diffusion skills framework

2026-03-18 22:08:13 +00:00 · 2026-03-17 13:34:25 +08:00
parent 7a80f10fa4
commit f88b99cb4f
11 changed files with 422 additions and 138 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -2,6 +2,7 @@
 /models
 /scripts
 /diffusers
+/.vscode
 *.pkl
 *.safetensors
 *.pth
--- a/diffsynth/diffusion/base_pipeline.py
+++ b/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 .skills import load_skill_model, load_skill_data_processor


 class PipelineUnit:
@@ -338,6 +339,14 @@ class BasePipeline(torch.nn.Module):
        else:
            noise_pred = noise_pred_posi
        return noise_pred
+    
+
+    def load_training_skill_model(self, model_config: ModelConfig = None):
+        if model_config is not None:
+            model_config.download_if_necessary()
+            self.skill_model = load_skill_model(model_config.path, torch_dtype=self.torch_dtype, device=self.device)
+            self.skill_data_processor = load_skill_data_processor(model_config.path)()
+


 class PipelineUnitGraph:
--- a/diffsynth/diffusion/parsers.py
+++ b/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_skill_model_config(parser: argparse.ArgumentParser):
+    parser.add_argument("--skill_model_id_or_path", type=str, default=None, help="Model ID of path of skill 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_skill_model_config(parser)
    return parser
--- a/diffsynth/diffusion/skills.py
+++ b/diffsynth/diffusion/skills.py
@@ -0,0 +1,137 @@
+import torch, os, importlib, warnings, json
+from typing import Dict, List, Tuple, Union
+from ..core import ModelConfig, load_model
+from ..core.device.npu_compatible_device import get_device_type
+
+
+SkillCache = Dict[str, Tuple[torch.Tensor, torch.Tensor]]
+
+
+class SkillModel(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+    
+    @torch.no_grad()
+    def process_inputs(self, pipe=None, **kwargs):
+        return {}
+    
+    def forward(self, **kwargs) -> SkillCache:
+        raise NotImplementedError()
+
+
+class MultiSkillModel(SkillModel):
+    def __init__(self, models: List[SkillModel]):
+        super().__init__()
+        if not isinstance(models, list):
+            models = [models]
+        self.models = torch.nn.ModuleList(models)
+
+    def merge(self, kv_cache_list: List[SkillCache]) -> SkillCache:
+        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
+    
+    @torch.no_grad()
+    def process_inputs(self, pipe=None, inputs: List[Dict] = None, **kwargs):
+        return [(i["model_id"], self.models[i["model_id"]].process_inputs(pipe=pipe, **i)) for i in inputs]
+    
+    def forward(self, inputs: List[Tuple[int, Dict]], **kwargs) -> SkillCache:
+        kv_cache_list = []
+        for model_id, model_inputs in inputs:
+            kv_cache = self.models[model_id](**model_inputs)
+            kv_cache_list.append(kv_cache)
+        return self.merge(kv_cache_list)
+
+
+def load_skill_model(path, torch_dtype=torch.bfloat16, device="cuda", verbose=1):
+    spec = importlib.util.spec_from_file_location("skill_model", os.path.join(path, "model.py"))
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    model = load_model(
+        model_class=getattr(module, 'SKILL_MODEL'),
+        config=getattr(module, 'SKILL_MODEL_CONFIG') if hasattr(module, 'SKILL_MODEL_CONFIG') else None,
+        path=os.path.join(path, getattr(module, 'SKILL_MODEL_PATH')),
+        torch_dtype=torch_dtype,
+        device=device,
+    )
+    if verbose > 0:
+        metadata = {
+            "model_architecture": getattr(module, 'SKILL_MODEL').__name__,
+            "code_path": os.path.join(path, "model.py"),
+            "weight_path": os.path.join(path, getattr(module, 'SKILL_MODEL_PATH')),
+        }
+        print(f"Skill model loaded: {json.dumps(metadata, indent=4)}")
+    return model
+
+
+def load_skill_data_processor(path):
+    spec = importlib.util.spec_from_file_location("skill_model", os.path.join(path, "model.py"))
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    if hasattr(module, 'SKILL_DATA_PROCESSOR'):
+        processor = getattr(module, 'SKILL_DATA_PROCESSOR')
+        return processor
+    else:
+        return None
+
+
+class SkillsPipeline(MultiSkillModel):
+    def __init__(self, models: List[SkillModel]):
+        super().__init__(models)
+
+    @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("SkillsPipeline 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:
+            SkillsPipeline.check_vram_config(model_config)
+            model_config.download_if_necessary()
+            model = load_skill_model(model_config.path, torch_dtype=torch_dtype, device=device)
+            models.append(model)
+        pipe = SkillsPipeline(models)
+        return pipe
+    
+    def call_single_side(self, pipe = None, inputs: List[Dict] = None):
+        inputs = self.process_inputs(pipe=pipe, inputs=inputs)
+        skill_cache = self.forward(inputs)
+        return skill_cache
+        
+    @torch.no_grad()
+    def __call__(
+        self,
+        pipe = None,
+        inputs: List[Dict] = None,
+        positive_inputs: List[Dict] = None,
+        negative_inputs: List[Dict] = None,
+    ):
+        shared_cache = self.call_single_side(pipe=pipe, inputs=inputs or [])
+        positive_cache = self.call_single_side(pipe=pipe, inputs=positive_inputs or [])
+        negative_cache = self.call_single_side(pipe=pipe, inputs=negative_inputs or [])
+        positive_cache = self.merge([positive_cache, shared_cache])
+        negative_cache = self.merge([negative_cache, shared_cache])
+        return {"skill_cache": positive_cache, "negative_skill_cache": negative_cache}
--- a/diffsynth/diffusion/training_module.py
+++ b/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 +28,40 @@ class GeneralUnit_RemoveCache(PipelineUnit):
        return inputs_shared, inputs_posi, inputs_nega


+class GeneralUnit_SkillProcessInputs(PipelineUnit):
+    # Only used for training
+    def __init__(self, data_processor):
+        super().__init__(
+            input_params=("skill_inputs",),
+            output_params=("skill_inputs",),
+        )
+        self.data_processor = data_processor
+    
+    def process(self, pipe, skill_inputs):
+        if not hasattr(pipe, "skill_model"):
+            return {}
+        if self.data_processor is not None:
+            skill_inputs = self.data_processor(**skill_inputs)
+        skill_inputs = pipe.skill_model.process_inputs(pipe=pipe, **skill_inputs)
+        return {"skill_inputs": skill_inputs}
+
+
+class GeneralUnit_SkillForward(PipelineUnit):
+    # Only used for training
+    def __init__(self):
+        super().__init__(
+            input_params=("skill_inputs",),
+            output_params=("skill_cache",),
+            onload_model_names=("skill_model",)
+        )
+    
+    def process(self, pipe, skill_inputs):
+        if not hasattr(pipe, "skill_model"):
+            return {}
+        skill_cache = pipe.skill_model.forward(**skill_inputs)
+        return {"skill_cache": skill_cache}
+
+
 class DiffusionTrainingModule(torch.nn.Module):
    def __init__(self):
        super().__init__()
@@ -209,6 +244,16 @@ class DiffusionTrainingModule(torch.nn.Module):
        else:
            lora_target_modules = lora_target_modules.split(",")
        return lora_target_modules
+    
+
+    def load_training_skill_model(self, pipe, path_or_model_id):
+        if path_or_model_id is None:
+            return pipe
+        model_config = self.parse_path_or_model_id(path_or_model_id)
+        pipe.load_training_skill_model(model_config)
+        pipe.units.append(GeneralUnit_SkillProcessInputs(pipe.skill_data_processor))
+        pipe.units.append(GeneralUnit_SkillForward())
+        return pipe


    def switch_pipe_to_training_mode(
--- a/diffsynth/models/flux2_dit.py
+++ b/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:, ...]
--- a/diffsynth/pipelines/flux2_image.py
+++ b/diffsynth/pipelines/flux2_image.py
@@ -93,6 +93,9 @@ class Flux2ImagePipeline(BasePipeline):
        initial_noise: torch.Tensor = None,
        # Steps
        num_inference_steps: int = 30,
+        # KV Cache
+        skill_cache = None,
+        negative_skill_cache = None,
        # Progress bar
        progress_bar_cmd = tqdm,
    ):
@@ -101,9 +104,11 @@ class Flux2ImagePipeline(BasePipeline):
        # Parameters
        inputs_posi = {
            "prompt": prompt,
+            "skill_cache": skill_cache,
        }
        inputs_nega = {
            "negative_prompt": negative_prompt,
+            "skill_cache": negative_skill_cache,
        }
        inputs_shared = {
            "cfg_scale": cfg_scale, "embedded_guidance": embedded_guidance,
@@ -570,6 +575,7 @@ def model_fn_flux2(
    image_ids=None,
    edit_latents=None,
    edit_image_ids=None,
+    skill_cache=None,
    use_gradient_checkpointing=False,
    use_gradient_checkpointing_offload=False,
    **kwargs,
@@ -587,6 +593,7 @@ def model_fn_flux2(
        encoder_hidden_states=prompt_embeds,
        txt_ids=text_ids,
        img_ids=image_ids,
+        kv_cache=skill_cache,
        use_gradient_checkpointing=use_gradient_checkpointing,
        use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
    )
--- a/examples/flux2/model_inference/FLUX.2-klein-base-4B-skills.py
+++ b/examples/flux2/model_inference/FLUX.2-klein-base-4B-skills.py
@@ -0,0 +1,56 @@
+from diffsynth.diffusion.skills import SkillsPipeline
+from diffsynth.pipelines.flux2_image import Flux2ImagePipeline, ModelConfig
+import torch
+from PIL import Image
+
+
+pipe = Flux2ImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        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-base-4B", origin_file_pattern="transformer/*.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/"),
+)
+skills = SkillsPipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="DiffSynth-Studio/F2KB4B-Skills-ControlNet"),
+        ModelConfig(model_id="DiffSynth-Studio/F2KB4B-Skills-Brightness"),
+    ],
+)
+skill_cache = skills(
+    positive_inputs = [
+        {
+            "model_id": 0,
+            "image": Image.open("xxx.jpg"),
+            "prompt": "一位长发少女，四周环绕着魔法粒子",
+        },
+        {
+            "model_id": 1,
+            "scale": 0.6,
+        },
+    ],
+    negative_inputs = [
+        {
+            "model_id": 0,
+            "image": Image.open("xxx.jpg"),
+            "prompt": "一位长发少女，四周环绕着魔法粒子",
+        },
+        {
+            "model_id": 1,
+            "scale": 0.5,
+        },
+    ],
+    pipe=pipe,
+)
+image = pipe(
+    prompt="一位长发少女，四周环绕着魔法粒子",
+    seed=0, rand_device="cuda", num_inference_steps=50, cfg_scale=4,
+    height=1024, width=1024,
+    **skill_cache,
+)
+image.save("image.jpg")
--- a/examples/flux2/model_training/full/FLUX.2-klein-base-4B-skills.sh
+++ b/examples/flux2/model_training/full/FLUX.2-klein-base-4B-skills.sh
@@ -0,0 +1,16 @@
+accelerate launch examples/flux2/model_training/train.py \
+  --dataset_base_path /mnt/nas1/duanzhongjie.dzj/dataset/ImagePulseV2 \
+  --dataset_metadata_path /mnt/nas1/duanzhongjie.dzj/dataset/ImagePulseV2/metadata_example_ti2ti.jsonl \
+  --extra_inputs "skill_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" \
+  --skill_model_id_or_path "models/base" \
+  --tokenizer_path "black-forest-labs/FLUX.2-klein-4B:tokenizer/" \
+  --learning_rate 1e-4 \
+  --num_epochs 999 \
+  --remove_prefix_in_ckpt "pipe.skill_model." \
+  --output_path "./models/train/FLUX.2-klein-base-4B-skills_full" \
+  --trainable_models "skill_model" \
+  --use_gradient_checkpointing \
+  --save_steps 200
--- a/examples/flux2/model_training/scripts/convert_base_model_to_skill_model.py
+++ b/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")
--- a/examples/flux2/model_training/train.py
+++ b/examples/flux2/model_training/train.py
@@ -18,6 +18,7 @@ class Flux2ImageTrainingModule(DiffusionTrainingModule):
        extra_inputs=None,
        fp8_models=None,
        offload_models=None,
+        skill_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_skill_model(self.pipe, skill_model_id_or_path)
        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,
+        skill_model_id_or_path=args.skill_model_id_or_path,
        task=args.task,
        device="cpu" if args.initialize_model_on_cpu else accelerator.device,
    )