            kv_list = [kv_cache.get(name) for kv_cache in kv_cache_list]
+            kv_list = [kv for kv in kv_list if kv is not None]
+            if len(kv_list) > 0:
+                k = torch.concat([kv[0] for kv in kv_list], dim=1)
+                v = torch.concat([kv[1] for kv in kv_list], dim=1)
+                kv_cache_merged[name] = (k, v)
+        return kv_cache_merged
+    
+    def merge_template_cache(self, template_cache_list):
+        params = sorted(list(set(sum([list(template_cache.keys()) for template_cache in template_cache_list], []))))
+        template_cache_merged = {}
+        for param in params:
+            data = [template_cache[param] for template_cache in template_cache_list if param in template_cache]
+            if param == "kv_cache":
+                data = self.merge_kv_cache(data)
+            elif len(data) == 1:
+                data = data[0]
+            else:
+                print(f"Conflict detected: `{param}` appears in the outputs of multiple Template models. Only the first one will be retained.")
+                data = data[0]
+            template_cache_merged[param] = data
+        return template_cache_merged
+
+    @staticmethod
+    def check_vram_config(model_config: ModelConfig):
+        params = [
+            model_config.offload_device, model_config.offload_dtype,
+            model_config.onload_device, model_config.onload_dtype,
+            model_config.preparing_device, model_config.preparing_dtype,
+            model_config.computation_device, model_config.computation_dtype,
+        ]
+        for param in params:
+            if param is not None:
+                warnings.warn("TemplatePipeline doesn't support VRAM management. VRAM config will be ignored.")
+
+    @staticmethod
+    def from_pretrained(
+        torch_dtype: torch.dtype = torch.bfloat16,
+        device: Union[str, torch.device] = get_device_type(),
+        model_configs: list[ModelConfig] = [],
+    ):
+        models = []
+        for model_config in model_configs:
+            TemplatePipeline.check_vram_config(model_config)
+            model_config.download_if_necessary()
+            model = load_template_model(model_config.path, torch_dtype=torch_dtype, device=device)
+            models.append(model)
+        pipe = TemplatePipeline(models)
+        return pipe
+    
+    @torch.no_grad()
+    def process_inputs(self, inputs: List[Dict], pipe=None, **kwargs):
+        return [(i.get("model_id", 0), self.models[i.get("model_id", 0)].process_inputs(pipe=pipe, **i)) for i in inputs]
+    
+    def forward(self, inputs: List[Tuple[int, Dict]], pipe=None, **kwargs):
+        template_cache = []
+        for model_id, model_inputs in inputs:
+            kv_cache = self.models[model_id](pipe=pipe, **model_inputs)
+            template_cache.append(kv_cache)
+        return template_cache
+    
+    def call_single_side(self, pipe=None, inputs: List[Dict] = None):
+        inputs = self.process_inputs(pipe=pipe, inputs=inputs)
+        template_cache = self.forward(pipe=pipe, inputs=inputs)
+        template_cache = self.merge_template_cache(template_cache)
+        return template_cache
+        
+    @torch.no_grad()
+    def __call__(
+        self,
+        pipe=None,
+        template_inputs: List[Dict] = None,
+        negative_template_inputs: List[Dict] = None,
+        **kwargs,
+    ):
+        template_cache = self.call_single_side(pipe=pipe, inputs=template_inputs or [])
+        negative_template_cache = self.call_single_side(pipe=pipe, inputs=negative_template_inputs or [])
+        required_params = list(inspect.signature(pipe.__call__).parameters.keys())
+        for param in template_cache:
+            if param in required_params:
+                kwargs[param] = template_cache[param]
+            else:
+                print(f"`{param}` is not included in the inputs of `{pipe.__class__.__name__}`. This parameter will be ignored.")
+        for param in negative_template_cache:
+            if "negative_" + param in required_params:
+                kwargs["negative_" + param] = negative_template_cache[param]
+            else:
+                print(f"`{'negative_' + param}` is not included in the inputs of `{pipe.__class__.__name__}`. This parameter will be ignored.")
+        return pipe(**kwargs)

diffsynth/diffusion/training_module.py

@@ -6,6 +6,7 @@ from peft import LoraConfig, inject_adapter_in_model
 
 
 class GeneralUnit_RemoveCache(PipelineUnit):
+    # Only used for training
     def __init__(self, required_params=tuple(), force_remove_params_shared=tuple(), force_remove_params_posi=tuple(), force_remove_params_nega=tuple()):
         super().__init__(take_over=True)
         self.required_params = required_params
@@ -27,6 +28,47 @@ class GeneralUnit_RemoveCache(PipelineUnit):
         return inputs_shared, inputs_posi, inputs_nega
 
 
+class GeneralUnit_TemplateProcessInputs(PipelineUnit):
+    # Only used for training
+    def __init__(self, data_processor):
+        super().__init__(
+            input_params=("template_inputs",),
+            output_params=("template_inputs",),
+        )
+        self.data_processor = data_processor
+    
+    def process(self, pipe, template_inputs):
+        if not hasattr(pipe, "template_model"):
+            return {}
+        if self.data_processor is not None:
+            template_inputs = self.data_processor(**template_inputs)
+        template_inputs = pipe.template_model.process_inputs(pipe=pipe, **template_inputs)
+        return {"template_inputs": template_inputs}
+
+
+class GeneralUnit_TemplateForward(PipelineUnit):
+    # Only used for training
+    def __init__(self, use_gradient_checkpointing=False, use_gradient_checkpointing_offload=False):
+        super().__init__(
+            input_params=("template_inputs",),
+            output_params=("kv_cache",),
+            onload_model_names=("template_model",)
+        )
+        self.use_gradient_checkpointing = use_gradient_checkpointing
+        self.use_gradient_checkpointing_offload = use_gradient_checkpointing_offload
+    
+    def process(self, pipe, template_inputs):
+        if not hasattr(pipe, "template_model"):
+            return {}
+        template_cache = pipe.template_model.forward(
+            **template_inputs,
+            pipe=pipe,
+            use_gradient_checkpointing=self.use_gradient_checkpointing,
+            use_gradient_checkpointing_offload=self.use_gradient_checkpointing_offload,
+        )
+        return template_cache
+
+
 class DiffusionTrainingModule(torch.nn.Module):
     def __init__(self):
         super().__init__()
@@ -211,6 +253,16 @@ class DiffusionTrainingModule(torch.nn.Module):
         return lora_target_modules
     
 
+    def load_training_template_model(self, pipe, path_or_model_id, use_gradient_checkpointing=False, use_gradient_checkpointing_offload=False):
+        if path_or_model_id is None:
+            return pipe
+        model_config = self.parse_path_or_model_id(path_or_model_id)
+        pipe.load_training_template_model(model_config)
+        pipe.units.append(GeneralUnit_TemplateProcessInputs(pipe.template_data_processor))
+        pipe.units.append(GeneralUnit_TemplateForward(use_gradient_checkpointing, use_gradient_checkpointing_offload))
+        return pipe
+
+
     def switch_pipe_to_training_mode(
         self,
         pipe,

diffsynth/models/flux2_dit.py

@@ -364,78 +364,7 @@ class Flux2FeedForward(nn.Module):
         return x
 
 
-class Flux2AttnProcessor:
-    _attention_backend = None
-    _parallel_config = None
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError(f"{self.__class__.__name__} requires PyTorch 2.0. Please upgrade your pytorch version.")
-
-    def __call__(
-        self,
-        attn: "Flux2Attention",
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: torch.Tensor = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-    ) -> torch.Tensor:
-        query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections(
-            attn, hidden_states, encoder_hidden_states
-        )
-
-        query = query.unflatten(-1, (attn.heads, -1))
-        key = key.unflatten(-1, (attn.heads, -1))
-        value = value.unflatten(-1, (attn.heads, -1))
-
-        query = attn.norm_q(query)
-        key = attn.norm_k(key)
-
-        if attn.added_kv_proj_dim is not None:
-            encoder_query = encoder_query.unflatten(-1, (attn.heads, -1))
-            encoder_key = encoder_key.unflatten(-1, (attn.heads, -1))
-            encoder_value = encoder_value.unflatten(-1, (attn.heads, -1))
-
-            encoder_query = attn.norm_added_q(encoder_query)
-            encoder_key = attn.norm_added_k(encoder_key)
-
-            query = torch.cat([encoder_query, query], dim=1)
-            key = torch.cat([encoder_key, key], dim=1)
-            value = torch.cat([encoder_value, value], dim=1)
-
-        if image_rotary_emb is not None:
-            query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
-            key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
-
-        query, key, value = query.to(hidden_states.dtype), key.to(hidden_states.dtype), value.to(hidden_states.dtype)
-        hidden_states = attention_forward(
-            query,
-            key,
-            value,
-            q_pattern="b s n d", k_pattern="b s n d", v_pattern="b s n d", out_pattern="b s n d",
-        )
-        hidden_states = hidden_states.flatten(2, 3)
-        hidden_states = hidden_states.to(query.dtype)
-
-        if encoder_hidden_states is not None:
-            encoder_hidden_states, hidden_states = hidden_states.split_with_sizes(
-                [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1
-            )
-            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
-
-        hidden_states = attn.to_out[0](hidden_states)
-        hidden_states = attn.to_out[1](hidden_states)
-
-        if encoder_hidden_states is not None:
-            return hidden_states, encoder_hidden_states
-        else:
-            return hidden_states
-
-
 class Flux2Attention(torch.nn.Module):
-    _default_processor_cls = Flux2AttnProcessor
-    _available_processors = [Flux2AttnProcessor]
-
     def __init__(
         self,
         query_dim: int,
@@ -449,7 +378,6 @@ class Flux2Attention(torch.nn.Module):
         eps: float = 1e-5,
         out_dim: int = None,
         elementwise_affine: bool = True,
-        processor=None,
     ):
         super().__init__()
 
@@ -485,59 +413,45 @@ class Flux2Attention(torch.nn.Module):
             self.add_v_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias)
             self.to_add_out = torch.nn.Linear(self.inner_dim, query_dim, bias=out_bias)
 
-        if processor is None:
-            processor = self._default_processor_cls()
-        self.processor = processor
-
     def forward(
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
         image_rotary_emb: Optional[torch.Tensor] = None,
+        kv_cache = None,
         **kwargs,
     ) -> torch.Tensor:
-        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
-        kwargs = {k: w for k, w in kwargs.items() if k in attn_parameters}
-        return self.processor(self, hidden_states, encoder_hidden_states, attention_mask, image_rotary_emb, **kwargs)
-
-
-class Flux2ParallelSelfAttnProcessor:
-    _attention_backend = None
-    _parallel_config = None
-
-    def __init__(self):
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError(f"{self.__class__.__name__} requires PyTorch 2.0. Please upgrade your pytorch version.")
-
-    def __call__(
-        self,
-        attn: "Flux2ParallelSelfAttention",
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        image_rotary_emb: Optional[torch.Tensor] = None,
-    ) -> torch.Tensor:
-        # Parallel in (QKV + MLP in) projection
-        hidden_states = attn.to_qkv_mlp_proj(hidden_states)
-        qkv, mlp_hidden_states = torch.split(
-            hidden_states, [3 * attn.inner_dim, attn.mlp_hidden_dim * attn.mlp_mult_factor], dim=-1
+        query, key, value, encoder_query, encoder_key, encoder_value = _get_qkv_projections(
+            self, hidden_states, encoder_hidden_states
         )
 
-        # Handle the attention logic
-        query, key, value = qkv.chunk(3, dim=-1)
+        query = query.unflatten(-1, (self.heads, -1))
+        key = key.unflatten(-1, (self.heads, -1))
+        value = value.unflatten(-1, (self.heads, -1))
 
-        query = query.unflatten(-1, (attn.heads, -1))
-        key = key.unflatten(-1, (attn.heads, -1))
-        value = value.unflatten(-1, (attn.heads, -1))
+        query = self.norm_q(query)
+        key = self.norm_k(key)
 
-        query = attn.norm_q(query)
-        key = attn.norm_k(key)
+        if self.added_kv_proj_dim is not None:
+            encoder_query = encoder_query.unflatten(-1, (self.heads, -1))
+            encoder_key = encoder_key.unflatten(-1, (self.heads, -1))
+            encoder_value = encoder_value.unflatten(-1, (self.heads, -1))
+
+            encoder_query = self.norm_added_q(encoder_query)
+            encoder_key = self.norm_added_k(encoder_key)
+
+            query = torch.cat([encoder_query, query], dim=1)
+            key = torch.cat([encoder_key, key], dim=1)
+            value = torch.cat([encoder_value, value], dim=1)
 
         if image_rotary_emb is not None:
             query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
             key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
 
-        query, key, value = query.to(hidden_states.dtype), key.to(hidden_states.dtype), value.to(hidden_states.dtype)
+        if kv_cache is not None:
+            key = torch.concat([key, kv_cache[0]], dim=1)
+            value = torch.concat([value, kv_cache[1]], dim=1)
         hidden_states = attention_forward(
             query,
             key,
@@ -547,30 +461,22 @@ class Flux2ParallelSelfAttnProcessor:
         hidden_states = hidden_states.flatten(2, 3)
         hidden_states = hidden_states.to(query.dtype)
 
-        # Handle the feedforward (FF) logic
-        mlp_hidden_states = attn.mlp_act_fn(mlp_hidden_states)
+        if encoder_hidden_states is not None:
+            encoder_hidden_states, hidden_states = hidden_states.split_with_sizes(
+                [encoder_hidden_states.shape[1], hidden_states.shape[1] - encoder_hidden_states.shape[1]], dim=1
+            )
+            encoder_hidden_states = self.to_add_out(encoder_hidden_states)
 
-        # Concatenate and parallel output projection
-        hidden_states = torch.cat([hidden_states, mlp_hidden_states], dim=-1)
-        hidden_states = attn.to_out(hidden_states)
+        hidden_states = self.to_out[0](hidden_states)
+        hidden_states = self.to_out[1](hidden_states)
 
-        return hidden_states
+        if encoder_hidden_states is not None:
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
 
 
 class Flux2ParallelSelfAttention(torch.nn.Module):
-    """
-    Flux 2 parallel self-attention for the Flux 2 single-stream transformer blocks.
-
-    This implements a parallel transformer block, where the attention QKV projections are fused to the feedforward (FF)
-    input projections, and the attention output projections are fused to the FF output projections. See the [ViT-22B
-    paper](https://arxiv.org/abs/2302.05442) for a visual depiction of this type of transformer block.
-    """
-
-    _default_processor_cls = Flux2ParallelSelfAttnProcessor
-    _available_processors = [Flux2ParallelSelfAttnProcessor]
-    # Does not support QKV fusion as the QKV projections are always fused
-    _supports_qkv_fusion = False
-
     def __init__(
         self,
         query_dim: int,
@@ -614,20 +520,54 @@ class Flux2ParallelSelfAttention(torch.nn.Module):
         # Fused attention output projection + MLP output projection
         self.to_out = torch.nn.Linear(self.inner_dim + self.mlp_hidden_dim, self.out_dim, bias=out_bias)
 
-        if processor is None:
-            processor = self._default_processor_cls()
-        self.processor = processor
-
     def forward(
         self,
         hidden_states: torch.Tensor,
         attention_mask: Optional[torch.Tensor] = None,
         image_rotary_emb: Optional[torch.Tensor] = None,
+        kv_cache = None,
         **kwargs,
     ) -> torch.Tensor:
-        attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
-        kwargs = {k: w for k, w in kwargs.items() if k in attn_parameters}
-        return self.processor(self, hidden_states, attention_mask, image_rotary_emb, **kwargs)
+        # Parallel in (QKV + MLP in) projection
+        hidden_states = self.to_qkv_mlp_proj(hidden_states)
+        qkv, mlp_hidden_states = torch.split(
+            hidden_states, [3 * self.inner_dim, self.mlp_hidden_dim * self.mlp_mult_factor], dim=-1
+        )
+
+        # Handle the attention logic
+        query, key, value = qkv.chunk(3, dim=-1)
+
+        query = query.unflatten(-1, (self.heads, -1))
+        key = key.unflatten(-1, (self.heads, -1))
+        value = value.unflatten(-1, (self.heads, -1))
+
+        query = self.norm_q(query)
+        key = self.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb, sequence_dim=1)
+            key = apply_rotary_emb(key, image_rotary_emb, sequence_dim=1)
+
+        if kv_cache is not None:
+            key = torch.concat([key, kv_cache[0]], dim=1)
+            value = torch.concat([value, kv_cache[1]], dim=1)
+        hidden_states = attention_forward(
+            query,
+            key,
+            value,
+            q_pattern="b s n d", k_pattern="b s n d", v_pattern="b s n d", out_pattern="b s n d",
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # Handle the feedforward (FF) logic
+        mlp_hidden_states = self.mlp_act_fn(mlp_hidden_states)
+
+        # Concatenate and parallel output projection
+        hidden_states = torch.cat([hidden_states, mlp_hidden_states], dim=-1)
+        hidden_states = self.to_out(hidden_states)
+
+        return hidden_states
 
 
 class Flux2SingleTransformerBlock(nn.Module):
@@ -657,7 +597,6 @@ class Flux2SingleTransformerBlock(nn.Module):
             eps=eps,
             mlp_ratio=mlp_ratio,
             mlp_mult_factor=2,
-            processor=Flux2ParallelSelfAttnProcessor(),
         )
 
     def forward(
@@ -669,6 +608,7 @@ class Flux2SingleTransformerBlock(nn.Module):
         joint_attention_kwargs: Optional[Dict[str, Any]] = None,
         split_hidden_states: bool = False,
         text_seq_len: Optional[int] = None,
+        kv_cache = None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         # If encoder_hidden_states is None, hidden_states is assumed to have encoder_hidden_states already
         # concatenated
@@ -685,6 +625,7 @@ class Flux2SingleTransformerBlock(nn.Module):
         attn_output = self.attn(
             hidden_states=norm_hidden_states,
             image_rotary_emb=image_rotary_emb,
+            kv_cache=kv_cache,
             **joint_attention_kwargs,
         )
 
@@ -725,7 +666,6 @@ class Flux2TransformerBlock(nn.Module):
             added_proj_bias=bias,
             out_bias=bias,
             eps=eps,
-            processor=Flux2AttnProcessor(),
         )
 
         self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
@@ -742,6 +682,7 @@ class Flux2TransformerBlock(nn.Module):
         temb_mod_params_txt: Tuple[Tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
         joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        kv_cache = None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         joint_attention_kwargs = joint_attention_kwargs or {}
 
@@ -762,6 +703,7 @@ class Flux2TransformerBlock(nn.Module):
             hidden_states=norm_hidden_states,
             encoder_hidden_states=norm_encoder_hidden_states,
             image_rotary_emb=image_rotary_emb,
+            kv_cache=kv_cache,
             **joint_attention_kwargs,
         )
 
@@ -969,6 +911,7 @@ class Flux2DiT(torch.nn.Module):
         txt_ids: torch.Tensor = None,
         guidance: torch.Tensor = None,
         joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        kv_cache = None,
         use_gradient_checkpointing=False,
         use_gradient_checkpointing_offload=False,
     ):
@@ -1013,7 +956,7 @@ class Flux2DiT(torch.nn.Module):
         )
 
         # 4. Double Stream Transformer Blocks
-        for index_block, block in enumerate(self.transformer_blocks):
+        for block_id, block in enumerate(self.transformer_blocks):
             encoder_hidden_states, hidden_states = gradient_checkpoint_forward(
                 block,
                 use_gradient_checkpointing=use_gradient_checkpointing,
@@ -1024,12 +967,13 @@ class Flux2DiT(torch.nn.Module):
                 temb_mod_params_txt=double_stream_mod_txt,
                 image_rotary_emb=concat_rotary_emb,
                 joint_attention_kwargs=joint_attention_kwargs,
+                kv_cache=None if kv_cache is None else kv_cache.get(f"double_{block_id}"),
             )
         # Concatenate text and image streams for single-block inference
         hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
         # 5. Single Stream Transformer Blocks
-        for index_block, block in enumerate(self.single_transformer_blocks):
+        for block_id, block in enumerate(self.single_transformer_blocks):
             hidden_states = gradient_checkpoint_forward(
                 block,
                 use_gradient_checkpointing=use_gradient_checkpointing,
@@ -1039,6 +983,7 @@ class Flux2DiT(torch.nn.Module):
                 temb_mod_params=single_stream_mod,
                 image_rotary_emb=concat_rotary_emb,
                 joint_attention_kwargs=joint_attention_kwargs,
+                kv_cache=None if kv_cache is None else kv_cache.get(f"single_{block_id}"),
             )
         # Remove text tokens from concatenated stream
         hidden_states = hidden_states[:, num_txt_tokens:, ...]

diffsynth/pipelines/flux2_image.py

@@ -40,6 +40,7 @@ class Flux2ImagePipeline(BasePipeline):
             Flux2Unit_InputImageEmbedder(),
             Flux2Unit_EditImageEmbedder(),
             Flux2Unit_ImageIDs(),
+            Flux2Unit_Inpaint(),
         ]
         self.model_fn = model_fn_flux2
     
@@ -93,6 +94,16 @@ class Flux2ImagePipeline(BasePipeline):
         initial_noise: torch.Tensor = None,
         # Steps
         num_inference_steps: int = 30,
+        # KV Cache
+        kv_cache = None,
+        negative_kv_cache = None,
+        # LoRA
+        lora = None,
+        negative_lora = None,
+        # Inpaint
+        inpaint_mask: Image.Image = None,
+        inpaint_blur_size: int = None,
+        inpaint_blur_sigma: float = None,
         # Progress bar
         progress_bar_cmd = tqdm,
     ):
@@ -101,9 +112,11 @@ class Flux2ImagePipeline(BasePipeline):
         # Parameters
         inputs_posi = {
             "prompt": prompt,
+            "kv_cache": kv_cache,
         }
         inputs_nega = {
             "negative_prompt": negative_prompt,
+            "kv_cache": negative_kv_cache,
         }
         inputs_shared = {
             "cfg_scale": cfg_scale, "embedded_guidance": embedded_guidance,
@@ -112,6 +125,9 @@ class Flux2ImagePipeline(BasePipeline):
             "height": height, "width": width,
             "seed": seed, "rand_device": rand_device, "initial_noise": initial_noise,
             "num_inference_steps": num_inference_steps,
+            "positive_only_lora": lora,
+            "negative_only_lora": negative_lora,
+            "inpaint_mask": inpaint_mask, "inpaint_blur_size": inpaint_blur_size, "inpaint_blur_sigma": inpaint_blur_sigma,
         }
         for unit in self.units:
             inputs_shared, inputs_posi, inputs_nega = self.unit_runner(unit, self, inputs_shared, inputs_posi, inputs_nega)
@@ -560,6 +576,26 @@ class Flux2Unit_ImageIDs(PipelineUnit):
         return {"image_ids": image_ids}
 
 
+class Flux2Unit_Inpaint(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("inpaint_mask", "height", "width", "inpaint_blur_size", "inpaint_blur_sigma"),
+            output_params=("inpaint_mask",),
+        )
+
+    def process(self, pipe: Flux2ImagePipeline, inpaint_mask, height, width, inpaint_blur_size, inpaint_blur_sigma):
+        if inpaint_mask is None:
+            return {}
+        inpaint_mask = pipe.preprocess_image(inpaint_mask.convert("RGB").resize((width // 16, height // 16)), min_value=0, max_value=1)
+        inpaint_mask = inpaint_mask.mean(dim=1, keepdim=True)
+        if inpaint_blur_size is not None and inpaint_blur_sigma is not None:
+            from torchvision.transforms import GaussianBlur
+            blur = GaussianBlur(kernel_size=inpaint_blur_size * 2 + 1, sigma=inpaint_blur_sigma)
+            inpaint_mask = blur(inpaint_mask)
+        inpaint_mask = rearrange(inpaint_mask, "B C H W -> B (H W) C")
+        return {"inpaint_mask": inpaint_mask}
+
+
 def model_fn_flux2(
     dit: Flux2DiT,
     latents=None,
@@ -570,6 +606,7 @@ def model_fn_flux2(
     image_ids=None,
     edit_latents=None,
     edit_image_ids=None,
+    kv_cache=None,
     use_gradient_checkpointing=False,
     use_gradient_checkpointing_offload=False,
     **kwargs,
@@ -587,6 +624,7 @@ def model_fn_flux2(
         encoder_hidden_states=prompt_embeds,
         txt_ids=text_ids,
         img_ids=image_ids,
+        kv_cache=kv_cache,
         use_gradient_checkpointing=use_gradient_checkpointing,
         use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
     )

examples/flux2/model_inference/Template-KleinBase4B.py

@@ -0,0 +1,256 @@
+from diffsynth.diffusion.template import TemplatePipeline
+from diffsynth.pipelines.flux2_image import Flux2ImagePipeline, ModelConfig
+import torch
+from PIL import Image
+import numpy as np
+
+def load_template_pipeline(model_ids):
+    template = TemplatePipeline.from_pretrained(
+        torch_dtype=torch.bfloat16,
+        device="cuda",
+        model_configs=[ModelConfig(model_id=model_id) for model_id in model_ids],
+    )
+    return template
+
+# Base Model
+pipe = Flux2ImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-base-4B", origin_file_pattern="transformer/*.safetensors"),
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="text_encoder/*.safetensors"),
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
+    ],
+    tokenizer_config=ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="tokenizer/"),
+)
+# image = pipe(
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+# )
+# image.save("image_base.jpg")
+
+# template = load_template_pipeline(["DiffSynth-Studio/Template-KleinBase4B-Brightness"])
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{"scale": 0.7}],
+#     negative_template_inputs = [{"scale": 0.5}]
+# )
+# image.save("image_Brightness_light.jpg")
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{"scale": 0.5}],
+#     negative_template_inputs = [{"scale": 0.5}]
+# )
+# image.save("image_Brightness_normal.jpg")
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{"scale": 0.3}],
+#     negative_template_inputs = [{"scale": 0.5}]
+# )
+# image.save("image_Brightness_dark.jpg")
+
+# template = load_template_pipeline(["DiffSynth-Studio/Template-KleinBase4B-ControlNet"])
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone, bathed in bright sunshine.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "image": Image.open("data/assets/image_depth.jpg"),
+#         "prompt": "A cat is sitting on a stone, bathed in bright sunshine.",
+#     }],
+#     negative_template_inputs = [{
+#         "image": Image.open("data/assets/image_depth.jpg"),
+#         "prompt": "",
+#     }],
+# )
+# image.save("image_ControlNet_sunshine.jpg")
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone, surrounded by colorful magical particles.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "image": Image.open("data/assets/image_depth.jpg"),
+#         "prompt": "A cat is sitting on a stone, surrounded by colorful magical particles.",
+#     }],
+#     negative_template_inputs = [{
+#         "image": Image.open("data/assets/image_depth.jpg"),
+#         "prompt": "",
+#     }],
+# )
+# image.save("image_ControlNet_magic.jpg")
+
+# template = load_template_pipeline(["DiffSynth-Studio/Template-KleinBase4B-Edit"])
+# image = template(
+#     pipe,
+#     prompt="Put a hat on this cat.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "image": Image.open("data/assets/image_reference.jpg"),
+#         "prompt": "Put a hat on this cat.",
+#     }],
+#     negative_template_inputs = [{
+#         "image": Image.open("data/assets/image_reference.jpg"),
+#         "prompt": "",
+#     }],
+# )
+# image.save("image_Edit_hat.jpg")
+# image = template(
+#     pipe,
+#     prompt="Make the cat turn its head to look to the right.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "image": Image.open("data/assets/image_reference.jpg"),
+#         "prompt": "Make the cat turn its head to look to the right.",
+#     }],
+#     negative_template_inputs = [{
+#         "image": Image.open("data/assets/image_reference.jpg"),
+#         "prompt": "",
+#     }],
+# )
+# image.save("image_Edit_head.jpg")
+
+# template = load_template_pipeline(["DiffSynth-Studio/Template-KleinBase4B-Upscaler"])
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "image": Image.open("data/assets/image_lowres_512.jpg"),
+#         "prompt": "A cat is sitting on a stone.",
+#     }],
+#     negative_template_inputs = [{
+#         "image": Image.open("data/assets/image_lowres_512.jpg"),
+#         "prompt": "",
+#     }],
+# )
+# image.save("image_Upscaler_1.png")
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "image": Image.open("data/assets/image_lowres_100.jpg"),
+#         "prompt": "A cat is sitting on a stone.",
+#     }],
+#     negative_template_inputs = [{
+#         "image": Image.open("data/assets/image_lowres_100.jpg"),
+#         "prompt": "",
+#     }],
+# )
+# image.save("image_Upscaler_2.png")
+
+# template = load_template_pipeline(["DiffSynth-Studio/Template-KleinBase4B-SoftRGB"])
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "R": 128/255,
+#         "G": 128/255,
+#         "B": 128/255
+#     }],
+# )
+# image.save("image_rgb_normal.jpg")
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "R": 208/255,
+#         "G": 185/255,
+#         "B": 138/255
+#     }],
+# )
+# image.save("image_rgb_warm.jpg")
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "R": 94/255,
+#         "G": 163/255,
+#         "B": 174/255
+#     }],
+# )
+# image.save("image_rgb_cold.jpg")
+
+# template = load_template_pipeline(["DiffSynth-Studio/Template-KleinBase4B-PandaMeme"])
+# image = template(
+#     pipe,
+#     prompt="A meme with a sleepy expression.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{}],
+#     negative_template_inputs = [{}],
+# )
+# image.save("image_PandaMeme_sleepy.jpg")
+# image = template(
+#     pipe,
+#     prompt="A meme with a happy expression.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{}],
+#     negative_template_inputs = [{}],
+# )
+# image.save("image_PandaMeme_happy.jpg")
+# image = template(
+#     pipe,
+#     prompt="A meme with a surprised expression.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{}],
+#     negative_template_inputs = [{}],
+# )
+# image.save("image_PandaMeme_surprised.jpg")
+
+# template = load_template_pipeline(["DiffSynth-Studio/Template-KleinBase4B-Sharpness"])
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{"scale": 0.1}],
+#     negative_template_inputs = [{"scale": 0.5}],
+# )
+# image.save("image_Sharpness_0.1.jpg")
+# image = template(
+#     pipe,
+#     prompt="A cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{"scale": 0.8}],
+#     negative_template_inputs = [{"scale": 0.5}],
+# )
+# image.save("image_Sharpness_0.8.jpg")
+
+# template = load_template_pipeline(["DiffSynth-Studio/Template-KleinBase4B-Inpaint"])
+# image = template(
+#     pipe,
+#     prompt="An orange cat is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "image": Image.open("data/assets/image_reference.jpg"),
+#         "mask": Image.open("data/assets/image_mask_1.jpg"),
+#         "force_inpaint": True,
+#     }],
+#     negative_template_inputs = [{
+#         "image": Image.open("data/assets/image_reference.jpg"),
+#         "mask": Image.open("data/assets/image_mask_1.jpg"),
+#     }],
+# )
+# image.save("image_Inpaint_1.jpg")
+# image = template(
+#     pipe,
+#     prompt="A cat wearing sunglasses is sitting on a stone.",
+#     seed=0, cfg_scale=4, num_inference_steps=50,
+#     template_inputs = [{
+#         "image": Image.open("data/assets/image_reference.jpg"),
+#         "mask": Image.open("data/assets/image_mask_2.jpg"),
+#     }],
+#     negative_template_inputs = [{
+#         "image": Image.open("data/assets/image_reference.jpg"),
+#         "mask": Image.open("data/assets/image_mask_2.jpg"),
+#     }],
+# )
+# image.save("image_Inpaint_2.jpg")

examples/flux2/model_training/full/Template-KleinBase4B.sh

@@ -0,0 +1,17 @@
+accelerate launch examples/flux2/model_training/train.py \
+  --dataset_base_path xxx \
+  --dataset_metadata_path xxx/metadata.jsonl \
+  --extra_inputs "template_inputs" \
+  --max_pixels 1048576 \
+  --dataset_repeat 1 \
+  --model_id_with_origin_paths "black-forest-labs/FLUX.2-klein-4B:text_encoder/*.safetensors,black-forest-labs/FLUX.2-klein-base-4B:transformer/*.safetensors,black-forest-labs/FLUX.2-klein-4B:vae/diffusion_pytorch_model.safetensors" \
+  --template_model_id_or_path "xxx" \
+  --tokenizer_path "black-forest-labs/FLUX.2-klein-4B:tokenizer/" \
+  --learning_rate 1e-4 \
+  --num_epochs 999 \
+  --remove_prefix_in_ckpt "pipe.template_model." \
+  --output_path "./models/train/Template-KleinBase4B_full" \
+  --trainable_models "template_model" \
+  --save_steps 1000 \
+  --use_gradient_checkpointing \
+  --find_unused_parameters

examples/flux2/model_training/scripts/convert_base_model_to_skill_model.py

@@ -0,0 +1,60 @@
+from diffsynth import load_state_dict
+from safetensors.torch import save_file
+import torch
+
+
+def Flux2DiTStateDictConverter(state_dict):
+    rename_dict = {
+        "time_guidance_embed.timestep_embedder.linear_1.weight": "time_guidance_embed.timestep_embedder.0.weight",
+        "time_guidance_embed.timestep_embedder.linear_2.weight": "time_guidance_embed.timestep_embedder.2.weight",
+        "x_embedder.weight": "img_embedder.weight",
+        "context_embedder.weight": "txt_embedder.weight",
+    }
+    state_dict_ = {}
+    for name in state_dict:
+        if name in rename_dict:
+            state_dict_[rename_dict[name]] = state_dict[name]
+        elif name.startswith("transformer_blocks"):
+            if name.endswith("attn.to_q.weight"):
+                state_dict_[name.replace("to_q", "img_to_qkv").replace(".attn.", ".")] = torch.concat([
+                    state_dict[name.replace("to_q", "to_q")],
+                    state_dict[name.replace("to_q", "to_k")],
+                    state_dict[name.replace("to_q", "to_v")],
+                ], dim=0)
+            elif name.endswith("attn.to_k.weight") or name.endswith("attn.to_v.weight"):
+                continue
+            elif name.endswith("attn.to_out.0.weight"):
+                state_dict_[name.replace("attn.to_out.0.weight", "img_to_out.weight")] = state_dict[name]
+            elif name.endswith("attn.norm_q.weight"):
+                state_dict_[name.replace("attn.norm_q.weight", "img_norm_q.weight")] = state_dict[name]
+            elif name.endswith("attn.norm_k.weight"):
+                state_dict_[name.replace("attn.norm_k.weight", "img_norm_k.weight")] = state_dict[name]
+            elif name.endswith("attn.norm_added_q.weight"):
+                state_dict_[name.replace("attn.norm_added_q.weight", "txt_norm_q.weight")] = state_dict[name]
+            elif name.endswith("attn.norm_added_k.weight"):
+                state_dict_[name.replace("attn.norm_added_k.weight", "txt_norm_k.weight")] = state_dict[name]
+            elif name.endswith("attn.to_add_out.weight"):
+                state_dict_[name.replace("attn.to_add_out.weight", "txt_to_out.weight")] = state_dict[name]
+            elif name.endswith("attn.add_q_proj.weight"):
+                state_dict_[name.replace("add_q_proj", "txt_to_qkv").replace(".attn.", ".")] = torch.concat([
+                    state_dict[name.replace("add_q_proj", "add_q_proj")],
+                    state_dict[name.replace("add_q_proj", "add_k_proj")],
+                    state_dict[name.replace("add_q_proj", "add_v_proj")],
+                ], dim=0)
+            elif ".ff." in name:
+                state_dict_[name.replace(".ff.", ".img_ff.")] = state_dict[name]
+            elif ".ff_context." in name:
+                state_dict_[name.replace(".ff_context.", ".txt_ff.")] = state_dict[name]
+            elif name.endswith("attn.add_k_proj.weight") or name.endswith("attn.add_v_proj.weight"):
+                continue
+            else:
+                state_dict_[name] = state_dict[name]
+        elif name.startswith("single_transformer_blocks"):
+            state_dict_[name.replace(".attn.", ".")] = state_dict[name]
+        else:
+            state_dict_[name] = state_dict[name]
+    return state_dict_
+
+
+state_dict = load_state_dict("xxx.safetensors")
+save_file(state_dict, "yyy.safetensors")

examples/flux2/model_training/train.py

@@ -18,6 +18,7 @@ class Flux2ImageTrainingModule(DiffusionTrainingModule):
         extra_inputs=None,
         fp8_models=None,
         offload_models=None,
+        template_model_id_or_path=None,
         device="cpu",
         task="sft",
     ):
@@ -26,6 +27,7 @@ class Flux2ImageTrainingModule(DiffusionTrainingModule):
         model_configs = self.parse_model_configs(model_paths, model_id_with_origin_paths, fp8_models=fp8_models, offload_models=offload_models, device=device)
         tokenizer_config = self.parse_path_or_model_id(tokenizer_path, default_value=ModelConfig(model_id="black-forest-labs/FLUX.2-dev", origin_file_pattern="tokenizer/"))
         self.pipe = Flux2ImagePipeline.from_pretrained(torch_dtype=torch.bfloat16, device=device, model_configs=model_configs, tokenizer_config=tokenizer_config)
+        self.pipe = self.load_training_template_model(self.pipe, template_model_id_or_path, args.use_gradient_checkpointing, args.use_gradient_checkpointing_offload)
         self.pipe = self.split_pipeline_units(task, self.pipe, trainable_models, lora_base_model)
 
         # Training mode
@@ -126,6 +128,7 @@ if __name__ == "__main__":
         extra_inputs=args.extra_inputs,
         fp8_models=args.fp8_models,
         offload_models=args.offload_models,
+        template_model_id_or_path=args.template_model_id_or_path,
         task=args.task,
         device="cpu" if args.initialize_model_on_cpu else accelerator.device,
     )