diffusers 0.30.3__py3-none-any.whl → 0.32.0__py3-none-any.whl

diffusers/models/transformers/transformer_flux.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Black Forest Labs, The HuggingFace Team. All rights reserved.
+# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -13,61 +13,35 @@
 # limitations under the License.
 
 
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, Optional, Tuple, Union
 
+import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...loaders import FluxTransformer2DLoadersMixin, FromOriginalModelMixin, PeftAdapterMixin
 from ...models.attention import FeedForward
-from ...models.attention_processor import Attention, FluxAttnProcessor2_0, FluxSingleAttnProcessor2_0
+from ...models.attention_processor import (
+    Attention,
+    AttentionProcessor,
+    FluxAttnProcessor2_0,
+    FluxAttnProcessor2_0_NPU,
+    FusedFluxAttnProcessor2_0,
+)
 from ...models.modeling_utils import ModelMixin
 from ...models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
 from ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.import_utils import is_torch_npu_available
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings
+from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
 from ..modeling_outputs import Transformer2DModelOutput
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-# YiYi to-do: refactor rope related functions/classes
-def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
-    assert dim % 2 == 0, "The dimension must be even."
-
-    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
-    omega = 1.0 / (theta**scale)
-
-    batch_size, seq_length = pos.shape
-    out = torch.einsum("...n,d->...nd", pos, omega)
-    cos_out = torch.cos(out)
-    sin_out = torch.sin(out)
-
-    stacked_out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
-    out = stacked_out.view(batch_size, -1, dim // 2, 2, 2)
-    return out.float()
-
-
-# YiYi to-do: refactor rope related functions/classes
-class EmbedND(nn.Module):
-    def __init__(self, dim: int, theta: int, axes_dim: List[int]):
-        super().__init__()
-        self.dim = dim
-        self.theta = theta
-        self.axes_dim = axes_dim
-
-    def forward(self, ids: torch.Tensor) -> torch.Tensor:
-        n_axes = ids.shape[-1]
-        emb = torch.cat(
-            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
-            dim=-3,
-        )
-        return emb.unsqueeze(1)
-
-
 @maybe_allow_in_graph
 class FluxSingleTransformerBlock(nn.Module):
     r"""
@@ -92,7 +66,10 @@ class FluxSingleTransformerBlock(nn.Module):
         self.act_mlp = nn.GELU(approximate="tanh")
         self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
 
-        processor = FluxSingleAttnProcessor2_0()
+        if is_torch_npu_available():
+            processor = FluxAttnProcessor2_0_NPU()
+        else:
+            processor = FluxAttnProcessor2_0()
         self.attn = Attention(
             query_dim=dim,
             cross_attention_dim=None,
@@ -111,14 +88,16 @@ class FluxSingleTransformerBlock(nn.Module):
         hidden_states: torch.FloatTensor,
         temb: torch.FloatTensor,
         image_rotary_emb=None,
+        joint_attention_kwargs=None,
     ):
         residual = hidden_states
         norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
         mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
-
+        joint_attention_kwargs = joint_attention_kwargs or {}
         attn_output = self.attn(
             hidden_states=norm_hidden_states,
             image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
         )
 
         hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
@@ -189,20 +168,27 @@ class FluxTransformerBlock(nn.Module):
         encoder_hidden_states: torch.FloatTensor,
         temb: torch.FloatTensor,
         image_rotary_emb=None,
+        joint_attention_kwargs=None,
     ):
         norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
 
         norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
             encoder_hidden_states, emb=temb
         )
-
+        joint_attention_kwargs = joint_attention_kwargs or {}
         # Attention.
-        attn_output, context_attn_output = self.attn(
+        attention_outputs = self.attn(
             hidden_states=norm_hidden_states,
             encoder_hidden_states=norm_encoder_hidden_states,
             image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
         )
 
+        if len(attention_outputs) == 2:
+            attn_output, context_attn_output = attention_outputs
+        elif len(attention_outputs) == 3:
+            attn_output, context_attn_output, ip_attn_output = attention_outputs
+
         # Process attention outputs for the `hidden_states`.
         attn_output = gate_msa.unsqueeze(1) * attn_output
         hidden_states = hidden_states + attn_output
@@ -214,6 +200,8 @@ class FluxTransformerBlock(nn.Module):
         ff_output = gate_mlp.unsqueeze(1) * ff_output
 
         hidden_states = hidden_states + ff_output
+        if len(attention_outputs) == 3:
+            hidden_states = hidden_states + ip_attn_output
 
         # Process attention outputs for the `encoder_hidden_states`.
 
@@ -231,7 +219,9 @@ class FluxTransformerBlock(nn.Module):
         return encoder_hidden_states, hidden_states
 
 
-class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
+class FluxTransformer2DModel(
+    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, FluxTransformer2DLoadersMixin
+):
     """
     The Transformer model introduced in Flux.
 
@@ -250,12 +240,14 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
     """
 
     _supports_gradient_checkpointing = True
+    _no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]
 
     @register_to_config
     def __init__(
         self,
         patch_size: int = 1,
         in_channels: int = 64,
+        out_channels: Optional[int] = None,
         num_layers: int = 19,
         num_single_layers: int = 38,
         attention_head_dim: int = 128,
@@ -263,13 +255,14 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         joint_attention_dim: int = 4096,
         pooled_projection_dim: int = 768,
         guidance_embeds: bool = False,
-        axes_dims_rope: List[int] = [16, 56, 56],
+        axes_dims_rope: Tuple[int] = (16, 56, 56),
     ):
         super().__init__()
-        self.out_channels = in_channels
+        self.out_channels = out_channels or in_channels
         self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
 
-        self.pos_embed = EmbedND(dim=self.inner_dim, theta=10000, axes_dim=axes_dims_rope)
+        self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
+
         text_time_guidance_cls = (
             CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
         )
@@ -278,7 +271,7 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         )
 
         self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.inner_dim)
-        self.x_embedder = torch.nn.Linear(self.config.in_channels, self.inner_dim)
+        self.x_embedder = nn.Linear(self.config.in_channels, self.inner_dim)
 
         self.transformer_blocks = nn.ModuleList(
             [
@@ -307,6 +300,106 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
 
         self.gradient_checkpointing = False
 
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedFluxAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
     def _set_gradient_checkpointing(self, module, value=False):
         if hasattr(module, "gradient_checkpointing"):
             module.gradient_checkpointing = value
@@ -321,7 +414,10 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         txt_ids: torch.Tensor = None,
         guidance: torch.Tensor = None,
         joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_block_samples=None,
+        controlnet_single_block_samples=None,
         return_dict: bool = True,
+        controlnet_blocks_repeat: bool = False,
     ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
         """
         The [`FluxTransformer2DModel`] forward method.
@@ -363,6 +459,7 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                 logger.warning(
                     "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
                 )
+
         hidden_states = self.x_embedder(hidden_states)
 
         timestep = timestep.to(hidden_states.dtype) * 1000
@@ -370,6 +467,7 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
             guidance = guidance.to(hidden_states.dtype) * 1000
         else:
             guidance = None
+
         temb = (
             self.time_text_embed(timestep, pooled_projections)
             if guidance is None
@@ -377,11 +475,29 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         )
         encoder_hidden_states = self.context_embedder(encoder_hidden_states)
 
-        ids = torch.cat((txt_ids, img_ids), dim=1)
+        if txt_ids.ndim == 3:
+            logger.warning(
+                "Passing `txt_ids` 3d torch.Tensor is deprecated."
+                "Please remove the batch dimension and pass it as a 2d torch Tensor"
+            )
+            txt_ids = txt_ids[0]
+        if img_ids.ndim == 3:
+            logger.warning(
+                "Passing `img_ids` 3d torch.Tensor is deprecated."
+                "Please remove the batch dimension and pass it as a 2d torch Tensor"
+            )
+            img_ids = img_ids[0]
+
+        ids = torch.cat((txt_ids, img_ids), dim=0)
         image_rotary_emb = self.pos_embed(ids)
 
+        if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs:
+            ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds")
+            ip_hidden_states = self.encoder_hid_proj(ip_adapter_image_embeds)
+            joint_attention_kwargs.update({"ip_hidden_states": ip_hidden_states})
+
         for index_block, block in enumerate(self.transformer_blocks):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):
@@ -408,12 +524,24 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     encoder_hidden_states=encoder_hidden_states,
                     temb=temb,
                     image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
                 )
 
+            # controlnet residual
+            if controlnet_block_samples is not None:
+                interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
+                interval_control = int(np.ceil(interval_control))
+                # For Xlabs ControlNet.
+                if controlnet_blocks_repeat:
+                    hidden_states = (
+                        hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)]
+                    )
+                else:
+                    hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
         hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
         for index_block, block in enumerate(self.single_transformer_blocks):
-            if self.training and self.gradient_checkpointing:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
                     def custom_forward(*inputs):
@@ -438,6 +566,16 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     hidden_states=hidden_states,
                     temb=temb,
                     image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+
+            # controlnet residual
+            if controlnet_single_block_samples is not None:
+                interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
+                interval_control = int(np.ceil(interval_control))
+                hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
+                    hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+                    + controlnet_single_block_samples[index_block // interval_control]
                 )
 
         hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]