Bug fixes. Added IP adapter training for Pixart

toolkit/models/vd_adapter.py

@@ -6,16 +6,103 @@ import torch.nn.functional as F
 import weakref
 from typing import Union, TYPE_CHECKING
 
+from diffusers import Transformer2DModel
 from transformers import T5EncoderModel, CLIPTextModel, CLIPTokenizer, T5Tokenizer, CLIPVisionModelWithProjection
 from toolkit.paths import REPOS_ROOT
 sys.path.append(REPOS_ROOT)
 
-from ipadapter.ip_adapter.attention_processor import AttnProcessor2_0
 
 if TYPE_CHECKING:
     from toolkit.stable_diffusion_model import StableDiffusion
     from toolkit.custom_adapter import CustomAdapter
 
+class AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
 
 class VisionDirectAdapterAttnProcessor(nn.Module):
     r"""
@@ -31,7 +118,7 @@ class VisionDirectAdapterAttnProcessor(nn.Module):
     """
 
     def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, adapter=None,
-                 adapter_hidden_size=None):
+                 adapter_hidden_size=None, has_bias=False, **kwargs):
         super().__init__()
 
         if not hasattr(F, "scaled_dot_product_attention"):
@@ -44,12 +131,13 @@ class VisionDirectAdapterAttnProcessor(nn.Module):
         self.cross_attention_dim = cross_attention_dim
         self.scale = scale
 
-        self.to_k_adapter = nn.Linear(adapter_hidden_size, hidden_size, bias=False)
-        self.to_v_adapter = nn.Linear(adapter_hidden_size, hidden_size, bias=False)
+        self.to_k_adapter = nn.Linear(adapter_hidden_size, hidden_size, bias=has_bias)
+        self.to_v_adapter = nn.Linear(adapter_hidden_size, hidden_size, bias=has_bias)
 
     @property
     def is_active(self):
         return self.adapter_ref().is_active
+        # return False
 
     @property
     def unconditional_embeds(self):
@@ -175,6 +263,7 @@ class VisionDirectAdapter(torch.nn.Module):
             vision_model: Union[CLIPVisionModelWithProjection],
     ):
         super(VisionDirectAdapter, self).__init__()
+        is_pixart = sd.is_pixart
         self.adapter_ref: weakref.ref = weakref.ref(adapter)
         self.sd_ref: weakref.ref = weakref.ref(sd)
         self.vision_model_ref: weakref.ref = weakref.ref(vision_model)
@@ -184,8 +273,22 @@ class VisionDirectAdapter(torch.nn.Module):
         # init adapter modules
         attn_procs = {}
         unet_sd = sd.unet.state_dict()
-        for name in sd.unet.attn_processors.keys():
-            cross_attention_dim = None if name.endswith("attn1.processor") else sd.unet.config['cross_attention_dim']
+
+        attn_processor_keys = []
+        if is_pixart:
+            transformer: Transformer2DModel = sd.unet
+            for i, module in transformer.transformer_blocks.named_children():
+
+                attn_processor_keys.append(f"transformer_blocks.{i}.attn1")
+
+                # cross attention
+                attn_processor_keys.append(f"transformer_blocks.{i}.attn2")
+
+        else:
+            attn_processor_keys = list(sd.unet.attn_processors.keys())
+
+        for name in attn_processor_keys:
+            cross_attention_dim = None if name.endswith("attn1.processor") or name.endswith("attn.1") else sd.unet.config['cross_attention_dim']
             if name.startswith("mid_block"):
                 hidden_size = sd.unet.config['block_out_channels'][-1]
             elif name.startswith("up_blocks"):
@@ -194,6 +297,8 @@ class VisionDirectAdapter(torch.nn.Module):
             elif name.startswith("down_blocks"):
                 block_id = int(name[len("down_blocks.")])
                 hidden_size = sd.unet.config['block_out_channels'][block_id]
+            elif name.startswith("transformer"):
+                hidden_size = sd.unet.config['cross_attention_dim']
             else:
                 # they didnt have this, but would lead to undefined below
                 raise ValueError(f"unknown attn processor name: {name}")
@@ -203,6 +308,12 @@ class VisionDirectAdapter(torch.nn.Module):
                 layer_name = name.split(".processor")[0]
                 to_k_adapter = unet_sd[layer_name + ".to_k.weight"]
                 to_v_adapter = unet_sd[layer_name + ".to_v.weight"]
+                # if is_pixart:
+                #     to_k_bias = unet_sd[layer_name + ".to_k.bias"]
+                #     to_v_bias = unet_sd[layer_name + ".to_v.bias"]
+                # else:
+                #     to_k_bias = None
+                #     to_v_bias = None
 
                 # add zero padding to the adapter
                 if to_k_adapter.shape[1] < self.token_size:
@@ -220,29 +331,65 @@ class VisionDirectAdapter(torch.nn.Module):
                     ],
                         dim=1
                     )
+                    # if is_pixart:
+                    #     to_k_bias = torch.cat([
+                    #         to_k_bias,
+                    #         torch.zeros(self.token_size - to_k_adapter.shape[1]).to(
+                    #             to_k_adapter.device, dtype=to_k_adapter.dtype)
+                    #     ],
+                    #         dim=0
+                    #     )
+                    #     to_v_bias = torch.cat([
+                    #         to_v_bias,
+                    #         torch.zeros(self.token_size - to_v_adapter.shape[1]).to(
+                    #             to_k_adapter.device, dtype=to_k_adapter.dtype)
+                    #     ],
+                    #         dim=0
+                    #     )
                 elif to_k_adapter.shape[1] > self.token_size:
                     to_k_adapter = to_k_adapter[:, :self.token_size]
                     to_v_adapter = to_v_adapter[:, :self.token_size]
+                    # if is_pixart:
+                    #     to_k_bias = to_k_bias[:self.token_size]
+                    #     to_v_bias = to_v_bias[:self.token_size]
                 else:
                     to_k_adapter = to_k_adapter
                     to_v_adapter = to_v_adapter
+                    # if is_pixart:
+                    #     to_k_bias = to_k_bias
+                    #     to_v_bias = to_v_bias
 
-                # todo resize to the TE hidden size
                 weights = {
-                    "to_k_adapter.weight": to_k_adapter,
-                    "to_v_adapter.weight": to_v_adapter,
+                    "to_k_adapter.weight": to_k_adapter * 0.01,
+                    "to_v_adapter.weight": to_v_adapter * 0.01,
                 }
+                # if is_pixart:
+                #     weights["to_k_adapter.bias"] = to_k_bias
+                #     weights["to_v_adapter.bias"] = to_v_bias
 
                 attn_procs[name] = VisionDirectAdapterAttnProcessor(
                     hidden_size=hidden_size,
                     cross_attention_dim=cross_attention_dim,
                     scale=1.0,
                     adapter=self,
-                    adapter_hidden_size=self.token_size
+                    adapter_hidden_size=self.token_size,
+                    has_bias=False,
                 )
                 attn_procs[name].load_state_dict(weights)
-        sd.unet.set_attn_processor(attn_procs)
-        self.adapter_modules = torch.nn.ModuleList(sd.unet.attn_processors.values())
+        if self.sd_ref().is_pixart:
+            # we have to set them ourselves
+            transformer: Transformer2DModel = sd.unet
+            for i, module in transformer.transformer_blocks.named_children():
+                module.attn1.processor = attn_procs[f"transformer_blocks.{i}.attn1"]
+                module.attn2.processor = attn_procs[f"transformer_blocks.{i}.attn2"]
+            self.adapter_modules = torch.nn.ModuleList([
+                transformer.transformer_blocks[i].attn1.processor for i in range(len(transformer.transformer_blocks))
+            ] + [
+                transformer.transformer_blocks[i].attn2.processor for i in range(len(transformer.transformer_blocks))
+            ])
+        else:
+            sd.unet.set_attn_processor(attn_procs)
+            self.adapter_modules = torch.nn.ModuleList(sd.unet.attn_processors.values())
 
     # make a getter to see if is active
     @property