Revert "【Hackathon 5th No.83】PaddleMIX ppdiffusers models模块功能升级同步HF" (P…

…addlePaddle#354)

Reverts PaddlePaddle#322
回滚该PR，等之后修复后再合入，下面是对应的问题。
1. fastdeploy的sd和sdxl 推理造成了错误。无法顺利加载vae的decoder模型，可能当前升级有部分修改，导致这个发生错误。
2.
sd的预训练和lora微调存在性能下降10%，可能怀疑是recompute粒度发生了修改，部分模块之前没有加recompute现在加上了recompute导致的。

ppdiffusers/ppdiffusers/__init__.py

@@ -63,7 +63,6 @@
         LVDMAutoencoderKL,
         LVDMUNet3DModel,
         ModelMixin,
-        MotionAdapter,
         MultiAdapter,
         PriorTransformer,
         T2IAdapter,
@@ -73,7 +72,6 @@
         UNet2DConditionModel,
         UNet2DModel,
         UNet3DConditionModel,
-        UNetMotionModel,
         VQModel,
     )
     from .optimization import (

ppdiffusers/ppdiffusers/models/__init__.py

@@ -37,7 +37,6 @@
     from .unet_2d import UNet2DModel
     from .unet_2d_condition import UNet2DConditionModel
     from .unet_3d_condition import UNet3DConditionModel
-    from .unet_motion_model import MotionAdapter, UNetMotionModel
     from .vq_model import VQModel
 
 try:

ppdiffusers/ppdiffusers/models/activations.py

@@ -13,94 +13,14 @@
 # limitations under the License.
 
 import paddle.nn as nn
-import paddle.nn.functional as F
 
-from ..utils import USE_PEFT_BACKEND
-from .lora import LoRACompatibleLinear
 
-ACTIVATION_FUNCTIONS = {
-    "swish": nn.Silu(),
-    "silu": nn.Silu(),
-    "mish": nn.Mish(),
-    "gelu": nn.GELU(),
-    "relu": nn.ReLU(),
-}
-
-
-def get_activation(act_fn: str) -> nn.Layer:
-    """Helper function to get activation function from string.
-
-    Args:
-        act_fn (str): Name of activation function.
-
-    Returns:
-        nn.Layer: Activation function.
-    """
-
-    act_fn = act_fn.lower()
-    if act_fn in ACTIVATION_FUNCTIONS:
-        return ACTIVATION_FUNCTIONS[act_fn]
+def get_activation(act_fn):
+    if act_fn in ["swish", "silu"]:
+        return nn.Silu()
+    elif act_fn == "mish":
+        return nn.Mish()
+    elif act_fn == "gelu":
+        return nn.GELU()
     else:
         raise ValueError(f"Unsupported activation function: {act_fn}")
-
-
-class GELU(nn.Layer):
-    r"""
-    GELU activation function with tanh approximation support with `approximate="tanh"`.
-
-    Parameters:
-        dim_in (`int`): The number of channels in the input.
-        dim_out (`int`): The number of channels in the output.
-        approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation.
-    """
-
-    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none"):
-        super().__init__()
-        self.proj = LoRACompatibleLinear(dim_in, dim_out)
-        self.approximate = approximate
-        self.approximate_bool = approximate == "tanh"
-
-    def forward(self, hidden_states):
-        hidden_states = self.proj(hidden_states)
-        hidden_states = F.gelu(hidden_states, approximate=self.approximate_bool)
-        return hidden_states
-
-
-class GEGLU(nn.Layer):
-    r"""
-    A [variant](https://arxiv.org/abs/2002.05202) of the gated linear unit activation function.
-
-    Parameters:
-        dim_in (`int`): The number of channels in the input.
-        dim_out (`int`): The number of channels in the output.
-    """
-
-    def __init__(self, dim_in: int, dim_out: int):
-        super().__init__()
-        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
-
-        self.proj = linear_cls(dim_in, dim_out * 2)
-
-    def forward(self, hidden_states, scale: float = 1.0):
-        args = () if USE_PEFT_BACKEND else (scale,)
-        hidden_states, gate = self.proj(hidden_states, *args).chunk(2, axis=-1)
-        return hidden_states * F.gelu(gate)
-
-
-class ApproximateGELU(nn.Layer):
-    r"""
-    The approximate form of the Gaussian Error Linear Unit (GELU). For more details, see section 2 of this
-    [paper](https://arxiv.org/abs/1606.08415).
-
-    Parameters:
-        dim_in (`int`): The number of channels in the input.
-        dim_out (`int`): The number of channels in the output.
-    """
-
-    def __init__(self, dim_in: int, dim_out: int):
-        super().__init__()
-        self.proj = nn.Linear(dim_in, dim_out)
-
-    def forward(self, x):
-        x = self.proj(x)
-        return x * F.sigmoid(1.702 * x)