fix: omnigen2

omnigen2/cache_functions/__init__.py

@@ -0,0 +1,3 @@
+from .cache_init import cache_init
+from .cal_type import cal_type
+from .force_scheduler import force_scheduler

omnigen2/cache_functions/cache_init.py

@@ -0,0 +1,38 @@
+# Modified from https://github.com/Shenyi-Z/TaylorSeer/blob/main/TaylorSeers-xDiT/taylorseer_flux/cache_functions/cache_init.py
+
+# Type hinting would cause circular import, self should be `OmniGen2Pipeline`
+def cache_init(self, num_steps: int):
+    '''
+    Initialization for cache.
+    '''
+    cache_dic = {}
+    cache = {}
+    cache_index = {}
+    cache[-1]={}
+    cache_index[-1]={}
+    cache_index['layer_index']={}
+    cache[-1]['layers_stream']={}
+    cache_dic['cache_counter'] = 0
+
+    for j in range(len(self.transformer.layers)):
+        cache[-1]['layers_stream'][j] = {}
+        cache_index[-1][j] = {}
+
+    cache_dic['Delta-DiT'] = False
+    cache_dic['cache_type'] = 'random'
+    cache_dic['cache_index'] = cache_index
+    cache_dic['cache'] = cache
+    cache_dic['fresh_ratio_schedule'] = 'ToCa' 
+    cache_dic['fresh_ratio'] = 0.0
+    cache_dic['fresh_threshold'] = 3
+    cache_dic['soft_fresh_weight'] = 0.0
+    cache_dic['taylor_cache'] = True
+    cache_dic['max_order'] = 4
+    cache_dic['first_enhance'] = 5
+
+    current = {}
+    current['activated_steps'] = [0]
+    current['step'] = 0
+    current['num_steps'] = num_steps
+
+    return cache_dic, current

omnigen2/cache_functions/cal_type.py

@@ -0,0 +1,41 @@
+# Copied from https://github.com/Shenyi-Z/TaylorSeer/blob/main/TaylorSeers-xDiT/taylorseer_flux/cache_functions/cal_type.py
+
+from .force_scheduler import force_scheduler
+
+def cal_type(cache_dic, current):
+    '''
+    Determine calculation type for this step
+    '''
+    if (cache_dic['fresh_ratio'] == 0.0) and (not cache_dic['taylor_cache']):
+        # FORA:Uniform
+        first_step = (current['step'] == 0)
+    else:
+        # ToCa: First enhanced
+        first_step = (current['step'] < cache_dic['first_enhance'])
+
+    if not first_step:
+        fresh_interval = cache_dic['cal_threshold']
+    else:
+        fresh_interval = cache_dic['fresh_threshold']
+
+    if (first_step) or (cache_dic['cache_counter'] == fresh_interval - 1 ):
+        current['type'] = 'full'
+        cache_dic['cache_counter'] = 0
+        current['activated_steps'].append(current['step'])
+        force_scheduler(cache_dic, current)
+    
+    elif (cache_dic['taylor_cache']):
+        cache_dic['cache_counter'] += 1
+        current['type'] = 'Taylor'
+        
+
+    elif (cache_dic['cache_counter'] % 2 == 1): # 0: ToCa-Aggresive-ToCa, 1: Aggresive-ToCa-Aggresive
+        cache_dic['cache_counter'] += 1
+        current['type'] = 'ToCa'
+    # 'cache_noise' 'ToCa' 'FORA'
+    elif cache_dic['Delta-DiT']:
+        cache_dic['cache_counter'] += 1
+        current['type'] = 'Delta-Cache'
+    else:
+        cache_dic['cache_counter'] += 1
+        current['type'] = 'ToCa'

omnigen2/cache_functions/force_scheduler.py

@@ -0,0 +1,19 @@
+# Copied from https://github.com/Shenyi-Z/TaylorSeer/blob/main/TaylorSeers-xDiT/taylorseer_flux/cache_functions/force_scheduler.py
+
+import torch
+
+def force_scheduler(cache_dic, current):
+    if cache_dic['fresh_ratio'] == 0:
+        # FORA
+        linear_step_weight = 0.0
+    else: 
+        # TokenCache
+        linear_step_weight = 0.0
+    step_factor = torch.tensor(1 - linear_step_weight + 2 * linear_step_weight * current['step'] / current['num_steps'])
+    threshold = torch.round(cache_dic['fresh_threshold'] / step_factor)
+
+    # no force constrain for sensitive steps, cause the performance is good enough.
+    # you may have a try.
+    
+    cache_dic['cal_threshold'] = threshold
+    #return threshold

omnigen2/dataset/omnigen2_test_dataset.py

@@ -0,0 +1,153 @@
+from typing import Optional
+
+import os
+import random
+import yaml
+import glob
+from PIL import Image
+
+import torch
+
+from datasets import load_dataset, concatenate_datasets
+
+from ..pipelines.omnigen2.pipeline_omnigen2 import OmniGen2ImageProcessor
+
+class OmniGen2TestDataset(torch.utils.data.Dataset):
+    SYSTEM_PROMPT = "You are a helpful assistant that generates high-quality images based on user instructions."
+
+    def __init__(
+        self,
+        config_path: str,
+        tokenizer,
+        use_chat_template: bool,
+        max_pixels: Optional[int] = None,
+        max_side_length: Optional[int] = None,
+        img_scale_num: int = 16,
+        align_res: bool = True
+    ):
+        
+        self.max_pixels = max_pixels
+        self.max_side_length = max_side_length
+        self.img_scale_num = img_scale_num
+        self.align_res = align_res
+
+        with open(config_path, "r") as f:
+            self.config = yaml.load(f, Loader=yaml.FullLoader)
+
+        self.use_chat_template = use_chat_template
+        self.image_processor = OmniGen2ImageProcessor(vae_scale_factor=img_scale_num, do_resize=True)
+
+        data = self._collect_annotations(self.config)
+
+        self.data = data
+        self.tokenizer = tokenizer
+        
+    def _collect_annotations(self, config):
+        json_datasets = []
+        for data in config['data']:
+            data_path, data_type = data['path'], data.get("type", "default")
+            if os.path.isdir(data_path):
+                jsonl_files = list(glob.glob(os.path.join(data_path, "**/*.jsonl"), recursive=True)) + list(glob.glob(os.path.join(data_path, "**/*.json"), recursive=True))
+                json_dataset = load_dataset('json', data_files=jsonl_files, cache_dir=None)['train']
+            else:
+                data_ext = os.path.splitext(data_path)[-1]
+                if data_ext in [".json", ".jsonl"]:
+                    json_dataset = load_dataset('json', data_files=data_path, cache_dir=None)['train']
+                elif data_ext in [".yml", ".yaml"]:
+                    with open(data_path, "r") as f:
+                        sub_config = yaml.load(f, Loader=yaml.FullLoader)
+                        json_dataset = self._collect_annotations(sub_config)
+                else:
+                    raise NotImplementedError(
+                        f'Unknown data file extension: "{data_ext}". '
+                        f"Currently, .json, .jsonl .yml .yaml are supported. "
+                        "If you are using a supported format, please set the file extension so that the proper parsing "
+                        "routine can be called."
+                    )
+            json_datasets.append(json_dataset)
+        
+        json_dataset = concatenate_datasets(json_datasets)
+        return json_dataset
+    
+    def apply_chat_template(self, instruction, system_prompt):
+        if self.use_chat_template:
+            prompt = [
+                {
+                    "role": "system",
+                    "content": system_prompt,
+                },
+                {"role": "user", "content": instruction},
+            ]
+            instruction = self.tokenizer.apply_chat_template(prompt, tokenize=False, add_generation_prompt=False)
+        return instruction
+    
+    def process_item(self, data_item):
+        assert data_item['instruction'] is not None
+        if 'input_images' in data_item and data_item['input_images'] is not None:
+            input_images_path = data_item['input_images']
+            input_images = []
+
+            for input_image_path in input_images_path:
+                input_image = Image.open(input_image_path).convert("RGB")
+                input_images.append(input_image)
+        else:
+            input_images_path, input_images = None, None
+
+        if input_images is not None and len(input_images) == 1 and self.align_res:
+            target_img_size = (input_images[0].width, input_images[0].height)
+        else:
+            target_img_size = data_item["target_img_size"]
+
+        w, h = target_img_size
+        cur_pixels = w * h
+        ratio = min(1, (self.max_pixels / cur_pixels) ** 0.5)
+
+        target_img_size = (int(w * ratio) // self.img_scale_num * self.img_scale_num, int(h * ratio) // self.img_scale_num * self.img_scale_num)
+
+        data = {
+            'task_type': data_item['task_type'],
+            'instruction': data_item['instruction'],
+            'input_images_path': input_images_path,
+            'input_images': input_images,
+            'target_img_size': target_img_size,
+        }
+        return data
+
+    def __getitem__(self, index):
+        data_item = self.data[index]
+        return self.process_item(data_item)
+        
+    def __len__(self):
+        return len(self.data)
+
+class OmniGen2Collator():
+    def __init__(self, tokenizer, max_token_len):
+        self.tokenizer = tokenizer
+        self.max_token_len = max_token_len
+
+    def __call__(self, batch):
+        task_type = [data['task_type'] for data in batch]
+        instruction = [data['instruction'] for data in batch]
+        input_images_path = [data['input_images_path'] for data in batch]
+        input_images = [data['input_images'] for data in batch]
+        output_image = [data['output_image'] for data in batch]
+        output_image_path = [data['output_image_path'] for data in batch]
+
+        text_inputs = self.tokenizer(
+            instruction,
+            padding="longest",
+            max_length=self.max_token_len,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        data = {
+            "task_type": task_type,
+            "text_ids": text_inputs.input_ids,
+            "text_mask": text_inputs.attention_mask,
+            "input_images": input_images, 
+            "input_images_path": input_images_path,
+            "output_image": output_image,
+            "output_image_path": output_image_path,
+        }
+        return data

omnigen2/dataset/omnigen2_train_dataset.py

@@ -0,0 +1,203 @@
+from typing import Optional, Union, List
+
+import os
+import random
+import yaml
+import glob
+from PIL import Image
+
+import torch
+from torchvision import transforms
+
+from datasets import load_dataset, concatenate_datasets
+
+from ..pipelines.omnigen2.pipeline_omnigen2 import OmniGen2ImageProcessor
+
+class OmniGen2TrainDataset(torch.utils.data.Dataset):
+    SYSTEM_PROMPT = "You are a helpful assistant that generates high-quality images based on user instructions."
+    SYSTEM_PROMPT_DROP = "You are a helpful assistant that generates images."
+
+    def __init__(
+        self,
+        config_path: str,
+        tokenizer,
+        use_chat_template: bool,
+        max_input_pixels: Optional[Union[int, List[int]]] = None,
+        max_output_pixels: Optional[int] = None,
+        max_side_length: Optional[int] = None,
+        img_scale_num: int = 16,
+        prompt_dropout_prob: float = 0.0,
+        ref_img_dropout_prob: float = 0.0,
+    ):
+        self.max_input_pixels = max_input_pixels
+        self.max_output_pixels = max_output_pixels
+
+        self.max_side_length = max_side_length
+        self.img_scale_num = img_scale_num
+        self.prompt_dropout_prob = prompt_dropout_prob
+        self.ref_img_dropout_prob = ref_img_dropout_prob
+
+        with open(config_path, "r") as f:
+            self.config = yaml.load(f, Loader=yaml.FullLoader)
+
+        self.use_chat_template = use_chat_template
+        self.image_processor = OmniGen2ImageProcessor(vae_scale_factor=img_scale_num, do_resize=True)
+
+        data = self._collect_annotations(self.config)
+
+        self.data = data
+        self.tokenizer = tokenizer
+        
+    def _collect_annotations(self, config):
+        total_samples = 0
+        total_ratio = 0
+        json_datasets = []
+        for data in config['data']:
+            data_path, data_type = data['path'], data.get("type", "default")
+            if os.path.isdir(data_path):
+                jsonl_files = list(glob.glob(os.path.join(data_path, "**/*.jsonl"), recursive=True)) + list(glob.glob(os.path.join(data_path, "**/*.json"), recursive=True))
+                json_dataset = load_dataset('json', data_files=jsonl_files, cache_dir=None)['train']
+            else:
+                data_ext = os.path.splitext(data_path)[-1]
+                if data_ext in [".json", ".jsonl"]:
+                    json_dataset = load_dataset('json', data_files=data_path, cache_dir=None)['train']
+                elif data_ext in [".yml", ".yaml"]:
+                    with open(data_path, "r") as f:
+                        sub_config = yaml.load(f, Loader=yaml.FullLoader)
+                        json_dataset = self._collect_annotations(sub_config)
+                else:
+                    raise NotImplementedError(
+                        f'Unknown data file extension: "{data_ext}". '
+                        f"Currently, .json, .jsonl .yml .yaml are supported. "
+                        "If you are using a supported format, please set the file extension so that the proper parsing "
+                        "routine can be called."
+                    )
+            total_ratio += data['ratio']
+            total_samples += len(json_dataset)
+            json_datasets.append(json_dataset)
+        
+        for json_dataset in json_datasets:
+            target_size = int(len(json_dataset) * data['ratio'] / total_ratio) # normalize the ratio
+            if target_size <= len(json_dataset):
+                # Random selection without replacement
+                indices = random.sample(range(len(json_dataset)), target_size)
+            else:
+                # Oversample with replacement
+                indices = random.choices(range(len(json_dataset)), k=target_size)
+            json_dataset = json_dataset.select(indices)
+            
+        json_dataset = concatenate_datasets(json_datasets)
+        return json_dataset
+    
+    def clean_data_item(self, data_item):
+        task_type = data_item['task_type']
+        prefixs = ["The image portrays ", "The image depicts ", "The image captures ", "The image highlights ", "The image shows ", "这张图片展示了"]
+        if "text_to_image" in task_type or "t2i" in task_type:
+            if random.random() < 0.5:
+                for p in prefixs:
+                    if p in data_item['instruction']:
+                        data_item['instruction'] = data_item['instruction'].replace(p, "")
+                        break
+        return data_item
+    
+    def apply_chat_template(self, instruction, system_prompt):
+        if self.use_chat_template:
+            prompt = [
+                {
+                    "role": "system",
+                    "content": system_prompt,
+                },
+                {"role": "user", "content": instruction},
+            ]
+            instruction = self.tokenizer.apply_chat_template(prompt, tokenize=False, add_generation_prompt=False)
+        return instruction
+    
+    def process_item(self, data_item):
+        assert data_item['instruction'] is not None
+        data_item = self.clean_data_item(data_item)
+
+        drop_prompt = random.random() < self.prompt_dropout_prob
+        drop_ref_img = drop_prompt and random.random() < self.ref_img_dropout_prob
+
+        if drop_prompt:
+            instruction = self.apply_chat_template("", self.SYSTEM_PROMPT_DROP)
+        else:
+            instruction = self.apply_chat_template(data_item['instruction'], self.SYSTEM_PROMPT)
+
+        if not drop_ref_img and 'input_images' in data_item and data_item['input_images'] is not None:
+            input_images_path = data_item['input_images']
+            input_images = []
+
+            max_input_pixels = self.max_input_pixels[len(input_images_path) - 1] if isinstance(self.max_input_pixels, list) else self.max_input_pixels
+
+            for input_image_path in input_images_path:
+                input_image = Image.open(input_image_path).convert("RGB")
+                input_image = self.image_processor.preprocess(input_image, max_pixels=max_input_pixels, max_side_length=self.max_side_length)
+                input_images.append(input_image)
+        else:
+            input_images_path, input_images = None, None
+
+        output_image_path = data_item['output_image']
+        output_image = Image.open(output_image_path).convert("RGB")
+        output_image = self.image_processor.preprocess(output_image, max_pixels=self.max_output_pixels, max_side_length=self.max_side_length)
+
+        data = {
+            'task_type': data_item['task_type'],
+            'instruction': instruction,
+            'input_images_path': input_images_path,
+            'input_images': input_images,
+            'output_image': output_image,
+            'output_image_path': output_image_path,
+        }
+        return data
+
+    def __getitem__(self, index):
+        max_retries = 12
+
+        current_index = index
+        for attempt in range(max_retries):
+            try:
+                data_item = self.data[current_index]
+                return self.process_item(data_item)
+            except Exception as e:
+                if attempt == max_retries - 1:
+                    raise e
+                else:
+                    # Try a different index for the next attempt
+                    current_index = random.randint(0, len(self.data) - 1)
+                    continue
+        
+    def __len__(self):
+        return len(self.data)
+
+class OmniGen2Collator():
+    def __init__(self, tokenizer, max_token_len):
+        self.tokenizer = tokenizer
+        self.max_token_len = max_token_len
+
+    def __call__(self, batch):
+        task_type = [data['task_type'] for data in batch]
+        instruction = [data['instruction'] for data in batch]
+        input_images_path = [data['input_images_path'] for data in batch]
+        input_images = [data['input_images'] for data in batch]
+        output_image = [data['output_image'] for data in batch]
+        output_image_path = [data['output_image_path'] for data in batch]
+
+        text_inputs = self.tokenizer(
+            instruction,
+            padding="longest",
+            max_length=self.max_token_len,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        data = {
+            "task_type": task_type,
+            "text_ids": text_inputs.input_ids,
+            "text_mask": text_inputs.attention_mask,
+            "input_images": input_images, 
+            "input_images_path": input_images_path,
+            "output_image": output_image,
+            "output_image_path": output_image_path,
+        }
+        return data

omnigen2/models/attention_processor.py

@@ -0,0 +1,357 @@
+"""
+OmniGen2 Attention Processor Module
+
+Copyright 2025 BAAI, The OmniGen2 Team and The HuggingFace 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.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+import warnings
+import math
+from typing import Optional, Tuple, Dict, Any
+
+import torch
+import torch.nn.functional as F
+from einops import repeat
+
+from ..utils.import_utils import is_flash_attn_available
+
+if is_flash_attn_available():
+    from flash_attn import flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input
+else:
+    warnings.warn("Cannot import flash_attn, install flash_attn to use Flash2Varlen attention for better performance")
+
+
+from diffusers.models.attention_processor import Attention
+from .embeddings import apply_rotary_emb
+
+
+class OmniGen2AttnProcessorFlash2Varlen:
+    """
+    Processor for implementing scaled dot-product attention with flash attention and variable length sequences.
+    
+    This processor implements:
+    - Flash attention with variable length sequences
+    - Rotary position embeddings (RoPE)
+    - Query-Key normalization
+    - Proportional attention scaling
+    
+    Args:
+        None
+    """
+
+    def __init__(self) -> None:
+        """Initialize the attention processor."""
+        if not is_flash_attn_available():
+            raise ImportError(
+                "OmniGen2AttnProcessorFlash2Varlen requires flash_attn. "
+                "Please install flash_attn."
+            )
+
+    def _upad_input(
+        self,
+        query_layer: torch.Tensor,
+        key_layer: torch.Tensor,
+        value_layer: torch.Tensor,
+        attention_mask: torch.Tensor,
+        query_length: int,
+        num_heads: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, Tuple[torch.Tensor, torch.Tensor], Tuple[int, int]]:
+        """
+        Unpad the input tensors for flash attention.
+
+        Args:
+            query_layer: Query tensor of shape (batch_size, seq_len, num_heads, head_dim)
+            key_layer: Key tensor of shape (batch_size, seq_len, num_kv_heads, head_dim)
+            value_layer: Value tensor of shape (batch_size, seq_len, num_kv_heads, head_dim)
+            attention_mask: Attention mask tensor of shape (batch_size, seq_len)
+            query_length: Length of the query sequence
+            num_heads: Number of attention heads
+
+        Returns:
+            Tuple containing:
+                - Unpadded query tensor
+                - Unpadded key tensor
+                - Unpadded value tensor
+                - Query indices
+                - Tuple of cumulative sequence lengths for query and key
+                - Tuple of maximum sequence lengths for query and key
+        """
+        def _get_unpad_data(attention_mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, int]:
+            """Helper function to get unpadding data from attention mask."""
+            seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+            indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+            max_seqlen_in_batch = seqlens_in_batch.max().item()
+            cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+            return indices, cu_seqlens, max_seqlen_in_batch
+
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        # Unpad key and value layers
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
+            indices_k,
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim),
+            indices_k,
+        )
+
+        # Handle different query length cases
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim),
+                indices_k,
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+            
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        base_sequence_length: Optional[int] = None,
+    ) -> torch.Tensor:
+        """
+        Process attention computation with flash attention.
+
+        Args:
+            attn: Attention module
+            hidden_states: Hidden states tensor of shape (batch_size, seq_len, hidden_dim)
+            encoder_hidden_states: Encoder hidden states tensor
+            attention_mask: Optional attention mask tensor
+            image_rotary_emb: Optional rotary embeddings for image tokens
+            base_sequence_length: Optional base sequence length for proportional attention
+
+        Returns:
+            torch.Tensor: Processed hidden states after attention computation
+        """
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        # Get Query-Key-Value Pair
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query_dim = query.shape[-1]
+        inner_dim = key.shape[-1]
+        head_dim = query_dim // attn.heads
+        dtype = query.dtype
+
+        # Get key-value heads
+        kv_heads = inner_dim // head_dim
+
+        # Reshape tensors for attention computation
+        query = query.view(batch_size, -1, attn.heads, head_dim)
+        key = key.view(batch_size, -1, kv_heads, head_dim)
+        value = value.view(batch_size, -1, kv_heads, head_dim)
+
+        # Apply Query-Key normalization
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply Rotary Position Embeddings
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb, use_real=False)
+            key = apply_rotary_emb(key, image_rotary_emb, use_real=False)
+
+        query, key = query.to(dtype), key.to(dtype)
+
+        # Calculate attention scale
+        if base_sequence_length is not None:
+            softmax_scale = math.sqrt(math.log(sequence_length, base_sequence_length)) * attn.scale
+        else:
+            softmax_scale = attn.scale
+
+        # Unpad input for flash attention
+        (
+            query_states,
+            key_states,
+            value_states,
+            indices_q,
+            cu_seq_lens,
+            max_seq_lens,
+        ) = self._upad_input(query, key, value, attention_mask, sequence_length, attn.heads)
+
+        cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+        max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+        # Handle different number of heads
+        if kv_heads < attn.heads:
+            key_states = repeat(key_states, "l h c -> l (h k) c", k=attn.heads // kv_heads)
+            value_states = repeat(value_states, "l h c -> l (h k) c", k=attn.heads // kv_heads)
+
+        # Apply flash attention
+        attn_output_unpad = flash_attn_varlen_func(
+            query_states,
+            key_states,
+            value_states,
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+            max_seqlen_q=max_seqlen_in_batch_q,
+            max_seqlen_k=max_seqlen_in_batch_k,
+            dropout_p=0.0,
+            causal=False,
+            softmax_scale=softmax_scale,
+        )
+
+        # Pad output and apply final transformations
+        hidden_states = pad_input(attn_output_unpad, indices_q, batch_size, sequence_length)
+        hidden_states = hidden_states.flatten(-2)
+        hidden_states = hidden_states.type_as(query)
+
+        # Apply output projection
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        
+        return hidden_states
+
+
+class OmniGen2AttnProcessor:
+    """
+    Processor for implementing scaled dot-product attention with flash attention and variable length sequences.
+    
+    This processor is optimized for PyTorch 2.0 and implements:
+    - Flash attention with variable length sequences
+    - Rotary position embeddings (RoPE)
+    - Query-Key normalization
+    - Proportional attention scaling
+    
+    Args:
+        None
+        
+    Raises:
+        ImportError: If PyTorch version is less than 2.0
+    """
+
+    def __init__(self) -> None:
+        """Initialize the attention processor."""
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "OmniGen2AttnProcessorFlash2Varlen requires PyTorch 2.0. "
+                "Please upgrade PyTorch to version 2.0 or later."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+        base_sequence_length: Optional[int] = None,
+    ) -> torch.Tensor:
+        """
+        Process attention computation with flash attention.
+
+        Args:
+            attn: Attention module
+            hidden_states: Hidden states tensor of shape (batch_size, seq_len, hidden_dim)
+            encoder_hidden_states: Encoder hidden states tensor
+            attention_mask: Optional attention mask tensor
+            image_rotary_emb: Optional rotary embeddings for image tokens
+            base_sequence_length: Optional base sequence length for proportional attention
+
+        Returns:
+            torch.Tensor: Processed hidden states after attention computation
+        """
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        # Get Query-Key-Value Pair
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query_dim = query.shape[-1]
+        inner_dim = key.shape[-1]
+        head_dim = query_dim // attn.heads
+        dtype = query.dtype
+
+        # Get key-value heads
+        kv_heads = inner_dim // head_dim
+
+        # Reshape tensors for attention computation
+        query = query.view(batch_size, -1, attn.heads, head_dim)
+        key = key.view(batch_size, -1, kv_heads, head_dim)
+        value = value.view(batch_size, -1, kv_heads, head_dim)
+
+        # Apply Query-Key normalization
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply Rotary Position Embeddings
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb, use_real=False)
+            key = apply_rotary_emb(key, image_rotary_emb, use_real=False)
+
+        query, key = query.to(dtype), key.to(dtype)
+
+        # Calculate attention scale
+        if base_sequence_length is not None:
+            softmax_scale = math.sqrt(math.log(sequence_length, base_sequence_length)) * attn.scale
+        else:
+            softmax_scale = attn.scale
+
+        # scaled_dot_product_attention expects attention_mask shape to be
+        # (batch, heads, source_length, target_length)
+        if attention_mask is not None:
+            attention_mask = attention_mask.bool().view(batch_size, 1, 1, -1)
+
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+        
+        # explicitly repeat key and value to match query length, otherwise using enable_gqa=True results in MATH backend of sdpa in our test of pytorch2.6
+        key = key.repeat_interleave(query.size(-3) // key.size(-3), -3)
+        value = value.repeat_interleave(query.size(-3) // value.size(-3), -3)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, scale=softmax_scale
+        )
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.type_as(query)
+
+        # Apply output projection
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        
+        return hidden_states

omnigen2/models/embeddings.py

@@ -0,0 +1,126 @@
+# Copyright 2024 The HuggingFace 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+
+from diffusers.models.activations import get_activation
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        time_embed_dim: int,
+        act_fn: str = "silu",
+        out_dim: int = None,
+        post_act_fn: Optional[str] = None,
+        cond_proj_dim=None,
+        sample_proj_bias=True,
+    ):
+        super().__init__()
+
+        self.linear_1 = nn.Linear(in_channels, time_embed_dim, sample_proj_bias)
+
+        if cond_proj_dim is not None:
+            self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False)
+        else:
+            self.cond_proj = None
+
+        self.act = get_activation(act_fn)
+
+        if out_dim is not None:
+            time_embed_dim_out = out_dim
+        else:
+            time_embed_dim_out = time_embed_dim
+        self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias)
+
+        if post_act_fn is None:
+            self.post_act = None
+        else:
+            self.post_act = get_activation(post_act_fn)
+
+        self.initialize_weights()
+        
+    def initialize_weights(self):
+        nn.init.normal_(self.linear_1.weight, std=0.02)
+        nn.init.zeros_(self.linear_1.bias)
+        nn.init.normal_(self.linear_2.weight, std=0.02)
+        nn.init.zeros_(self.linear_2.bias)
+        
+    def forward(self, sample, condition=None):
+        if condition is not None:
+            sample = sample + self.cond_proj(condition)
+        sample = self.linear_1(sample)
+
+        if self.act is not None:
+            sample = self.act(sample)
+
+        sample = self.linear_2(sample)
+
+        if self.post_act is not None:
+            sample = self.post_act(sample)
+        return sample
+
+
+def apply_rotary_emb(
+    x: torch.Tensor,
+    freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+    use_real: bool = True,
+    use_real_unbind_dim: int = -1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
+    to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
+    reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
+    tensors contain rotary embeddings and are returned as real tensors.
+
+    Args:
+        x (`torch.Tensor`):
+            Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
+        freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+    """
+    if use_real:
+        cos, sin = freqs_cis  # [S, D]
+        cos = cos[None, None]
+        sin = sin[None, None]
+        cos, sin = cos.to(x.device), sin.to(x.device)
+
+        if use_real_unbind_dim == -1:
+            # Used for flux, cogvideox, hunyuan-dit
+            x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+            x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+        elif use_real_unbind_dim == -2:
+            # Used for Stable Audio, OmniGen and CogView4
+            x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2)  # [B, S, H, D//2]
+            x_rotated = torch.cat([-x_imag, x_real], dim=-1)
+        else:
+            raise ValueError(f"`use_real_unbind_dim={use_real_unbind_dim}` but should be -1 or -2.")
+
+        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+        return out
+    else:
+        # used for lumina
+        # x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
+        x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], x.shape[-1] // 2, 2))
+        freqs_cis = freqs_cis.unsqueeze(2)
+        x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3)
+
+        return x_out.type_as(x)

omnigen2/models/transformers/__init__.py

@@ -0,0 +1,3 @@
+from .transformer_omnigen2 import OmniGen2Transformer2DModel
+
+__all__ = ["OmniGen2Transformer2DModel"]

omnigen2/models/transformers/block_lumina2.py

@@ -0,0 +1,218 @@
+
+# Copyright 2024 Alpha-VLLM Authors and The HuggingFace 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+
+from diffusers.models.embeddings import Timesteps
+from ..embeddings import TimestepEmbedding
+
+from ...utils.import_utils import is_flash_attn_available, is_triton_available
+
+if is_triton_available():
+    from ...ops.triton.layer_norm import RMSNorm
+else:
+    from torch.nn import RMSNorm
+    warnings.warn("Cannot import triton, install triton to use fused RMSNorm for better performance")
+    
+if is_flash_attn_available():
+    from flash_attn.ops.activations import swiglu
+else:
+    from .components import swiglu
+    warnings.warn("Cannot import flash_attn, install flash_attn to use fused SwiGLU for better performance")
+
+# try:
+#     from flash_attn.ops.activations import swiglu as fused_swiglu
+#     FUSEDSWIGLU_AVALIBLE = True
+# except ImportError:
+    
+#     FUSEDSWIGLU_AVALIBLE = False
+#     warnings.warn("Cannot import apex RMSNorm, switch to vanilla implementation")
+        
+class LuminaRMSNormZero(nn.Module):
+    """
+    Norm layer adaptive RMS normalization zero.
+
+    Parameters:
+        embedding_dim (`int`): The size of each embedding vector.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        norm_eps: float,
+        norm_elementwise_affine: bool,
+    ):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(
+            min(embedding_dim, 1024),
+            4 * embedding_dim,
+            bias=True,
+        )
+        
+        self.norm = RMSNorm(embedding_dim, eps=norm_eps)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        emb: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        scale_msa, gate_msa, scale_mlp, gate_mlp = emb.chunk(4, dim=1)
+        x = self.norm(x) * (1 + scale_msa[:, None])
+        return x, gate_msa, scale_mlp, gate_mlp
+    
+
+class LuminaLayerNormContinuous(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        conditioning_embedding_dim: int,
+        # NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters
+        # because the output is immediately scaled and shifted by the projected conditioning embeddings.
+        # Note that AdaLayerNorm does not let the norm layer have scale and shift parameters.
+        # However, this is how it was implemented in the original code, and it's rather likely you should
+        # set `elementwise_affine` to False.
+        elementwise_affine=True,
+        eps=1e-5,
+        bias=True,
+        norm_type="layer_norm",
+        out_dim: Optional[int] = None,
+    ):
+        super().__init__()
+
+        # AdaLN
+        self.silu = nn.SiLU()
+        self.linear_1 = nn.Linear(conditioning_embedding_dim, embedding_dim, bias=bias)
+
+        if norm_type == "layer_norm":
+            self.norm = nn.LayerNorm(embedding_dim, eps, elementwise_affine, bias)
+        elif norm_type == "rms_norm":
+            self.norm = RMSNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
+        else:
+            raise ValueError(f"unknown norm_type {norm_type}")
+
+        self.linear_2 = None
+        if out_dim is not None:
+            self.linear_2 = nn.Linear(embedding_dim, out_dim, bias=bias)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        conditioning_embedding: torch.Tensor,
+    ) -> torch.Tensor:
+        # convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32 (needed for hunyuanDiT)
+        emb = self.linear_1(self.silu(conditioning_embedding).to(x.dtype))
+        scale = emb
+        x = self.norm(x) * (1 + scale)[:, None, :]
+
+        if self.linear_2 is not None:
+            x = self.linear_2(x)
+
+        return x
+    
+
+class LuminaFeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        hidden_size (`int`):
+            The dimensionality of the hidden layers in the model. This parameter determines the width of the model's
+            hidden representations.
+        intermediate_size (`int`): The intermediate dimension of the feedforward layer.
+        multiple_of (`int`, *optional*): Value to ensure hidden dimension is a multiple
+            of this value.
+        ffn_dim_multiplier (float, *optional*): Custom multiplier for hidden
+            dimension. Defaults to None.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        inner_dim: int,
+        multiple_of: Optional[int] = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+    ):
+        super().__init__()
+        self.swiglu = swiglu
+        
+        # custom hidden_size factor multiplier
+        if ffn_dim_multiplier is not None:
+            inner_dim = int(ffn_dim_multiplier * inner_dim)
+        inner_dim = multiple_of * ((inner_dim + multiple_of - 1) // multiple_of)
+
+        self.linear_1 = nn.Linear(
+            dim,
+            inner_dim,
+            bias=False,
+        )
+        self.linear_2 = nn.Linear(
+            inner_dim,
+            dim,
+            bias=False,
+        )
+        self.linear_3 = nn.Linear(
+            dim,
+            inner_dim,
+            bias=False,
+        )
+
+    def forward(self, x):
+        h1, h2 = self.linear_1(x), self.linear_3(x)
+        return self.linear_2(self.swiglu(h1, h2))
+
+
+class Lumina2CombinedTimestepCaptionEmbedding(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int = 4096,
+        text_feat_dim: int = 2048,
+        frequency_embedding_size: int = 256,
+        norm_eps: float = 1e-5,
+        timestep_scale: float = 1.0,
+    ) -> None:
+        super().__init__()
+
+        self.time_proj = Timesteps(
+            num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0.0, scale=timestep_scale
+        )
+
+        self.timestep_embedder = TimestepEmbedding(
+            in_channels=frequency_embedding_size, time_embed_dim=min(hidden_size, 1024)
+        )
+
+        self.caption_embedder = nn.Sequential(
+            RMSNorm(text_feat_dim, eps=norm_eps),
+            nn.Linear(text_feat_dim, hidden_size, bias=True),
+        )
+        
+        self._initialize_weights()
+
+    def _initialize_weights(self):
+        nn.init.trunc_normal_(self.caption_embedder[1].weight, std=0.02)
+        nn.init.zeros_(self.caption_embedder[1].bias)
+
+    def forward(
+        self, timestep: torch.Tensor, text_hidden_states: torch.Tensor, dtype: torch.dtype
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        timestep_proj = self.time_proj(timestep).to(dtype=dtype)
+        time_embed = self.timestep_embedder(timestep_proj)
+        caption_embed = self.caption_embedder(text_hidden_states)
+        return time_embed, caption_embed

omnigen2/models/transformers/components.py

@@ -0,0 +1,4 @@
+import torch.nn.functional as F
+
+def swiglu(x, y):
+    return F.silu(x.float(), inplace=False).to(x.dtype) * y

omnigen2/models/transformers/repo.py

@@ -0,0 +1,129 @@
+from typing import List, Tuple
+
+import torch
+import torch.nn as nn
+
+from einops import repeat
+from diffusers.models.embeddings import get_1d_rotary_pos_embed
+
+class OmniGen2RotaryPosEmbed(nn.Module):
+    def __init__(self, theta: int,
+                 axes_dim: Tuple[int, int, int],
+                 axes_lens: Tuple[int, int, int] = (300, 512, 512),
+                 patch_size: int = 2):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+        self.axes_lens = axes_lens
+        self.patch_size = patch_size
+
+    @staticmethod
+    def get_freqs_cis(axes_dim: Tuple[int, int, int],
+                      axes_lens: Tuple[int, int, int],
+                      theta: int) -> List[torch.Tensor]:
+        freqs_cis = []
+        freqs_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64
+        for i, (d, e) in enumerate(zip(axes_dim, axes_lens)):
+            emb = get_1d_rotary_pos_embed(d, e, theta=theta, freqs_dtype=freqs_dtype)
+            freqs_cis.append(emb)
+        return freqs_cis
+
+    def _get_freqs_cis(self, freqs_cis, ids: torch.Tensor) -> torch.Tensor:
+        device = ids.device
+        if ids.device.type == "mps":
+            ids = ids.to("cpu")
+
+        result = []
+        for i in range(len(self.axes_dim)):
+            freqs = freqs_cis[i].to(ids.device)
+            index = ids[:, :, i : i + 1].repeat(1, 1, freqs.shape[-1]).to(torch.int64)
+            result.append(torch.gather(freqs.unsqueeze(0).repeat(index.shape[0], 1, 1), dim=1, index=index))
+        return torch.cat(result, dim=-1).to(device)
+
+    def forward(
+        self,
+        freqs_cis,
+        attention_mask,
+        l_effective_ref_img_len,
+        l_effective_img_len,
+        ref_img_sizes,
+        img_sizes,
+        device
+    ):
+        batch_size = len(attention_mask)
+        p = self.patch_size
+
+        encoder_seq_len = attention_mask.shape[1]
+        l_effective_cap_len = attention_mask.sum(dim=1).tolist()
+
+        seq_lengths = [cap_len + sum(ref_img_len) + img_len for cap_len, ref_img_len, img_len in zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len)]
+
+        max_seq_len = max(seq_lengths)
+        max_ref_img_len = max([sum(ref_img_len) for ref_img_len in l_effective_ref_img_len])
+        max_img_len = max(l_effective_img_len)
+
+        # Create position IDs
+        position_ids = torch.zeros(batch_size, max_seq_len, 3, dtype=torch.int32, device=device)
+
+        for i, (cap_seq_len, seq_len) in enumerate(zip(l_effective_cap_len, seq_lengths)):
+            # add text position ids
+            position_ids[i, :cap_seq_len] = repeat(torch.arange(cap_seq_len, dtype=torch.int32, device=device), "l -> l 3")
+
+            pe_shift = cap_seq_len
+            pe_shift_len = cap_seq_len
+
+            if ref_img_sizes[i] is not None:
+                for ref_img_size, ref_img_len in zip(ref_img_sizes[i], l_effective_ref_img_len[i]):
+                    H, W = ref_img_size
+                    ref_H_tokens, ref_W_tokens = H // p, W // p
+                    assert ref_H_tokens * ref_W_tokens == ref_img_len
+                    # add image position ids
+
+                    row_ids = repeat(torch.arange(ref_H_tokens, dtype=torch.int32, device=device), "h -> h w", w=ref_W_tokens).flatten()
+                    col_ids = repeat(torch.arange(ref_W_tokens, dtype=torch.int32, device=device), "w -> h w", h=ref_H_tokens).flatten()
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 0] = pe_shift
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 1] = row_ids
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 2] = col_ids
+
+                    pe_shift += max(ref_H_tokens, ref_W_tokens)
+                    pe_shift_len += ref_img_len
+
+            H, W = img_sizes[i]
+            H_tokens, W_tokens = H // p, W // p
+            assert H_tokens * W_tokens == l_effective_img_len[i]
+
+            row_ids = repeat(torch.arange(H_tokens, dtype=torch.int32, device=device), "h -> h w", w=W_tokens).flatten()
+            col_ids = repeat(torch.arange(W_tokens, dtype=torch.int32, device=device), "w -> h w", h=H_tokens).flatten()
+
+            assert pe_shift_len + l_effective_img_len[i] == seq_len
+            position_ids[i, pe_shift_len: seq_len, 0] = pe_shift
+            position_ids[i, pe_shift_len: seq_len, 1] = row_ids
+            position_ids[i, pe_shift_len: seq_len, 2] = col_ids
+
+        # Get combined rotary embeddings
+        freqs_cis = self._get_freqs_cis(freqs_cis, position_ids)
+        
+        # create separate rotary embeddings for captions and images
+        cap_freqs_cis = torch.zeros(
+            batch_size, encoder_seq_len, freqs_cis.shape[-1], device=device, dtype=freqs_cis.dtype
+        )
+        ref_img_freqs_cis = torch.zeros(
+            batch_size, max_ref_img_len, freqs_cis.shape[-1], device=device, dtype=freqs_cis.dtype
+        )
+        img_freqs_cis = torch.zeros(
+            batch_size, max_img_len, freqs_cis.shape[-1], device=device, dtype=freqs_cis.dtype
+        )
+
+        for i, (cap_seq_len, ref_img_len, img_len, seq_len) in enumerate(zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len, seq_lengths)):
+            cap_freqs_cis[i, :cap_seq_len] = freqs_cis[i, :cap_seq_len]
+            ref_img_freqs_cis[i, :sum(ref_img_len)] = freqs_cis[i, cap_seq_len:cap_seq_len + sum(ref_img_len)]
+            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_seq_len + sum(ref_img_len):cap_seq_len + sum(ref_img_len) + img_len]
+
+        return (
+            cap_freqs_cis,
+            ref_img_freqs_cis,
+            img_freqs_cis,
+            freqs_cis,
+            l_effective_cap_len,
+            seq_lengths,
+        )

omnigen2/models/transformers/transformer_omnigen2.py

@@ -0,0 +1,716 @@
+import warnings
+import itertools
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+
+from einops import rearrange
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import PeftAdapterMixin
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.attention_processor import Attention
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+
+from ..attention_processor import OmniGen2AttnProcessorFlash2Varlen, OmniGen2AttnProcessor
+from .repo import OmniGen2RotaryPosEmbed
+from .block_lumina2 import LuminaLayerNormContinuous, LuminaRMSNormZero, LuminaFeedForward, Lumina2CombinedTimestepCaptionEmbedding
+
+from ...utils.import_utils import is_triton_available, is_flash_attn_available
+from ...utils.teacache_util import TeaCacheParams
+
+if is_triton_available():
+    from ...ops.triton.layer_norm import RMSNorm
+else:
+    from torch.nn import RMSNorm
+
+from ...taylorseer_utils import derivative_approximation, taylor_formula, taylor_cache_init
+from ...cache_functions import cache_init, cal_type
+
+logger = logging.get_logger(__name__)
+
+class OmniGen2TransformerBlock(nn.Module):
+    """
+    Transformer block for OmniGen2 model.
+    
+    This block implements a transformer layer with:
+    - Multi-head attention with flash attention
+    - Feed-forward network with SwiGLU activation
+    - RMS normalization
+    - Optional modulation for conditional generation
+    
+    Args:
+        dim: Dimension of the input and output tensors
+        num_attention_heads: Number of attention heads
+        num_kv_heads: Number of key-value heads
+        multiple_of: Multiple of which the hidden dimension should be
+        ffn_dim_multiplier: Multiplier for the feed-forward network dimension
+        norm_eps: Epsilon value for normalization layers
+        modulation: Whether to use modulation for conditional generation
+        use_fused_rms_norm: Whether to use fused RMS normalization
+        use_fused_swiglu: Whether to use fused SwiGLU activation
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        num_kv_heads: int,
+        multiple_of: int,
+        ffn_dim_multiplier: float,
+        norm_eps: float,
+        modulation: bool = True,
+    ) -> None:
+        """Initialize the transformer block."""
+        super().__init__()
+        self.head_dim = dim // num_attention_heads
+        self.modulation = modulation
+
+        try:
+            processor = OmniGen2AttnProcessorFlash2Varlen()
+        except ImportError:
+            processor = OmniGen2AttnProcessor()
+
+        # Initialize attention layer
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,
+            dim_head=dim // num_attention_heads,
+            qk_norm="rms_norm",
+            heads=num_attention_heads,
+            kv_heads=num_kv_heads,
+            eps=1e-5,
+            bias=False,
+            out_bias=False,
+            processor=processor,
+        )
+
+        # Initialize feed-forward network
+        self.feed_forward = LuminaFeedForward(
+            dim=dim,
+            inner_dim=4 * dim,
+            multiple_of=multiple_of,
+            ffn_dim_multiplier=ffn_dim_multiplier
+        )
+
+        # Initialize normalization layers
+        if modulation:
+            self.norm1 = LuminaRMSNormZero(
+                embedding_dim=dim,
+                norm_eps=norm_eps,
+                norm_elementwise_affine=True
+            )
+        else:
+            self.norm1 = RMSNorm(dim, eps=norm_eps)
+
+        self.ffn_norm1 = RMSNorm(dim, eps=norm_eps)
+        self.norm2 = RMSNorm(dim, eps=norm_eps)
+        self.ffn_norm2 = RMSNorm(dim, eps=norm_eps)
+
+        self.initialize_weights()
+
+    def initialize_weights(self) -> None:
+        """
+        Initialize the weights of the transformer block.
+        
+        Uses Xavier uniform initialization for linear layers and zero initialization for biases.
+        """
+        nn.init.xavier_uniform_(self.attn.to_q.weight)
+        nn.init.xavier_uniform_(self.attn.to_k.weight)
+        nn.init.xavier_uniform_(self.attn.to_v.weight)
+        nn.init.xavier_uniform_(self.attn.to_out[0].weight)
+
+        nn.init.xavier_uniform_(self.feed_forward.linear_1.weight)
+        nn.init.xavier_uniform_(self.feed_forward.linear_2.weight)
+        nn.init.xavier_uniform_(self.feed_forward.linear_3.weight)
+        
+        if self.modulation:
+            nn.init.zeros_(self.norm1.linear.weight)
+            nn.init.zeros_(self.norm1.linear.bias)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        image_rotary_emb: torch.Tensor,
+        temb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Forward pass of the transformer block.
+
+        Args:
+            hidden_states: Input hidden states tensor
+            attention_mask: Attention mask tensor
+            image_rotary_emb: Rotary embeddings for image tokens
+            temb: Optional timestep embedding tensor
+
+        Returns:
+            torch.Tensor: Output hidden states after transformer block processing
+        """
+        enable_taylorseer = getattr(self, 'enable_taylorseer', False)
+        if enable_taylorseer:
+            if self.modulation:
+                if temb is None:
+                    raise ValueError("temb must be provided when modulation is enabled")
+                    
+                if self.current['type'] == 'full':
+                    self.current['module'] = 'total'
+                    taylor_cache_init(cache_dic=self.cache_dic, current=self.current)
+
+                    norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
+                    attn_output = self.attn(
+                        hidden_states=norm_hidden_states,
+                        encoder_hidden_states=norm_hidden_states,
+                        attention_mask=attention_mask,
+                        image_rotary_emb=image_rotary_emb,
+                    )
+                    hidden_states = hidden_states + gate_msa.unsqueeze(1).tanh() * self.norm2(attn_output)
+                    mlp_output = self.feed_forward(self.ffn_norm1(hidden_states) * (1 + scale_mlp.unsqueeze(1)))
+                    hidden_states = hidden_states + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(mlp_output)
+
+                    derivative_approximation(cache_dic=self.cache_dic, current=self.current, feature=hidden_states)
+
+                elif self.current['type'] == 'Taylor': 
+                    self.current['module'] = 'total'
+                    hidden_states = taylor_formula(cache_dic=self.cache_dic, current=self.current)
+            else:
+                norm_hidden_states = self.norm1(hidden_states)
+                attn_output = self.attn(
+                    hidden_states=norm_hidden_states,
+                    encoder_hidden_states=norm_hidden_states,
+                    attention_mask=attention_mask,
+                    image_rotary_emb=image_rotary_emb,
+                )
+                hidden_states = hidden_states + self.norm2(attn_output)
+                mlp_output = self.feed_forward(self.ffn_norm1(hidden_states))
+                hidden_states = hidden_states + self.ffn_norm2(mlp_output)
+        else:
+            if self.modulation:
+                if temb is None:
+                    raise ValueError("temb must be provided when modulation is enabled")
+                    
+                norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
+                attn_output = self.attn(
+                    hidden_states=norm_hidden_states,
+                    encoder_hidden_states=norm_hidden_states,
+                    attention_mask=attention_mask,
+                    image_rotary_emb=image_rotary_emb,
+                )
+                hidden_states = hidden_states + gate_msa.unsqueeze(1).tanh() * self.norm2(attn_output)
+                mlp_output = self.feed_forward(self.ffn_norm1(hidden_states) * (1 + scale_mlp.unsqueeze(1)))
+                hidden_states = hidden_states + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(mlp_output)
+            else:
+                norm_hidden_states = self.norm1(hidden_states)
+                attn_output = self.attn(
+                    hidden_states=norm_hidden_states,
+                    encoder_hidden_states=norm_hidden_states,
+                    attention_mask=attention_mask,
+                    image_rotary_emb=image_rotary_emb,
+                )
+                hidden_states = hidden_states + self.norm2(attn_output)
+                mlp_output = self.feed_forward(self.ffn_norm1(hidden_states))
+                hidden_states = hidden_states + self.ffn_norm2(mlp_output)
+
+        return hidden_states
+
+
+class OmniGen2Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
+    """
+    OmniGen2 Transformer 2D Model.
+    
+    A transformer-based diffusion model for image generation with:
+    - Patch-based image processing
+    - Rotary position embeddings
+    - Multi-head attention
+    - Conditional generation support
+    
+    Args:
+        patch_size: Size of image patches
+        in_channels: Number of input channels
+        out_channels: Number of output channels (defaults to in_channels)
+        hidden_size: Size of hidden layers
+        num_layers: Number of transformer layers
+        num_refiner_layers: Number of refiner layers
+        num_attention_heads: Number of attention heads
+        num_kv_heads: Number of key-value heads
+        multiple_of: Multiple of which the hidden dimension should be
+        ffn_dim_multiplier: Multiplier for feed-forward network dimension
+        norm_eps: Epsilon value for normalization layers
+        axes_dim_rope: Dimensions for rotary position embeddings
+        axes_lens: Lengths for rotary position embeddings
+        text_feat_dim: Dimension of text features
+        timestep_scale: Scale factor for timestep embeddings
+        use_fused_rms_norm: Whether to use fused RMS normalization
+        use_fused_swiglu: Whether to use fused SwiGLU activation
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["Omnigen2TransformerBlock"]
+    _skip_layerwise_casting_patterns = ["x_embedder", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        out_channels: Optional[int] = None,
+        hidden_size: int = 2304,
+        num_layers: int = 26,
+        num_refiner_layers: int = 2,
+        num_attention_heads: int = 24,
+        num_kv_heads: int = 8,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        norm_eps: float = 1e-5,
+        axes_dim_rope: Tuple[int, int, int] = (32, 32, 32),
+        axes_lens: Tuple[int, int, int] = (300, 512, 512),
+        text_feat_dim: int = 1024,
+        timestep_scale: float = 1.0
+    ) -> None:
+        """Initialize the OmniGen2 transformer model."""
+        super().__init__()
+
+        # Validate configuration
+        if (hidden_size // num_attention_heads) != sum(axes_dim_rope):
+            raise ValueError(
+                f"hidden_size // num_attention_heads ({hidden_size // num_attention_heads}) "
+                f"must equal sum(axes_dim_rope) ({sum(axes_dim_rope)})"
+            )
+        
+        self.out_channels = out_channels or in_channels
+
+        # Initialize embeddings
+        self.rope_embedder = OmniGen2RotaryPosEmbed(
+            theta=10000,
+            axes_dim=axes_dim_rope,
+            axes_lens=axes_lens,
+            patch_size=patch_size,
+        )
+
+        self.x_embedder = nn.Linear(
+            in_features=patch_size * patch_size * in_channels,
+            out_features=hidden_size,
+        )
+
+        self.ref_image_patch_embedder = nn.Linear(
+            in_features=patch_size * patch_size * in_channels,
+            out_features=hidden_size,
+        )
+
+        self.time_caption_embed = Lumina2CombinedTimestepCaptionEmbedding(
+            hidden_size=hidden_size,
+            text_feat_dim=text_feat_dim,
+            norm_eps=norm_eps,
+            timestep_scale=timestep_scale
+        )
+
+        # Initialize transformer blocks
+        self.noise_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size,
+                num_attention_heads,
+                num_kv_heads,
+                multiple_of,
+                ffn_dim_multiplier,
+                norm_eps,
+                modulation=True
+            )
+            for _ in range(num_refiner_layers)
+        ])
+
+        self.ref_image_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size,
+                num_attention_heads,
+                num_kv_heads,
+                multiple_of,
+                ffn_dim_multiplier,
+                norm_eps,
+                modulation=True
+            )
+            for _ in range(num_refiner_layers)
+        ])
+
+        self.context_refiner = nn.ModuleList(
+            [
+                OmniGen2TransformerBlock(
+                    hidden_size,
+                    num_attention_heads,
+                    num_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    modulation=False
+                )
+                for _ in range(num_refiner_layers)
+            ]
+        )
+
+        # 3. Transformer blocks
+        self.layers = nn.ModuleList(
+            [
+                OmniGen2TransformerBlock(
+                    hidden_size,
+                    num_attention_heads,
+                    num_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    modulation=True
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output norm & projection
+        self.norm_out = LuminaLayerNormContinuous(
+            embedding_dim=hidden_size,
+            conditioning_embedding_dim=min(hidden_size, 1024),
+            elementwise_affine=False,
+            eps=1e-6,
+            bias=True,
+            out_dim=patch_size * patch_size * self.out_channels
+        )
+        
+        # Add learnable embeddings to distinguish different images
+        self.image_index_embedding = nn.Parameter(torch.randn(5, hidden_size)) # support max 5 ref images
+
+        self.gradient_checkpointing = False
+
+        self.initialize_weights()
+
+        # TeaCache settings
+        self.enable_teacache = False
+        self.teacache_rel_l1_thresh = 0.05
+        self.teacache_params = TeaCacheParams()
+
+        coefficients = [-5.48259225, 11.48772289, -4.47407401, 2.47730926, -0.03316487]
+        self.rescale_func = np.poly1d(coefficients)
+
+    def initialize_weights(self) -> None:
+        """
+        Initialize the weights of the model.
+        
+        Uses Xavier uniform initialization for linear layers.
+        """
+        nn.init.xavier_uniform_(self.x_embedder.weight)
+        nn.init.constant_(self.x_embedder.bias, 0.0)
+
+        nn.init.xavier_uniform_(self.ref_image_patch_embedder.weight)
+        nn.init.constant_(self.ref_image_patch_embedder.bias, 0.0)
+
+        nn.init.zeros_(self.norm_out.linear_1.weight)
+        nn.init.zeros_(self.norm_out.linear_1.bias)
+        nn.init.zeros_(self.norm_out.linear_2.weight)
+        nn.init.zeros_(self.norm_out.linear_2.bias)
+        
+        nn.init.normal_(self.image_index_embedding, std=0.02)
+
+    def img_patch_embed_and_refine(
+        self,
+        hidden_states,
+        ref_image_hidden_states,
+        padded_img_mask,
+        padded_ref_img_mask,
+        noise_rotary_emb,
+        ref_img_rotary_emb,
+        l_effective_ref_img_len,
+        l_effective_img_len,
+        temb
+    ):
+        batch_size = len(hidden_states)
+        max_combined_img_len = max([img_len + sum(ref_img_len) for img_len, ref_img_len in zip(l_effective_img_len, l_effective_ref_img_len)])
+    
+        hidden_states = self.x_embedder(hidden_states)
+        ref_image_hidden_states = self.ref_image_patch_embedder(ref_image_hidden_states)
+        
+        for i in range(batch_size):
+            shift = 0
+            for j, ref_img_len in enumerate(l_effective_ref_img_len[i]):
+                ref_image_hidden_states[i, shift:shift + ref_img_len, :] = ref_image_hidden_states[i, shift:shift + ref_img_len, :] + self.image_index_embedding[j]
+                shift += ref_img_len
+
+        for layer in self.noise_refiner:
+            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb)
+
+        flat_l_effective_ref_img_len = list(itertools.chain(*l_effective_ref_img_len))
+        num_ref_images = len(flat_l_effective_ref_img_len)
+        max_ref_img_len = max(flat_l_effective_ref_img_len)
+
+        batch_ref_img_mask = ref_image_hidden_states.new_zeros(num_ref_images, max_ref_img_len, dtype=torch.bool)
+        batch_ref_image_hidden_states = ref_image_hidden_states.new_zeros(num_ref_images, max_ref_img_len, self.config.hidden_size)
+        batch_ref_img_rotary_emb = hidden_states.new_zeros(num_ref_images, max_ref_img_len, ref_img_rotary_emb.shape[-1], dtype=ref_img_rotary_emb.dtype)
+        batch_temb = temb.new_zeros(num_ref_images, *temb.shape[1:], dtype=temb.dtype)
+
+        # sequence of ref imgs to batch
+        idx = 0
+        for i in range(batch_size):
+            shift = 0
+            for ref_img_len in l_effective_ref_img_len[i]:
+                batch_ref_img_mask[idx, :ref_img_len] = True
+                batch_ref_image_hidden_states[idx, :ref_img_len] = ref_image_hidden_states[i, shift:shift + ref_img_len]
+                batch_ref_img_rotary_emb[idx, :ref_img_len] = ref_img_rotary_emb[i, shift:shift + ref_img_len]
+                batch_temb[idx] = temb[i]
+                shift += ref_img_len
+                idx += 1
+
+        # refine ref imgs separately
+        for layer in self.ref_image_refiner:
+            batch_ref_image_hidden_states = layer(batch_ref_image_hidden_states, batch_ref_img_mask, batch_ref_img_rotary_emb, batch_temb)
+
+        # batch of ref imgs to sequence
+        idx = 0
+        for i in range(batch_size):
+            shift = 0
+            for ref_img_len in l_effective_ref_img_len[i]:
+                ref_image_hidden_states[i, shift:shift + ref_img_len] = batch_ref_image_hidden_states[idx, :ref_img_len]
+                shift += ref_img_len
+                idx += 1
+            
+        combined_img_hidden_states = hidden_states.new_zeros(batch_size, max_combined_img_len, self.config.hidden_size)
+        for i, (ref_img_len, img_len) in enumerate(zip(l_effective_ref_img_len, l_effective_img_len)):
+            combined_img_hidden_states[i, :sum(ref_img_len)] = ref_image_hidden_states[i, :sum(ref_img_len)]
+            combined_img_hidden_states[i, sum(ref_img_len):sum(ref_img_len) + img_len] = hidden_states[i, :img_len]
+
+        return combined_img_hidden_states
+
+    def flat_and_pad_to_seq(self, hidden_states, ref_image_hidden_states):
+        batch_size = len(hidden_states)
+        p = self.config.patch_size
+        device = hidden_states[0].device
+
+        img_sizes = [(img.size(1), img.size(2)) for img in hidden_states]
+        l_effective_img_len = [(H // p) * (W // p) for (H, W) in img_sizes]
+
+        if ref_image_hidden_states is not None:
+            ref_img_sizes = [[(img.size(1), img.size(2)) for img in imgs] if imgs is not None else None for imgs in ref_image_hidden_states]
+            l_effective_ref_img_len = [[(ref_img_size[0] // p) * (ref_img_size[1] // p) for ref_img_size in _ref_img_sizes] if _ref_img_sizes is not None else [0] for _ref_img_sizes in ref_img_sizes]
+        else:
+            ref_img_sizes = [None for _ in range(batch_size)]
+            l_effective_ref_img_len = [[0] for _ in range(batch_size)]
+
+        max_ref_img_len = max([sum(ref_img_len) for ref_img_len in l_effective_ref_img_len])
+        max_img_len = max(l_effective_img_len)
+
+        # ref image patch embeddings
+        flat_ref_img_hidden_states = []
+        for i in range(batch_size):
+            if ref_img_sizes[i] is not None:
+                imgs = []
+                for ref_img in ref_image_hidden_states[i]:
+                    C, H, W = ref_img.size()
+                    ref_img = rearrange(ref_img, 'c (h p1) (w p2) -> (h w) (p1 p2 c)', p1=p, p2=p)
+                    imgs.append(ref_img)
+
+                img = torch.cat(imgs, dim=0)
+                flat_ref_img_hidden_states.append(img)
+            else:
+                flat_ref_img_hidden_states.append(None)
+
+        # image patch embeddings
+        flat_hidden_states = []
+        for i in range(batch_size):
+            img = hidden_states[i]
+            C, H, W = img.size()
+            
+            img = rearrange(img, 'c (h p1) (w p2) -> (h w) (p1 p2 c)', p1=p, p2=p)
+            flat_hidden_states.append(img)
+        
+        padded_ref_img_hidden_states = torch.zeros(batch_size, max_ref_img_len, flat_hidden_states[0].shape[-1], device=device, dtype=flat_hidden_states[0].dtype)
+        padded_ref_img_mask = torch.zeros(batch_size, max_ref_img_len, dtype=torch.bool, device=device)
+        for i in range(batch_size):
+            if ref_img_sizes[i] is not None:
+                padded_ref_img_hidden_states[i, :sum(l_effective_ref_img_len[i])] = flat_ref_img_hidden_states[i]
+                padded_ref_img_mask[i, :sum(l_effective_ref_img_len[i])] = True
+
+        padded_hidden_states = torch.zeros(batch_size, max_img_len, flat_hidden_states[0].shape[-1], device=device, dtype=flat_hidden_states[0].dtype)
+        padded_img_mask = torch.zeros(batch_size, max_img_len, dtype=torch.bool, device=device)
+        for i in range(batch_size):
+            padded_hidden_states[i, :l_effective_img_len[i]] = flat_hidden_states[i]
+            padded_img_mask[i, :l_effective_img_len[i]] = True
+
+        return (
+            padded_hidden_states,
+            padded_ref_img_hidden_states,
+            padded_img_mask,
+            padded_ref_img_mask,
+            l_effective_ref_img_len,
+            l_effective_img_len,
+            ref_img_sizes,
+            img_sizes,
+        )
+    
+    def forward(
+        self,
+        hidden_states: Union[torch.Tensor, List[torch.Tensor]],
+        timestep: torch.Tensor,
+        text_hidden_states: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        text_attention_mask: torch.Tensor,
+        ref_image_hidden_states: Optional[List[List[torch.Tensor]]] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = False,
+    ) -> Union[torch.Tensor, Transformer2DModelOutput]:
+        enable_taylorseer = getattr(self, 'enable_taylorseer', False)
+        if enable_taylorseer:
+            cal_type(self.cache_dic, self.current)
+        
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        # 1. Condition, positional & patch embedding
+        batch_size = len(hidden_states)
+        is_hidden_states_tensor = isinstance(hidden_states, torch.Tensor)
+
+        if is_hidden_states_tensor:
+            assert hidden_states.ndim == 4
+            hidden_states = [_hidden_states for _hidden_states in hidden_states]
+
+        device = hidden_states[0].device
+
+        temb, text_hidden_states = self.time_caption_embed(timestep, text_hidden_states, hidden_states[0].dtype)
+
+        (
+            hidden_states,
+            ref_image_hidden_states,
+            img_mask,
+            ref_img_mask,
+            l_effective_ref_img_len,
+            l_effective_img_len,
+            ref_img_sizes,
+            img_sizes,
+        ) = self.flat_and_pad_to_seq(hidden_states, ref_image_hidden_states)
+        
+        (
+            context_rotary_emb,
+            ref_img_rotary_emb,
+            noise_rotary_emb,
+            rotary_emb,
+            encoder_seq_lengths,
+            seq_lengths,
+        ) = self.rope_embedder(
+            freqs_cis,
+            text_attention_mask,
+            l_effective_ref_img_len,
+            l_effective_img_len,
+            ref_img_sizes,
+            img_sizes,
+            device,
+        )
+
+        # 2. Context refinement
+        for layer in self.context_refiner:
+            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb)
+        
+        combined_img_hidden_states = self.img_patch_embed_and_refine(
+            hidden_states,
+            ref_image_hidden_states,
+            img_mask,
+            ref_img_mask,
+            noise_rotary_emb,
+            ref_img_rotary_emb,
+            l_effective_ref_img_len,
+            l_effective_img_len,
+            temb,
+        )
+
+        # 3. Joint Transformer blocks
+        max_seq_len = max(seq_lengths)
+
+        attention_mask = hidden_states.new_zeros(batch_size, max_seq_len, dtype=torch.bool)
+        joint_hidden_states = hidden_states.new_zeros(batch_size, max_seq_len, self.config.hidden_size)
+        for i, (encoder_seq_len, seq_len) in enumerate(zip(encoder_seq_lengths, seq_lengths)):
+            attention_mask[i, :seq_len] = True
+            joint_hidden_states[i, :encoder_seq_len] = text_hidden_states[i, :encoder_seq_len]
+            joint_hidden_states[i, encoder_seq_len:seq_len] = combined_img_hidden_states[i, :seq_len - encoder_seq_len]
+
+        hidden_states = joint_hidden_states
+
+        if self.enable_teacache:
+            teacache_hidden_states = hidden_states.clone()
+            teacache_temb = temb.clone()
+            modulated_inp, _, _, _ = self.layers[0].norm1(teacache_hidden_states, teacache_temb)
+            if self.teacache_params.is_first_or_last_step:
+                should_calc = True
+                self.teacache_params.accumulated_rel_l1_distance = 0
+            else:
+                self.teacache_params.accumulated_rel_l1_distance += self.rescale_func(
+                    ((modulated_inp - self.teacache_params.previous_modulated_inp).abs().mean() \
+                        / self.teacache_params.previous_modulated_inp.abs().mean()).cpu().item()
+                )
+                if self.teacache_params.accumulated_rel_l1_distance < self.teacache_rel_l1_thresh:
+                    should_calc = False
+                else:
+                    should_calc = True
+                    self.teacache_params.accumulated_rel_l1_distance = 0
+            self.teacache_params.previous_modulated_inp = modulated_inp
+
+        if self.enable_teacache:
+            if not should_calc:
+                hidden_states += self.teacache_params.previous_residual
+            else:
+                ori_hidden_states = hidden_states.clone()
+                for layer_idx, layer in enumerate(self.layers):
+                    if torch.is_grad_enabled() and self.gradient_checkpointing:
+                        hidden_states = self._gradient_checkpointing_func(
+                            layer, hidden_states, attention_mask, rotary_emb, temb
+                        )
+                    else:
+                        hidden_states = layer(hidden_states, attention_mask, rotary_emb, temb)
+                self.teacache_params.previous_residual = hidden_states - ori_hidden_states
+        else:
+            if enable_taylorseer:
+                self.current['stream'] = 'layers_stream'
+
+            for layer_idx, layer in enumerate(self.layers):
+                if enable_taylorseer:
+                    layer.current = self.current
+                    layer.cache_dic = self.cache_dic
+                    layer.enable_taylorseer = True
+                    self.current['layer'] = layer_idx
+
+                if torch.is_grad_enabled() and self.gradient_checkpointing:
+                    hidden_states = self._gradient_checkpointing_func(
+                        layer, hidden_states, attention_mask, rotary_emb, temb
+                    )
+                else:
+                    hidden_states = layer(hidden_states, attention_mask, rotary_emb, temb)
+
+        # 4. Output norm & projection
+        hidden_states = self.norm_out(hidden_states, temb)
+
+        p = self.config.patch_size
+        output = []
+        for i, (img_size, img_len, seq_len) in enumerate(zip(img_sizes, l_effective_img_len, seq_lengths)):
+            height, width = img_size
+            output.append(rearrange(hidden_states[i][seq_len - img_len:seq_len], '(h w) (p1 p2 c) -> c (h p1) (w p2)', h=height // p, w=width // p, p1=p, p2=p))
+        if is_hidden_states_tensor:
+            output = torch.stack(output, dim=0)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+            
+        if enable_taylorseer:
+            self.current['step'] += 1
+
+        if not return_dict:
+            return output
+        return Transformer2DModelOutput(sample=output)

omnigen2/ops/triton/layer_norm.py

@@ -0,0 +1,1257 @@
+# Copyright (c) 2024, Tri Dao.
+# Implement dropout + residual + layer_norm / rms_norm.
+
+# Based on the Triton LayerNorm tutorial: https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
+# For the backward pass, we keep weight_grad and bias_grad in registers and accumulate.
+# This is faster for dimensions up to 8k, but after that it's much slower due to register spilling.
+# The models we train have hidden dim up to 8k anyway (e.g. Llama 70B), so this is fine.
+
+import math
+
+import torch
+import torch.nn.functional as F
+
+import triton
+import triton.language as tl
+
+
+from typing import Callable
+
+
+def custom_amp_decorator(dec: Callable, cuda_amp_deprecated: bool):
+    def decorator(*args, **kwargs):
+        if cuda_amp_deprecated:
+            kwargs["device_type"] = "cuda"
+        return dec(*args, **kwargs)
+    return decorator
+
+
+if hasattr(torch.amp, "custom_fwd"): # type: ignore[attr-defined]
+    deprecated = True
+    from torch.amp import custom_fwd, custom_bwd # type: ignore[attr-defined]
+else:
+    deprecated = False
+    from torch.cuda.amp import custom_fwd, custom_bwd
+
+custom_fwd = custom_amp_decorator(custom_fwd, deprecated)
+custom_bwd = custom_amp_decorator(custom_bwd, deprecated)
+
+
+def triton_autotune_configs():
+    # Return configs with a valid warp count for the current device
+    configs=[]
+    # Maximum threads per block is architecture-dependent in theory, but in reality all are 1024
+    max_threads_per_block=1024
+    # Default to warp size 32 if not defined by device
+    warp_size=getattr(torch.cuda.get_device_properties(torch.cuda.current_device()), "warp_size", 32)
+    # Autotune for warp counts which are powers of 2 and do not exceed thread per block limit
+    warp_count=1
+    while warp_count*warp_size <= max_threads_per_block:
+        configs.append(triton.Config({}, num_warps=warp_count))
+        warp_count*=2
+    return configs
+
+def layer_norm_ref(
+    x,
+    weight,
+    bias,
+    residual=None,
+    x1=None,
+    weight1=None,
+    bias1=None,
+    eps=1e-6,
+    dropout_p=0.0,
+    rowscale=None,
+    prenorm=False,
+    zero_centered_weight=False,
+    dropout_mask=None,
+    dropout_mask1=None,
+    upcast=False,
+):
+    dtype = x.dtype
+    if upcast:
+        x = x.float()
+        weight = weight.float()
+        bias = bias.float() if bias is not None else None
+        residual = residual.float() if residual is not None else residual
+        x1 = x1.float() if x1 is not None else None
+        weight1 = weight1.float() if weight1 is not None else None
+        bias1 = bias1.float() if bias1 is not None else None
+    if zero_centered_weight:
+        weight = weight + 1.0
+        if weight1 is not None:
+            weight1 = weight1 + 1.0
+    if x1 is not None:
+        assert rowscale is None, "rowscale is not supported with parallel LayerNorm"
+    if rowscale is not None:
+        x = x * rowscale[..., None]
+    if dropout_p > 0.0:
+        if dropout_mask is not None:
+            x = x.masked_fill(~dropout_mask, 0.0) / (1.0 - dropout_p)
+        else:
+            x = F.dropout(x, p=dropout_p)
+        if x1 is not None:
+            if dropout_mask1 is not None:
+                x1 = x1.masked_fill(~dropout_mask1, 0.0) / (1.0 - dropout_p)
+            else:
+                x1 = F.dropout(x1, p=dropout_p)
+    if x1 is not None:
+        x = x + x1
+    if residual is not None:
+        x = (x + residual).to(x.dtype)
+    out = F.layer_norm(x.to(weight.dtype), x.shape[-1:], weight=weight, bias=bias, eps=eps).to(
+        dtype
+    )
+    if weight1 is None:
+        return out if not prenorm else (out, x)
+    else:
+        out1 = F.layer_norm(
+            x.to(weight1.dtype), x.shape[-1:], weight=weight1, bias=bias1, eps=eps
+        ).to(dtype)
+        return (out, out1) if not prenorm else (out, out1, x)
+
+
+def rms_norm_ref(
+    x,
+    weight,
+    bias,
+    residual=None,
+    x1=None,
+    weight1=None,
+    bias1=None,
+    eps=1e-6,
+    dropout_p=0.0,
+    rowscale=None,
+    prenorm=False,
+    zero_centered_weight=False,
+    dropout_mask=None,
+    dropout_mask1=None,
+    upcast=False,
+):
+    dtype = x.dtype
+    if upcast:
+        x = x.float()
+        weight = weight.float()
+        bias = bias.float() if bias is not None else None
+        residual = residual.float() if residual is not None else residual
+        x1 = x1.float() if x1 is not None else None
+        weight1 = weight1.float() if weight1 is not None else None
+        bias1 = bias1.float() if bias1 is not None else None
+    if zero_centered_weight:
+        weight = weight + 1.0
+        if weight1 is not None:
+            weight1 = weight1 + 1.0
+    if x1 is not None:
+        assert rowscale is None, "rowscale is not supported with parallel LayerNorm"
+    if rowscale is not None:
+        x = x * rowscale[..., None]
+    if dropout_p > 0.0:
+        if dropout_mask is not None:
+            x = x.masked_fill(~dropout_mask, 0.0) / (1.0 - dropout_p)
+        else:
+            x = F.dropout(x, p=dropout_p)
+        if x1 is not None:
+            if dropout_mask1 is not None:
+                x1 = x1.masked_fill(~dropout_mask1, 0.0) / (1.0 - dropout_p)
+            else:
+                x1 = F.dropout(x1, p=dropout_p)
+    if x1 is not None:
+        x = x + x1
+    if residual is not None:
+        x = (x + residual).to(x.dtype)
+    rstd = 1 / torch.sqrt((x.square()).mean(dim=-1, keepdim=True) + eps)
+    out = ((x * rstd * weight) + bias if bias is not None else (x * rstd * weight)).to(dtype)
+    if weight1 is None:
+        return out if not prenorm else (out, x)
+    else:
+        out1 = ((x * rstd * weight1) + bias1 if bias1 is not None else (x * rstd * weight1)).to(
+            dtype
+        )
+        return (out, out1) if not prenorm else (out, out1, x)
+
+
+@triton.autotune(
+    configs=triton_autotune_configs(),
+    key=["N", "HAS_RESIDUAL", "STORE_RESIDUAL_OUT", "IS_RMS_NORM", "HAS_BIAS"],
+)
+# @triton.heuristics({"HAS_BIAS": lambda args: args["B"] is not None})
+# @triton.heuristics({"HAS_RESIDUAL": lambda args: args["RESIDUAL"] is not None})
+@triton.heuristics({"HAS_X1": lambda args: args["X1"] is not None})
+@triton.heuristics({"HAS_W1": lambda args: args["W1"] is not None})
+@triton.heuristics({"HAS_B1": lambda args: args["B1"] is not None})
+@triton.jit
+def _layer_norm_fwd_1pass_kernel(
+    X,  # pointer to the input
+    Y,  # pointer to the output
+    W,  # pointer to the weights
+    B,  # pointer to the biases
+    RESIDUAL,  # pointer to the residual
+    X1,
+    W1,
+    B1,
+    Y1,
+    RESIDUAL_OUT,  # pointer to the residual
+    ROWSCALE,
+    SEEDS,  # Dropout seeds for each row
+    DROPOUT_MASK,
+    Mean,  # pointer to the mean
+    Rstd,  # pointer to the 1/std
+    stride_x_row,  # how much to increase the pointer when moving by 1 row
+    stride_y_row,
+    stride_res_row,
+    stride_res_out_row,
+    stride_x1_row,
+    stride_y1_row,
+    M,  # number of rows in X
+    N,  # number of columns in X
+    eps,  # epsilon to avoid division by zero
+    dropout_p,  # Dropout probability
+    zero_centered_weight,  # If true, add 1.0 to the weight
+    IS_RMS_NORM: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    HAS_RESIDUAL: tl.constexpr,
+    STORE_RESIDUAL_OUT: tl.constexpr,
+    HAS_BIAS: tl.constexpr,
+    HAS_DROPOUT: tl.constexpr,
+    STORE_DROPOUT_MASK: tl.constexpr,
+    HAS_ROWSCALE: tl.constexpr,
+    HAS_X1: tl.constexpr,
+    HAS_W1: tl.constexpr,
+    HAS_B1: tl.constexpr,
+):
+    # Map the program id to the row of X and Y it should compute.
+    row = tl.program_id(0)
+    X += row * stride_x_row
+    Y += row * stride_y_row
+    if HAS_RESIDUAL:
+        RESIDUAL += row * stride_res_row
+    if STORE_RESIDUAL_OUT:
+        RESIDUAL_OUT += row * stride_res_out_row
+    if HAS_X1:
+        X1 += row * stride_x1_row
+    if HAS_W1:
+        Y1 += row * stride_y1_row
+    # Compute mean and variance
+    cols = tl.arange(0, BLOCK_N)
+    x = tl.load(X + cols, mask=cols < N, other=0.0).to(tl.float32)
+    if HAS_ROWSCALE:
+        rowscale = tl.load(ROWSCALE + row).to(tl.float32)
+        x *= rowscale
+    if HAS_DROPOUT:
+        # Compute dropout mask
+        # 7 rounds is good enough, and reduces register pressure
+        keep_mask = tl.rand(tl.load(SEEDS + row).to(tl.uint32), cols, n_rounds=7) > dropout_p
+        x = tl.where(keep_mask, x / (1.0 - dropout_p), 0.0)
+        if STORE_DROPOUT_MASK:
+            tl.store(DROPOUT_MASK + row * N + cols, keep_mask, mask=cols < N)
+    if HAS_X1:
+        x1 = tl.load(X1 + cols, mask=cols < N, other=0.0).to(tl.float32)
+        if HAS_ROWSCALE:
+            rowscale = tl.load(ROWSCALE + M + row).to(tl.float32)
+            x1 *= rowscale
+        if HAS_DROPOUT:
+            # Compute dropout mask
+            # 7 rounds is good enough, and reduces register pressure
+            keep_mask = (
+                tl.rand(tl.load(SEEDS + M + row).to(tl.uint32), cols, n_rounds=7) > dropout_p
+            )
+            x1 = tl.where(keep_mask, x1 / (1.0 - dropout_p), 0.0)
+            if STORE_DROPOUT_MASK:
+                tl.store(DROPOUT_MASK + (M + row) * N + cols, keep_mask, mask=cols < N)
+        x += x1
+    if HAS_RESIDUAL:
+        residual = tl.load(RESIDUAL + cols, mask=cols < N, other=0.0).to(tl.float32)
+        x += residual
+    if STORE_RESIDUAL_OUT:
+        tl.store(RESIDUAL_OUT + cols, x, mask=cols < N)
+    if not IS_RMS_NORM:
+        mean = tl.sum(x, axis=0) / N
+        tl.store(Mean + row, mean)
+        xbar = tl.where(cols < N, x - mean, 0.0)
+        var = tl.sum(xbar * xbar, axis=0) / N
+    else:
+        xbar = tl.where(cols < N, x, 0.0)
+        var = tl.sum(xbar * xbar, axis=0) / N
+    rstd = 1 / tl.sqrt(var + eps)
+    tl.store(Rstd + row, rstd)
+    # Normalize and apply linear transformation
+    mask = cols < N
+    w = tl.load(W + cols, mask=mask).to(tl.float32)
+    if zero_centered_weight:
+        w += 1.0
+    if HAS_BIAS:
+        b = tl.load(B + cols, mask=mask).to(tl.float32)
+    x_hat = (x - mean) * rstd if not IS_RMS_NORM else x * rstd
+    y = x_hat * w + b if HAS_BIAS else x_hat * w
+    # Write output
+    tl.store(Y + cols, y, mask=mask)
+    if HAS_W1:
+        w1 = tl.load(W1 + cols, mask=mask).to(tl.float32)
+        if zero_centered_weight:
+            w1 += 1.0
+        if HAS_B1:
+            b1 = tl.load(B1 + cols, mask=mask).to(tl.float32)
+        y1 = x_hat * w1 + b1 if HAS_B1 else x_hat * w1
+        tl.store(Y1 + cols, y1, mask=mask)
+
+
+def _layer_norm_fwd(
+    x,
+    weight,
+    bias,
+    eps,
+    residual=None,
+    x1=None,
+    weight1=None,
+    bias1=None,
+    dropout_p=0.0,
+    rowscale=None,
+    out_dtype=None,
+    residual_dtype=None,
+    zero_centered_weight=False,
+    is_rms_norm=False,
+    return_dropout_mask=False,
+    out=None,
+    residual_out=None
+):
+    if residual is not None:
+        residual_dtype = residual.dtype
+    M, N = x.shape
+    assert x.stride(-1) == 1
+    if residual is not None:
+        assert residual.stride(-1) == 1
+        assert residual.shape == (M, N)
+    assert weight.shape == (N,)
+    assert weight.stride(-1) == 1
+    if bias is not None:
+        assert bias.stride(-1) == 1
+        assert bias.shape == (N,)
+    if x1 is not None:
+        assert x1.shape == x.shape
+        assert rowscale is None
+        assert x1.stride(-1) == 1
+    if weight1 is not None:
+        assert weight1.shape == (N,)
+        assert weight1.stride(-1) == 1
+    if bias1 is not None:
+        assert bias1.shape == (N,)
+        assert bias1.stride(-1) == 1
+    if rowscale is not None:
+        assert rowscale.is_contiguous()
+        assert rowscale.shape == (M,)
+    # allocate output
+    if out is None:
+        out = torch.empty_like(x, dtype=x.dtype if out_dtype is None else out_dtype)
+    else:
+        assert out.shape == x.shape
+    assert out.stride(-1) == 1
+    if weight1 is not None:
+        y1 = torch.empty_like(out)
+        assert y1.stride(-1) == 1
+    else:
+        y1 = None
+    if (
+        residual is not None
+        or (residual_dtype is not None and residual_dtype != x.dtype)
+        or dropout_p > 0.0
+        or rowscale is not None
+        or x1 is not None
+    ):
+        if residual_out is None:
+            residual_out = torch.empty(
+                M, N, device=x.device, dtype=residual_dtype if residual_dtype is not None else x.dtype
+            )
+        else:
+            assert residual_out.shape == x.shape
+        assert residual_out.stride(-1) == 1
+    else:
+        residual_out = None
+    mean = torch.empty((M,), dtype=torch.float32, device=x.device) if not is_rms_norm else None
+    rstd = torch.empty((M,), dtype=torch.float32, device=x.device)
+    if dropout_p > 0.0:
+        seeds = torch.randint(
+            2**32, (M if x1 is None else 2 * M,), device=x.device, dtype=torch.int64
+        )
+    else:
+        seeds = None
+    if return_dropout_mask and dropout_p > 0.0:
+        dropout_mask = torch.empty(M if x1 is None else 2 * M, N, device=x.device, dtype=torch.bool)
+    else:
+        dropout_mask = None
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
+    if N > BLOCK_N:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    with torch.cuda.device(x.device.index):
+        _layer_norm_fwd_1pass_kernel[(M,)](
+            x,
+            out,
+            weight,
+            bias,
+            residual,
+            x1,
+            weight1,
+            bias1,
+            y1,
+            residual_out,
+            rowscale,
+            seeds,
+            dropout_mask,
+            mean,
+            rstd,
+            x.stride(0),
+            out.stride(0),
+            residual.stride(0) if residual is not None else 0,
+            residual_out.stride(0) if residual_out is not None else 0,
+            x1.stride(0) if x1 is not None else 0,
+            y1.stride(0) if y1 is not None else 0,
+            M,
+            N,
+            eps,
+            dropout_p,
+            zero_centered_weight,
+            is_rms_norm,
+            BLOCK_N,
+            residual is not None,
+            residual_out is not None,
+            bias is not None,
+            dropout_p > 0.0,
+            dropout_mask is not None,
+            rowscale is not None,
+        )
+    # residual_out is None if residual is None and residual_dtype == input_dtype and dropout_p == 0.0
+    if dropout_mask is not None and x1 is not None:
+        dropout_mask, dropout_mask1 = dropout_mask.tensor_split(2, dim=0)
+    else:
+        dropout_mask1 = None
+    return (
+        out,
+        y1,
+        mean,
+        rstd,
+        residual_out if residual_out is not None else x,
+        seeds,
+        dropout_mask,
+        dropout_mask1,
+    )
+
+
+@triton.autotune(
+    configs=triton_autotune_configs(),
+    key=["N", "HAS_DRESIDUAL", "STORE_DRESIDUAL", "IS_RMS_NORM", "HAS_BIAS", "HAS_DROPOUT"],
+)
+# @triton.heuristics({"HAS_BIAS": lambda args: args["B"] is not None})
+# @triton.heuristics({"HAS_DRESIDUAL": lambda args: args["DRESIDUAL"] is not None})
+# @triton.heuristics({"STORE_DRESIDUAL": lambda args: args["DRESIDUAL_IN"] is not None})
+@triton.heuristics({"HAS_ROWSCALE": lambda args: args["ROWSCALE"] is not None})
+@triton.heuristics({"HAS_DY1": lambda args: args["DY1"] is not None})
+@triton.heuristics({"HAS_DX1": lambda args: args["DX1"] is not None})
+@triton.heuristics({"HAS_B1": lambda args: args["DB1"] is not None})
+@triton.heuristics({"RECOMPUTE_OUTPUT": lambda args: args["Y"] is not None})
+@triton.jit
+def _layer_norm_bwd_kernel(
+    X,  # pointer to the input
+    W,  # pointer to the weights
+    B,  # pointer to the biases
+    Y,  # pointer to the output to be recomputed
+    DY,  # pointer to the output gradient
+    DX,  # pointer to the input gradient
+    DW,  # pointer to the partial sum of weights gradient
+    DB,  # pointer to the partial sum of biases gradient
+    DRESIDUAL,
+    W1,
+    DY1,
+    DX1,
+    DW1,
+    DB1,
+    DRESIDUAL_IN,
+    ROWSCALE,
+    SEEDS,
+    Mean,  # pointer to the mean
+    Rstd,  # pointer to the 1/std
+    stride_x_row,  # how much to increase the pointer when moving by 1 row
+    stride_y_row,
+    stride_dy_row,
+    stride_dx_row,
+    stride_dres_row,
+    stride_dy1_row,
+    stride_dx1_row,
+    stride_dres_in_row,
+    M,  # number of rows in X
+    N,  # number of columns in X
+    eps,  # epsilon to avoid division by zero
+    dropout_p,
+    zero_centered_weight,
+    rows_per_program,
+    IS_RMS_NORM: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    HAS_DRESIDUAL: tl.constexpr,
+    STORE_DRESIDUAL: tl.constexpr,
+    HAS_BIAS: tl.constexpr,
+    HAS_DROPOUT: tl.constexpr,
+    HAS_ROWSCALE: tl.constexpr,
+    HAS_DY1: tl.constexpr,
+    HAS_DX1: tl.constexpr,
+    HAS_B1: tl.constexpr,
+    RECOMPUTE_OUTPUT: tl.constexpr,
+):
+    # Map the program id to the elements of X, DX, and DY it should compute.
+    row_block_id = tl.program_id(0)
+    row_start = row_block_id * rows_per_program
+    # Do not early exit if row_start >= M, because we need to write DW and DB
+    cols = tl.arange(0, BLOCK_N)
+    mask = cols < N
+    X += row_start * stride_x_row
+    if HAS_DRESIDUAL:
+        DRESIDUAL += row_start * stride_dres_row
+    if STORE_DRESIDUAL:
+        DRESIDUAL_IN += row_start * stride_dres_in_row
+    DY += row_start * stride_dy_row
+    DX += row_start * stride_dx_row
+    if HAS_DY1:
+        DY1 += row_start * stride_dy1_row
+    if HAS_DX1:
+        DX1 += row_start * stride_dx1_row
+    if RECOMPUTE_OUTPUT:
+        Y += row_start * stride_y_row
+    w = tl.load(W + cols, mask=mask).to(tl.float32)
+    if zero_centered_weight:
+        w += 1.0
+    if RECOMPUTE_OUTPUT and HAS_BIAS:
+        b = tl.load(B + cols, mask=mask, other=0.0).to(tl.float32)
+    if HAS_DY1:
+        w1 = tl.load(W1 + cols, mask=mask).to(tl.float32)
+        if zero_centered_weight:
+            w1 += 1.0
+    dw = tl.zeros((BLOCK_N,), dtype=tl.float32)
+    if HAS_BIAS:
+        db = tl.zeros((BLOCK_N,), dtype=tl.float32)
+    if HAS_DY1:
+        dw1 = tl.zeros((BLOCK_N,), dtype=tl.float32)
+        if HAS_B1:
+            db1 = tl.zeros((BLOCK_N,), dtype=tl.float32)
+    row_end = min((row_block_id + 1) * rows_per_program, M)
+    for row in range(row_start, row_end):
+        # Load data to SRAM
+        x = tl.load(X + cols, mask=mask, other=0).to(tl.float32)
+        dy = tl.load(DY + cols, mask=mask, other=0).to(tl.float32)
+        if HAS_DY1:
+            dy1 = tl.load(DY1 + cols, mask=mask, other=0).to(tl.float32)
+        if not IS_RMS_NORM:
+            mean = tl.load(Mean + row)
+        rstd = tl.load(Rstd + row)
+        # Compute dx
+        xhat = (x - mean) * rstd if not IS_RMS_NORM else x * rstd
+        xhat = tl.where(mask, xhat, 0.0)
+        if RECOMPUTE_OUTPUT:
+            y = xhat * w + b if HAS_BIAS else xhat * w
+            tl.store(Y + cols, y, mask=mask)
+        wdy = w * dy
+        dw += dy * xhat
+        if HAS_BIAS:
+            db += dy
+        if HAS_DY1:
+            wdy += w1 * dy1
+            dw1 += dy1 * xhat
+            if HAS_B1:
+                db1 += dy1
+        if not IS_RMS_NORM:
+            c1 = tl.sum(xhat * wdy, axis=0) / N
+            c2 = tl.sum(wdy, axis=0) / N
+            dx = (wdy - (xhat * c1 + c2)) * rstd
+        else:
+            c1 = tl.sum(xhat * wdy, axis=0) / N
+            dx = (wdy - xhat * c1) * rstd
+        if HAS_DRESIDUAL:
+            dres = tl.load(DRESIDUAL + cols, mask=mask, other=0).to(tl.float32)
+            dx += dres
+        # Write dx
+        if STORE_DRESIDUAL:
+            tl.store(DRESIDUAL_IN + cols, dx, mask=mask)
+        if HAS_DX1:
+            if HAS_DROPOUT:
+                keep_mask = (
+                    tl.rand(tl.load(SEEDS + M + row).to(tl.uint32), cols, n_rounds=7) > dropout_p
+                )
+                dx1 = tl.where(keep_mask, dx / (1.0 - dropout_p), 0.0)
+            else:
+                dx1 = dx
+            tl.store(DX1 + cols, dx1, mask=mask)
+        if HAS_DROPOUT:
+            keep_mask = tl.rand(tl.load(SEEDS + row).to(tl.uint32), cols, n_rounds=7) > dropout_p
+            dx = tl.where(keep_mask, dx / (1.0 - dropout_p), 0.0)
+        if HAS_ROWSCALE:
+            rowscale = tl.load(ROWSCALE + row).to(tl.float32)
+            dx *= rowscale
+        tl.store(DX + cols, dx, mask=mask)
+
+        X += stride_x_row
+        if HAS_DRESIDUAL:
+            DRESIDUAL += stride_dres_row
+        if STORE_DRESIDUAL:
+            DRESIDUAL_IN += stride_dres_in_row
+        if RECOMPUTE_OUTPUT:
+            Y += stride_y_row
+        DY += stride_dy_row
+        DX += stride_dx_row
+        if HAS_DY1:
+            DY1 += stride_dy1_row
+        if HAS_DX1:
+            DX1 += stride_dx1_row
+    tl.store(DW + row_block_id * N + cols, dw, mask=mask)
+    if HAS_BIAS:
+        tl.store(DB + row_block_id * N + cols, db, mask=mask)
+    if HAS_DY1:
+        tl.store(DW1 + row_block_id * N + cols, dw1, mask=mask)
+        if HAS_B1:
+            tl.store(DB1 + row_block_id * N + cols, db1, mask=mask)
+
+
+def _layer_norm_bwd(
+    dy,
+    x,
+    weight,
+    bias,
+    eps,
+    mean,
+    rstd,
+    dresidual=None,
+    dy1=None,
+    weight1=None,
+    bias1=None,
+    seeds=None,
+    dropout_p=0.0,
+    rowscale=None,
+    has_residual=False,
+    has_x1=False,
+    zero_centered_weight=False,
+    is_rms_norm=False,
+    x_dtype=None,
+    recompute_output=False,
+):
+    M, N = x.shape
+    assert x.stride(-1) == 1
+    assert dy.stride(-1) == 1
+    assert dy.shape == (M, N)
+    if dresidual is not None:
+        assert dresidual.stride(-1) == 1
+        assert dresidual.shape == (M, N)
+    assert weight.shape == (N,)
+    assert weight.stride(-1) == 1
+    if bias is not None:
+        assert bias.stride(-1) == 1
+        assert bias.shape == (N,)
+    if dy1 is not None:
+        assert weight1 is not None
+        assert dy1.shape == dy.shape
+        assert dy1.stride(-1) == 1
+    if weight1 is not None:
+        assert weight1.shape == (N,)
+        assert weight1.stride(-1) == 1
+    if bias1 is not None:
+        assert bias1.shape == (N,)
+        assert bias1.stride(-1) == 1
+    if seeds is not None:
+        assert seeds.is_contiguous()
+        assert seeds.shape == (M if not has_x1 else M * 2,)
+    if rowscale is not None:
+        assert rowscale.is_contiguous()
+        assert rowscale.shape == (M,)
+    # allocate output
+    dx = (
+        torch.empty_like(x)
+        if x_dtype is None
+        else torch.empty(M, N, dtype=x_dtype, device=x.device)
+    )
+    dresidual_in = (
+        torch.empty_like(x)
+        if has_residual
+        and (dx.dtype != x.dtype or dropout_p > 0.0 or rowscale is not None or has_x1)
+        else None
+    )
+    dx1 = torch.empty_like(dx) if (has_x1 and dropout_p > 0.0) else None
+    y = torch.empty(M, N, dtype=dy.dtype, device=dy.device) if recompute_output else None
+    if recompute_output:
+        assert weight1 is None, "recompute_output is not supported with parallel LayerNorm"
+
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BLOCK_N = min(MAX_FUSED_SIZE, triton.next_power_of_2(N))
+    if N > BLOCK_N:
+        raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
+    # Increasing the multiple (e.g. 8) will allow more thread blocks to be launched and hide the
+    # latency of the gmem reads/writes, but will increase the time of summing up dw / db.
+    sm_count = torch.cuda.get_device_properties(x.device).multi_processor_count * 8
+    _dw = torch.empty((sm_count, N), dtype=torch.float32, device=weight.device)
+    _db = (
+        torch.empty((sm_count, N), dtype=torch.float32, device=bias.device)
+        if bias is not None
+        else None
+    )
+    _dw1 = torch.empty_like(_dw) if weight1 is not None else None
+    _db1 = torch.empty_like(_db) if bias1 is not None else None
+    rows_per_program = math.ceil(M / sm_count)
+    grid = (sm_count,)
+    with torch.cuda.device(x.device.index):
+        _layer_norm_bwd_kernel[grid](
+            x,
+            weight,
+            bias,
+            y,
+            dy,
+            dx,
+            _dw,
+            _db,
+            dresidual,
+            weight1,
+            dy1,
+            dx1,
+            _dw1,
+            _db1,
+            dresidual_in,
+            rowscale,
+            seeds,
+            mean,
+            rstd,
+            x.stride(0),
+            0 if not recompute_output else y.stride(0),
+            dy.stride(0),
+            dx.stride(0),
+            dresidual.stride(0) if dresidual is not None else 0,
+            dy1.stride(0) if dy1 is not None else 0,
+            dx1.stride(0) if dx1 is not None else 0,
+            dresidual_in.stride(0) if dresidual_in is not None else 0,
+            M,
+            N,
+            eps,
+            dropout_p,
+            zero_centered_weight,
+            rows_per_program,
+            is_rms_norm,
+            BLOCK_N,
+            dresidual is not None,
+            dresidual_in is not None,
+            bias is not None,
+            dropout_p > 0.0,
+        )
+    dw = _dw.sum(0).to(weight.dtype)
+    db = _db.sum(0).to(bias.dtype) if bias is not None else None
+    dw1 = _dw1.sum(0).to(weight1.dtype) if weight1 is not None else None
+    db1 = _db1.sum(0).to(bias1.dtype) if bias1 is not None else None
+    # Don't need to compute dresidual_in separately in this case
+    if has_residual and dx.dtype == x.dtype and dropout_p == 0.0 and rowscale is None:
+        dresidual_in = dx
+    if has_x1 and dropout_p == 0.0:
+        dx1 = dx
+    return (
+        (dx, dw, db, dresidual_in, dx1, dw1, db1)
+        if not recompute_output
+        else (dx, dw, db, dresidual_in, dx1, dw1, db1, y)
+    )
+
+
+class LayerNormFn(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x,
+        weight,
+        bias,
+        residual=None,
+        x1=None,
+        weight1=None,
+        bias1=None,
+        eps=1e-6,
+        dropout_p=0.0,
+        rowscale=None,
+        prenorm=False,
+        residual_in_fp32=False,
+        zero_centered_weight=False,
+        is_rms_norm=False,
+        return_dropout_mask=False,
+        out=None,
+        residual_out=None
+    ):
+        x_shape_og = x.shape
+        # Check for zero sequence length
+        if x.numel() == 0:
+            ctx.zero_seq_length = True
+            # Only save minimal required tensors for backward
+            # ctx.save_for_backward(weight, bias, weight1, bias1)
+            ctx.x_shape_og = x_shape_og
+            ctx.weight_shape = weight.shape
+            ctx.weight_dtype = weight.dtype
+            ctx.weight_device = weight.device
+
+            ctx.has_bias = bias is not None
+            ctx.bias_shape = bias.shape if bias is not None else None
+            ctx.bias_dtype = bias.dtype if bias is not None else None
+            ctx.bias_device = bias.device if bias is not None else None
+
+            ctx.has_weight1 = weight1 is not None
+            ctx.weight1_shape = weight1.shape if weight1 is not None else None
+            ctx.weight1_dtype = weight1.dtype if weight1 is not None else None
+            ctx.weight1_device = weight1.device if weight1 is not None else None
+
+            ctx.has_bias1 = bias1 is not None
+            ctx.bias1_shape = bias1.shape if bias1 is not None else None
+            ctx.bias1_dtype = bias1.dtype if bias1 is not None else None
+            ctx.bias1_device = bias1.device if bias1 is not None else None
+
+            ctx.has_residual = residual is not None
+            ctx.has_x1 = x1 is not None
+            ctx.dropout_p = dropout_p
+
+            # Handle output tensors with correct dtype
+            y = x  # Preserve input tensor properties
+            y1 = torch.empty_like(x) if x1 is not None else None
+            
+            # Only create residual_out if prenorm is True
+            residual_out = torch.empty(x.shape, 
+                                    dtype=torch.float32 if residual_in_fp32 else x.dtype,
+                                    device=x.device) if prenorm else None
+            
+            # Handle dropout masks
+            dropout_mask = None
+            dropout_mask1 = None
+            if return_dropout_mask:
+                dropout_mask = torch.empty_like(x, dtype=torch.uint8)
+                if x1 is not None:
+                    dropout_mask1 = torch.empty_like(x, dtype=torch.uint8)
+
+            # Return based on configuration
+            if not return_dropout_mask:
+                if weight1 is None:
+                    return y if not prenorm else (y, residual_out)
+                else:
+                    return (y, y1) if not prenorm else (y, y1, residual_out)
+            else:
+                if weight1 is None:
+                    return ((y, dropout_mask, dropout_mask1) if not prenorm 
+                        else (y, residual_out, dropout_mask, dropout_mask1))
+                else:
+                    return ((y, y1, dropout_mask, dropout_mask1) if not prenorm 
+                        else (y, y1, residual_out, dropout_mask, dropout_mask1))
+
+        ctx.zero_seq_length = False  
+        # reshape input data into 2D tensor
+        x = x.reshape(-1, x.shape[-1])
+        if x.stride(-1) != 1:
+            x = x.contiguous()
+        if residual is not None:
+            assert residual.shape == x_shape_og
+            residual = residual.reshape(-1, residual.shape[-1])
+            if residual.stride(-1) != 1:
+                residual = residual.contiguous()
+        if x1 is not None:
+            assert x1.shape == x_shape_og
+            assert rowscale is None, "rowscale is not supported with parallel LayerNorm"
+            x1 = x1.reshape(-1, x1.shape[-1])
+            if x1.stride(-1) != 1:
+                x1 = x1.contiguous()
+        weight = weight.contiguous()
+        if bias is not None:
+            bias = bias.contiguous()
+        if weight1 is not None:
+            weight1 = weight1.contiguous()
+        if bias1 is not None:
+            bias1 = bias1.contiguous()
+        if rowscale is not None:
+            rowscale = rowscale.reshape(-1).contiguous()
+        residual_dtype = (
+            residual.dtype
+            if residual is not None
+            else (torch.float32 if residual_in_fp32 else None)
+        )
+        if out is not None:
+            out = out.reshape(-1, out.shape[-1])
+        if residual_out is not None:
+            residual_out = residual_out.reshape(-1, residual_out.shape[-1])
+        y, y1, mean, rstd, residual_out, seeds, dropout_mask, dropout_mask1 = _layer_norm_fwd(
+            x,
+            weight,
+            bias,
+            eps,
+            residual,
+            x1,
+            weight1,
+            bias1,
+            dropout_p=dropout_p,
+            rowscale=rowscale,
+            residual_dtype=residual_dtype,
+            zero_centered_weight=zero_centered_weight,
+            is_rms_norm=is_rms_norm,
+            return_dropout_mask=return_dropout_mask,
+            out=out,
+            residual_out=residual_out
+        )
+        ctx.save_for_backward(
+            residual_out, weight, bias, weight1, bias1, rowscale, seeds, mean, rstd
+        )
+        ctx.x_shape_og = x_shape_og
+        ctx.eps = eps
+        ctx.dropout_p = dropout_p
+        ctx.is_rms_norm = is_rms_norm
+        ctx.has_residual = residual is not None
+        ctx.has_x1 = x1 is not None
+        ctx.prenorm = prenorm
+        ctx.x_dtype = x.dtype
+        ctx.zero_centered_weight = zero_centered_weight
+        y = y.reshape(x_shape_og)
+        y1 = y1.reshape(x_shape_og) if y1 is not None else None
+        residual_out = residual_out.reshape(x_shape_og) if residual_out is not None else None
+        dropout_mask = dropout_mask.reshape(x_shape_og) if dropout_mask is not None else None
+        dropout_mask1 = dropout_mask1.reshape(x_shape_og) if dropout_mask1 is not None else None
+        if not return_dropout_mask:
+            if weight1 is None:
+                return y if not prenorm else (y, residual_out)
+            else:
+                return (y, y1) if not prenorm else (y, y1, residual_out)
+        else:
+            if weight1 is None:
+                return (
+                    (y, dropout_mask, dropout_mask1)
+                    if not prenorm
+                    else (y, residual_out, dropout_mask, dropout_mask1)
+                )
+            else:
+                return (
+                    (y, y1, dropout_mask, dropout_mask1)
+                    if not prenorm
+                    else (y, y1, residual_out, dropout_mask, dropout_mask1)
+                )
+
+    @staticmethod
+    def backward(ctx, dy, *args):
+        if ctx.zero_seq_length:
+            return (
+                torch.zeros(ctx.x_shape_og, dtype=dy.dtype, device=dy.device),
+                torch.zeros(ctx.weight_shape, dtype=ctx.weight_dtype, device=ctx.weight_device),
+                torch.zeros(ctx.bias_shape, dtype=ctx.bias_dtype, device=ctx.bias_device) if ctx.has_bias else None,
+                torch.zeros(ctx.x_shape_og, dtype=dy.dtype, device=dy.device) if ctx.has_residual else None,
+                torch.zeros(ctx.x_shape_og, dtype=dy.dtype, device=dy.device) if ctx.has_x1 and ctx.dropout_p > 0.0 else None,
+                torch.zeros(ctx.weight1_shape, dtype=ctx.weight1_dtype, device=ctx.weight1_device) if ctx.has_weight1 else None,
+                torch.zeros(ctx.bias1_shape, dtype=ctx.bias1_dtype, device=ctx.bias1_device) if ctx.has_bias1 else None,
+                None,
+                None,
+                None,
+                None,
+                None,
+                None,
+                None,
+                None,
+                None,
+                None,
+            )
+        
+        x, weight, bias, weight1, bias1, rowscale, seeds, mean, rstd = ctx.saved_tensors
+        dy = dy.reshape(-1, dy.shape[-1])
+        if dy.stride(-1) != 1:
+            dy = dy.contiguous()
+        assert dy.shape == x.shape
+        if weight1 is not None:
+            dy1, args = args[0], args[1:]
+            dy1 = dy1.reshape(-1, dy1.shape[-1])
+            if dy1.stride(-1) != 1:
+                dy1 = dy1.contiguous()
+            assert dy1.shape == x.shape
+        else:
+            dy1 = None
+        if ctx.prenorm:
+            dresidual = args[0]
+            dresidual = dresidual.reshape(-1, dresidual.shape[-1])
+            if dresidual.stride(-1) != 1:
+                dresidual = dresidual.contiguous()
+            assert dresidual.shape == x.shape
+        else:
+            dresidual = None
+        
+        dx, dw, db, dresidual_in, dx1, dw1, db1 = _layer_norm_bwd(
+            dy,
+            x,
+            weight,
+            bias,
+            ctx.eps,
+            mean,
+            rstd,
+            dresidual,
+            dy1,
+            weight1,
+            bias1,
+            seeds,
+            ctx.dropout_p,
+            rowscale,
+            ctx.has_residual,
+            ctx.has_x1,
+            ctx.zero_centered_weight,
+            ctx.is_rms_norm,
+            x_dtype=ctx.x_dtype,
+        )
+        return (
+            dx.reshape(ctx.x_shape_og),
+            dw,
+            db,
+            dresidual_in.reshape(ctx.x_shape_og) if ctx.has_residual else None,
+            dx1.reshape(ctx.x_shape_og) if dx1 is not None else None,
+            dw1,
+            db1,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+        )
+
+
+def layer_norm_fn(
+    x,
+    weight,
+    bias,
+    residual=None,
+    x1=None,
+    weight1=None,
+    bias1=None,
+    eps=1e-6,
+    dropout_p=0.0,
+    rowscale=None,
+    prenorm=False,
+    residual_in_fp32=False,
+    zero_centered_weight=False,
+    is_rms_norm=False,
+    return_dropout_mask=False,
+    out=None,
+    residual_out=None
+):
+    return LayerNormFn.apply(
+        x,
+        weight,
+        bias,
+        residual,
+        x1,
+        weight1,
+        bias1,
+        eps,
+        dropout_p,
+        rowscale,
+        prenorm,
+        residual_in_fp32,
+        zero_centered_weight,
+        is_rms_norm,
+        return_dropout_mask,
+        out,
+        residual_out
+    )
+
+
+def rms_norm_fn(
+    x,
+    weight,
+    bias,
+    residual=None,
+    x1=None,
+    weight1=None,
+    bias1=None,
+    eps=1e-6,
+    dropout_p=0.0,
+    rowscale=None,
+    prenorm=False,
+    residual_in_fp32=False,
+    zero_centered_weight=False,
+    return_dropout_mask=False,
+    out=None,
+    residual_out=None
+):
+    return LayerNormFn.apply(
+        x,
+        weight,
+        bias,
+        residual,
+        x1,
+        weight1,
+        bias1,
+        eps,
+        dropout_p,
+        rowscale,
+        prenorm,
+        residual_in_fp32,
+        zero_centered_weight,
+        True,
+        return_dropout_mask,
+        out,
+        residual_out
+    )
+
+
+class RMSNorm(torch.nn.Module):
+
+    def __init__(self, hidden_size, eps=1e-5, dropout_p=0.0, zero_centered_weight=False,
+                 device=None, dtype=None):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        self.eps = eps
+        if dropout_p > 0.0:
+            self.drop = torch.nn.Dropout(dropout_p)
+        else:
+            self.drop = None
+        self.zero_centered_weight = zero_centered_weight
+        self.weight = torch.nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
+        self.register_parameter("bias", None)
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        if not self.zero_centered_weight:
+            torch.nn.init.ones_(self.weight)
+        else:
+            torch.nn.init.zeros_(self.weight)
+
+    def forward(self, x, residual=None, prenorm=False, residual_in_fp32=False):
+        return rms_norm_fn(
+            x,
+            self.weight,
+            self.bias,
+            residual=residual,
+            eps=self.eps,
+            dropout_p=self.drop.p if self.drop is not None and self.training else 0.0,
+            prenorm=prenorm,
+            residual_in_fp32=residual_in_fp32,
+            zero_centered_weight=self.zero_centered_weight,
+        )
+
+
+class LayerNormLinearFn(torch.autograd.Function):
+    @staticmethod
+    @custom_fwd
+    def forward(
+        ctx,
+        x,
+        norm_weight,
+        norm_bias,
+        linear_weight,
+        linear_bias,
+        residual=None,
+        eps=1e-6,
+        prenorm=False,
+        residual_in_fp32=False,
+        is_rms_norm=False,
+    ):
+        x_shape_og = x.shape
+        # reshape input data into 2D tensor
+        x = x.reshape(-1, x.shape[-1])
+        if x.stride(-1) != 1:
+            x = x.contiguous()
+        if residual is not None:
+            assert residual.shape == x_shape_og
+            residual = residual.reshape(-1, residual.shape[-1])
+            if residual.stride(-1) != 1:
+                residual = residual.contiguous()
+        norm_weight = norm_weight.contiguous()
+        if norm_bias is not None:
+            norm_bias = norm_bias.contiguous()
+        residual_dtype = (
+            residual.dtype
+            if residual is not None
+            else (torch.float32 if residual_in_fp32 else None)
+        )
+        y, _, mean, rstd, residual_out, *rest = _layer_norm_fwd(
+            x,
+            norm_weight,
+            norm_bias,
+            eps,
+            residual,
+            out_dtype=None if not torch.is_autocast_enabled() else torch.get_autocast_dtype("cuda"),
+            residual_dtype=residual_dtype,
+            is_rms_norm=is_rms_norm,
+        )
+        y = y.reshape(x_shape_og)
+        dtype = torch.get_autocast_dtype("cuda") if torch.is_autocast_enabled() else y.dtype
+        linear_weight = linear_weight.to(dtype)
+        linear_bias = linear_bias.to(dtype) if linear_bias is not None else None
+        out = F.linear(y.to(linear_weight.dtype), linear_weight, linear_bias)
+        # We don't store y, will be recomputed in the backward pass to save memory
+        ctx.save_for_backward(residual_out, norm_weight, norm_bias, linear_weight, mean, rstd)
+        ctx.x_shape_og = x_shape_og
+        ctx.eps = eps
+        ctx.is_rms_norm = is_rms_norm
+        ctx.has_residual = residual is not None
+        ctx.prenorm = prenorm
+        ctx.x_dtype = x.dtype
+        ctx.linear_bias_is_none = linear_bias is None
+        return out if not prenorm else (out, residual_out.reshape(x_shape_og))
+
+    @staticmethod
+    @custom_bwd
+    def backward(ctx, dout, *args):
+        x, norm_weight, norm_bias, linear_weight, mean, rstd = ctx.saved_tensors
+        dout = dout.reshape(-1, dout.shape[-1])
+        dy = F.linear(dout, linear_weight.t())
+        dlinear_bias = None if ctx.linear_bias_is_none else dout.sum(0)
+        if dy.stride(-1) != 1:
+            dy = dy.contiguous()
+        assert dy.shape == x.shape
+        if ctx.prenorm:
+            dresidual = args[0]
+            dresidual = dresidual.reshape(-1, dresidual.shape[-1])
+            if dresidual.stride(-1) != 1:
+                dresidual = dresidual.contiguous()
+            assert dresidual.shape == x.shape
+        else:
+            dresidual = None
+        dx, dnorm_weight, dnorm_bias, dresidual_in, _, _, _, y = _layer_norm_bwd(
+            dy,
+            x,
+            norm_weight,
+            norm_bias,
+            ctx.eps,
+            mean,
+            rstd,
+            dresidual=dresidual,
+            has_residual=ctx.has_residual,
+            is_rms_norm=ctx.is_rms_norm,
+            x_dtype=ctx.x_dtype,
+            recompute_output=True,
+        )
+        dlinear_weight = torch.einsum("bo,bi->oi", dout, y)
+        return (
+            dx.reshape(ctx.x_shape_og),
+            dnorm_weight,
+            dnorm_bias,
+            dlinear_weight,
+            dlinear_bias,
+            dresidual_in.reshape(ctx.x_shape_og) if ctx.has_residual else None,
+            None,
+            None,
+            None,
+            None,
+        )
+
+
+def layer_norm_linear_fn(
+    x,
+    norm_weight,
+    norm_bias,
+    linear_weight,
+    linear_bias,
+    residual=None,
+    eps=1e-6,
+    prenorm=False,
+    residual_in_fp32=False,
+    is_rms_norm=False,
+):
+    return LayerNormLinearFn.apply(
+        x,
+        norm_weight,
+        norm_bias,
+        linear_weight,
+        linear_bias,
+        residual,
+        eps,
+        prenorm,
+        residual_in_fp32,
+        is_rms_norm,
+    )

omnigen2/optim/scheduler/cosine_lr.py

@@ -0,0 +1,118 @@
+""" Cosine Scheduler
+
+Cosine LR schedule with warmup, cycle/restarts, noise, k-decay.
+
+Hacked together by / Copyright 2021 Ross Wightman
+"""
+import logging
+import math
+import torch
+from typing import List
+
+from .scheduler import Scheduler
+
+
+_logger = logging.getLogger(__name__)
+
+
+class CosineLRScheduler(Scheduler):
+    """
+    Cosine decay with restarts.
+    This is described in the paper https://arxiv.org/abs/1608.03983.
+
+    Inspiration from
+    https://github.com/allenai/allennlp/blob/master/allennlp/training/learning_rate_schedulers/cosine.py
+
+    k-decay option based on `k-decay: A New Method For Learning Rate Schedule` - https://arxiv.org/abs/2004.05909
+    """
+
+    def __init__(
+            self,
+            optimizer: torch.optim.Optimizer,
+            t_initial: int,
+            lr_min: float = 0.,
+            cycle_mul: float = 1.,
+            cycle_decay: float = 1.,
+            cycle_limit: int = 1,
+            warmup_t=0,
+            warmup_lr_init=0,
+            warmup_prefix=False,
+            t_in_epochs=True,
+            noise_range_t=None,
+            noise_pct=0.67,
+            noise_std=1.0,
+            noise_seed=42,
+            k_decay=1.0,
+            initialize=True,
+    ) -> None:
+        super().__init__(
+            optimizer,
+            param_group_field="lr",
+            t_in_epochs=t_in_epochs,
+            noise_range_t=noise_range_t,
+            noise_pct=noise_pct,
+            noise_std=noise_std,
+            noise_seed=noise_seed,
+            initialize=initialize,
+        )
+
+        assert t_initial > 0
+        assert lr_min >= 0
+        if t_initial == 1 and cycle_mul == 1 and cycle_decay == 1:
+            _logger.warning(
+                "Cosine annealing scheduler will have no effect on the learning "
+                "rate since t_initial = t_mul = eta_mul = 1.")
+        self.t_initial = t_initial
+        self.lr_min = lr_min
+        self.cycle_mul = cycle_mul
+        self.cycle_decay = cycle_decay
+        self.cycle_limit = cycle_limit
+        self.warmup_t = warmup_t
+        self.warmup_lr_init = warmup_lr_init
+        self.warmup_prefix = warmup_prefix
+        self.k_decay = k_decay
+        if self.warmup_t:
+            self.warmup_steps = [(v - warmup_lr_init) / self.warmup_t for v in self.base_values]
+            super().update_groups(self.warmup_lr_init)
+        else:
+            self.warmup_steps = [1 for _ in self.base_values]
+
+        self._step_count = 0 # no use
+
+    def _get_lr(self, t: int) -> List[float]:
+
+        if t < self.warmup_t:
+            lrs = [self.warmup_lr_init + t * s for s in self.warmup_steps]
+        else:
+            if self.warmup_prefix:
+                t = t - self.warmup_t
+
+            if self.cycle_mul != 1:
+                i = math.floor(math.log(1 - t / self.t_initial * (1 - self.cycle_mul), self.cycle_mul))
+                t_i = self.cycle_mul ** i * self.t_initial
+                t_curr = t - (1 - self.cycle_mul ** i) / (1 - self.cycle_mul) * self.t_initial
+            else:
+                i = t // self.t_initial
+                t_i = self.t_initial
+                t_curr = t - (self.t_initial * i)
+
+            gamma = self.cycle_decay ** i
+            lr_max_values = [v * gamma for v in self.base_values]
+            k = self.k_decay
+
+            if i < self.cycle_limit:
+                lrs = [
+                    self.lr_min + 0.5 * (lr_max - self.lr_min) * (1 + math.cos(math.pi * t_curr ** k / t_i ** k))
+                    for lr_max in lr_max_values
+                ]
+            else:
+                lrs = [self.lr_min for _ in self.base_values]
+
+        return lrs
+
+    def get_cycle_length(self, cycles=0):
+        cycles = max(1, cycles or self.cycle_limit)
+        if self.cycle_mul == 1.0:
+            return self.t_initial * cycles
+        else:
+            return int(math.floor(-self.t_initial * (self.cycle_mul ** cycles - 1) / (1 - self.cycle_mul)))

omnigen2/optim/scheduler/scheduler.py

@@ -0,0 +1,131 @@
+import abc
+from abc import ABC
+from typing import Any, Dict, List, Optional
+
+import torch
+
+
+class Scheduler(ABC):
+    """ Parameter Scheduler Base Class
+    A scheduler base class that can be used to schedule any optimizer parameter groups.
+
+    Unlike the builtin PyTorch schedulers, this is intended to be consistently called
+    * At the END of each epoch, before incrementing the epoch count, to calculate next epoch's value
+    * At the END of each optimizer update, after incrementing the update count, to calculate next update's value
+
+    The schedulers built on this should try to remain as stateless as possible (for simplicity).
+
+    This family of schedulers is attempting to avoid the confusion of the meaning of 'last_epoch'
+    and -1 values for special behaviour. All epoch and update counts must be tracked in the training
+    code and explicitly passed in to the schedulers on the corresponding step or step_update call.
+
+    Based on ideas from:
+     * https://github.com/pytorch/fairseq/tree/master/fairseq/optim/lr_scheduler
+     * https://github.com/allenai/allennlp/tree/master/allennlp/training/learning_rate_schedulers
+    """
+
+    def __init__(
+            self,
+            optimizer: torch.optim.Optimizer,
+            param_group_field: str,
+            t_in_epochs: bool = True,
+            noise_range_t=None,
+            noise_type='normal',
+            noise_pct=0.67,
+            noise_std=1.0,
+            noise_seed=None,
+            initialize: bool = True,
+    ) -> None:
+        self.optimizer = optimizer
+        self.param_group_field = param_group_field
+        self._initial_param_group_field = f"initial_{param_group_field}"
+        if initialize:
+            for i, group in enumerate(self.optimizer.param_groups):
+                if param_group_field not in group:
+                    raise KeyError(f"{param_group_field} missing from param_groups[{i}]")
+                group.setdefault(self._initial_param_group_field, group[param_group_field])
+        else:
+            for i, group in enumerate(self.optimizer.param_groups):
+                if self._initial_param_group_field not in group:
+                    raise KeyError(f"{self._initial_param_group_field} missing from param_groups[{i}]")
+        self.base_values = [group[self._initial_param_group_field] for group in self.optimizer.param_groups]
+        self.metric = None  # any point to having this for all?
+        self.t_in_epochs = t_in_epochs
+        self.noise_range_t = noise_range_t
+        self.noise_pct = noise_pct
+        self.noise_type = noise_type
+        self.noise_std = noise_std
+        self.noise_seed = noise_seed if noise_seed is not None else 42
+        self.update_groups(self.base_values)
+
+    def state_dict(self) -> Dict[str, Any]:
+        return {key: value for key, value in self.__dict__.items() if key != 'optimizer'}
+
+    def load_state_dict(self, state_dict: Dict[str, Any]) -> None:
+        self.__dict__.update(state_dict)
+
+    def get_last_lr(self):
+        """ Return last computed learning rate by current scheduler.
+        """
+        return self._last_lr
+    
+    @abc.abstractmethod
+    def _get_lr(self, t: int) -> List[float]:
+        pass
+
+    def _get_values(self, t: int, on_epoch: bool = True) -> Optional[List[float]]:
+        return self._get_lr(t)
+
+    def step(self, epoch: int, metric: float = None) -> None:
+        self.metric = metric
+        values = self._get_values(epoch, on_epoch=True)
+        if values is not None:
+            values = self._add_noise(values, epoch)
+            self.update_groups(values)
+
+    # def step_update(self, num_updates: int, metric: float = None):
+    #     self.metric = metric
+    #     values = self._get_values(num_updates, on_epoch=False)
+    #     if values is not None:
+    #         values = self._add_noise(values, num_updates)
+    #         self.update_groups(values)
+
+    def update_groups(self, values):
+        if not isinstance(values, (list, tuple)):
+            values = [values] * len(self.optimizer.param_groups)
+        for param_group, value in zip(self.optimizer.param_groups, values):
+            if 'lr_scale' in param_group:
+                param_group[self.param_group_field] = value * param_group['lr_scale']
+            else:
+                param_group[self.param_group_field] = value
+
+        self._last_lr = [group[self.param_group_field] for group in self.optimizer.param_groups]
+
+    def _add_noise(self, lrs, t):
+        if self._is_apply_noise(t):
+            noise = self._calculate_noise(t)
+            lrs = [v + v * noise for v in lrs]
+        return lrs
+
+    def _is_apply_noise(self, t) -> bool:
+        """Return True if scheduler in noise range."""
+        apply_noise = False
+        if self.noise_range_t is not None:
+            if isinstance(self.noise_range_t, (list, tuple)):
+                apply_noise = self.noise_range_t[0] <= t < self.noise_range_t[1]
+            else:
+                apply_noise = t >= self.noise_range_t
+        return apply_noise
+
+    def _calculate_noise(self, t) -> float:
+        g = torch.Generator()
+        g.manual_seed(self.noise_seed + t)
+        if self.noise_type == 'normal':
+            while True:
+                # resample if noise out of percent limit, brute force but shouldn't spin much
+                noise = torch.randn(1, generator=g).item()
+                if abs(noise) < self.noise_pct:
+                    return noise
+        else:
+            noise = 2 * (torch.rand(1, generator=g).item() - 0.5) * self.noise_pct
+        return noise

omnigen2/optim/scheduler/step_lr.py

@@ -0,0 +1,63 @@
+""" Step Scheduler
+
+Basic step LR schedule with warmup, noise.
+
+Hacked together by / Copyright 2020 Ross Wightman
+"""
+import math
+import torch
+from typing import List
+
+
+from .scheduler import Scheduler
+
+
+class StepLRScheduler(Scheduler):
+    """
+    """
+
+    def __init__(
+            self,
+            optimizer: torch.optim.Optimizer,
+            decay_t: float,
+            decay_rate: float = 1.,
+            warmup_t=0,
+            warmup_lr_init=0,
+            warmup_prefix=True,
+            t_in_epochs=True,
+            noise_range_t=None,
+            noise_pct=0.67,
+            noise_std=1.0,
+            noise_seed=42,
+            initialize=True,
+    ) -> None:
+        super().__init__(
+            optimizer,
+            param_group_field="lr",
+            t_in_epochs=t_in_epochs,
+            noise_range_t=noise_range_t,
+            noise_pct=noise_pct,
+            noise_std=noise_std,
+            noise_seed=noise_seed,
+            initialize=initialize,
+        )
+
+        self.decay_t = decay_t
+        self.decay_rate = decay_rate
+        self.warmup_t = warmup_t
+        self.warmup_lr_init = warmup_lr_init
+        self.warmup_prefix = warmup_prefix
+        if self.warmup_t:
+            self.warmup_steps = [(v - warmup_lr_init) / self.warmup_t for v in self.base_values]
+            super().update_groups(self.warmup_lr_init)
+        else:
+            self.warmup_steps = [1 for _ in self.base_values]
+
+    def _get_lr(self, t: int) -> List[float]:
+        if t < self.warmup_t:
+            lrs = [self.warmup_lr_init + t * s for s in self.warmup_steps]
+        else:
+            if self.warmup_prefix:
+                t = t - self.warmup_t
+            lrs = [v * (self.decay_rate ** (t // self.decay_t)) for v in self.base_values]
+        return lrs

omnigen2/pipelines/image_processor.py

@@ -0,0 +1,266 @@
+# Copyright 2024 The HuggingFace 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor, is_valid_image_imagelist
+from diffusers.configuration_utils import register_to_config
+
+class OmniGen2ImageProcessor(VaeImageProcessor):
+    """
+    Image processor for PixArt image resize and crop.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept
+            `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method.
+        vae_scale_factor (`int`, *optional*, defaults to `8`):
+            VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
+        resample (`str`, *optional*, defaults to `lanczos`):
+            Resampling filter to use when resizing the image.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image to [-1,1].
+        do_binarize (`bool`, *optional*, defaults to `False`):
+            Whether to binarize the image to 0/1.
+        do_convert_rgb (`bool`, *optional*, defaults to be `False`):
+            Whether to convert the images to RGB format.
+        do_convert_grayscale (`bool`, *optional*, defaults to be `False`):
+            Whether to convert the images to grayscale format.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        do_resize: bool = True,
+        vae_scale_factor: int = 16,
+        resample: str = "lanczos",
+        max_pixels: Optional[int] = None,
+        max_side_length: Optional[int] = None,
+        do_normalize: bool = True,
+        do_binarize: bool = False,
+        do_convert_grayscale: bool = False,
+    ):
+        super().__init__(
+            do_resize=do_resize,
+            vae_scale_factor=vae_scale_factor,
+            resample=resample,
+            do_normalize=do_normalize,
+            do_binarize=do_binarize,
+            do_convert_grayscale=do_convert_grayscale,
+        )
+
+        self.max_pixels = max_pixels
+        self.max_side_length = max_side_length
+
+    def get_new_height_width(
+        self,
+        image: Union[PIL.Image.Image, np.ndarray, torch.Tensor],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        max_pixels: Optional[int] = None,
+        max_side_length: Optional[int] = None,
+    ) -> Tuple[int, int]:
+        r"""
+        Returns the height and width of the image, downscaled to the next integer multiple of `vae_scale_factor`.
+
+        Args:
+            image (`Union[PIL.Image.Image, np.ndarray, torch.Tensor]`):
+                The image input, which can be a PIL image, NumPy array, or PyTorch tensor. If it is a NumPy array, it
+                should have shape `[batch, height, width]` or `[batch, height, width, channels]`. If it is a PyTorch
+                tensor, it should have shape `[batch, channels, height, width]`.
+            height (`Optional[int]`, *optional*, defaults to `None`):
+                The height of the preprocessed image. If `None`, the height of the `image` input will be used.
+            width (`Optional[int]`, *optional*, defaults to `None`):
+                The width of the preprocessed image. If `None`, the width of the `image` input will be used.
+
+        Returns:
+            `Tuple[int, int]`:
+                A tuple containing the height and width, both resized to the nearest integer multiple of
+                `vae_scale_factor`.
+        """
+
+        if height is None:
+            if isinstance(image, PIL.Image.Image):
+                height = image.height
+            elif isinstance(image, torch.Tensor):
+                height = image.shape[2]
+            else:
+                height = image.shape[1]
+
+        if width is None:
+            if isinstance(image, PIL.Image.Image):
+                width = image.width
+            elif isinstance(image, torch.Tensor):
+                width = image.shape[3]
+            else:
+                width = image.shape[2]
+        
+        if max_side_length is None:
+            max_side_length = self.max_side_length
+        
+        if max_pixels is None:
+            max_pixels = self.max_pixels
+        
+        ratio = 1.0
+        if max_side_length is not None:
+            if height > width:
+                max_side_length_ratio = max_side_length / height
+            else:
+                max_side_length_ratio = max_side_length / width
+        
+        cur_pixels = height * width
+        max_pixels_ratio = (max_pixels / cur_pixels) ** 0.5
+        ratio = min(max_pixels_ratio, max_side_length_ratio, 1.0) # do not upscale input image
+
+        new_height, new_width = int(height * ratio) // self.config.vae_scale_factor * self.config.vae_scale_factor, int(width * ratio) // self.config.vae_scale_factor * self.config.vae_scale_factor
+        return new_height, new_width
+
+    def preprocess(
+        self,
+        image: PipelineImageInput,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        max_pixels: Optional[int] = None,
+        max_side_length: Optional[int] = None,
+        resize_mode: str = "default",  # "default", "fill", "crop"
+        crops_coords: Optional[Tuple[int, int, int, int]] = None,
+    ) -> torch.Tensor:
+        """
+        Preprocess the image input.
+
+        Args:
+            image (`PipelineImageInput`):
+                The image input, accepted formats are PIL images, NumPy arrays, PyTorch tensors; Also accept list of
+                supported formats.
+            height (`int`, *optional*):
+                The height in preprocessed image. If `None`, will use the `get_default_height_width()` to get default
+                height.
+            width (`int`, *optional*):
+                The width in preprocessed. If `None`, will use get_default_height_width()` to get the default width.
+            resize_mode (`str`, *optional*, defaults to `default`):
+                The resize mode, can be one of `default` or `fill`. If `default`, will resize the image to fit within
+                the specified width and height, and it may not maintaining the original aspect ratio. If `fill`, will
+                resize the image to fit within the specified width and height, maintaining the aspect ratio, and then
+                center the image within the dimensions, filling empty with data from image. If `crop`, will resize the
+                image to fit within the specified width and height, maintaining the aspect ratio, and then center the
+                image within the dimensions, cropping the excess. Note that resize_mode `fill` and `crop` are only
+                supported for PIL image input.
+            crops_coords (`List[Tuple[int, int, int, int]]`, *optional*, defaults to `None`):
+                The crop coordinates for each image in the batch. If `None`, will not crop the image.
+
+        Returns:
+            `torch.Tensor`:
+                The preprocessed image.
+        """
+        supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor)
+
+        # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image
+        if self.config.do_convert_grayscale and isinstance(image, (torch.Tensor, np.ndarray)) and image.ndim == 3:
+            if isinstance(image, torch.Tensor):
+                # if image is a pytorch tensor could have 2 possible shapes:
+                #    1. batch x height x width: we should insert the channel dimension at position 1
+                #    2. channel x height x width: we should insert batch dimension at position 0,
+                #       however, since both channel and batch dimension has same size 1, it is same to insert at position 1
+                #    for simplicity, we insert a dimension of size 1 at position 1 for both cases
+                image = image.unsqueeze(1)
+            else:
+                # if it is a numpy array, it could have 2 possible shapes:
+                #   1. batch x height x width: insert channel dimension on last position
+                #   2. height x width x channel: insert batch dimension on first position
+                if image.shape[-1] == 1:
+                    image = np.expand_dims(image, axis=0)
+                else:
+                    image = np.expand_dims(image, axis=-1)
+
+        if isinstance(image, list) and isinstance(image[0], np.ndarray) and image[0].ndim == 4:
+            warnings.warn(
+                "Passing `image` as a list of 4d np.ndarray is deprecated."
+                "Please concatenate the list along the batch dimension and pass it as a single 4d np.ndarray",
+                FutureWarning,
+            )
+            image = np.concatenate(image, axis=0)
+        if isinstance(image, list) and isinstance(image[0], torch.Tensor) and image[0].ndim == 4:
+            warnings.warn(
+                "Passing `image` as a list of 4d torch.Tensor is deprecated."
+                "Please concatenate the list along the batch dimension and pass it as a single 4d torch.Tensor",
+                FutureWarning,
+            )
+            image = torch.cat(image, axis=0)
+
+        if not is_valid_image_imagelist(image):
+            raise ValueError(
+                f"Input is in incorrect format. Currently, we only support {', '.join(str(x) for x in supported_formats)}"
+            )
+        if not isinstance(image, list):
+            image = [image]
+
+        if isinstance(image[0], PIL.Image.Image):
+            if crops_coords is not None:
+                image = [i.crop(crops_coords) for i in image]
+            if self.config.do_resize:
+                height, width = self.get_new_height_width(image[0], height, width, max_pixels, max_side_length)
+                image = [self.resize(i, height, width, resize_mode=resize_mode) for i in image]
+            if self.config.do_convert_rgb:
+                image = [self.convert_to_rgb(i) for i in image]
+            elif self.config.do_convert_grayscale:
+                image = [self.convert_to_grayscale(i) for i in image]
+            image = self.pil_to_numpy(image)  # to np
+            image = self.numpy_to_pt(image)  # to pt
+
+        elif isinstance(image[0], np.ndarray):
+            image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
+
+            image = self.numpy_to_pt(image)
+
+            height, width = self.get_new_height_width(image, height, width, max_pixels, max_side_length)
+            if self.config.do_resize:
+                image = self.resize(image, height, width)
+
+        elif isinstance(image[0], torch.Tensor):
+            image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
+
+            if self.config.do_convert_grayscale and image.ndim == 3:
+                image = image.unsqueeze(1)
+
+            channel = image.shape[1]
+            # don't need any preprocess if the image is latents
+            if channel == self.config.vae_latent_channels:
+                return image
+
+            height, width = self.get_new_height_width(image, height, width, max_pixels, max_side_length)
+            if self.config.do_resize:
+                image = self.resize(image, height, width)
+
+        # expected range [0,1], normalize to [-1,1]
+        do_normalize = self.config.do_normalize
+        if do_normalize and image.min() < 0:
+            warnings.warn(
+                "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] "
+                f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{image.min()},{image.max()}]",
+                FutureWarning,
+            )
+            do_normalize = False
+        if do_normalize:
+            image = self.normalize(image)
+
+        if self.config.do_binarize:
+            image = self.binarize(image)
+
+        return image

omnigen2/pipelines/lora_pipeline.py

@@ -0,0 +1,388 @@
+# Copyright 2024 The HuggingFace 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from typing import Callable, Dict, List, Optional, Union
+
+import torch
+from huggingface_hub.utils import validate_hf_hub_args
+
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    is_peft_available,
+    is_peft_version,
+    is_torch_version,
+    is_transformers_available,
+    is_transformers_version,
+    logging,
+)
+from diffusers.loaders.lora_base import (  # noqa
+    LoraBaseMixin,
+    _fetch_state_dict,
+)
+from diffusers.loaders.lora_conversion_utils import (
+    _convert_non_diffusers_lumina2_lora_to_diffusers,
+)
+
+
+_LOW_CPU_MEM_USAGE_DEFAULT_LORA = False
+if is_torch_version(">=", "1.9.0"):
+    if (
+        is_peft_available()
+        and is_peft_version(">=", "0.13.1")
+        and is_transformers_available()
+        and is_transformers_version(">", "4.45.2")
+    ):
+        _LOW_CPU_MEM_USAGE_DEFAULT_LORA = True
+
+
+logger = logging.get_logger(__name__)
+
+TRANSFORMER_NAME = "transformer"
+
+class OmniGen2LoraLoaderMixin(LoraBaseMixin):
+    r"""
+    Load LoRA layers into [`OmniGen2Transformer2DModel`]. Specific to [`OmniGen2Pipeline`].
+    """
+
+    _lora_loadable_modules = ["transformer"]
+    transformer_name = TRANSFORMER_NAME
+
+    @classmethod
+    @validate_hf_hub_args
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        Return state dict for lora weights and the network alphas.
+
+        <Tip warning={true}>
+
+        We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+
+        This function is experimental and might change in the future.
+
+        </Tip>
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # transformer and text encoder or both.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+
+        state_dict = _fetch_state_dict(
+            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
+            weight_name=weight_name,
+            use_safetensors=use_safetensors,
+            local_files_only=local_files_only,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            subfolder=subfolder,
+            user_agent=user_agent,
+            allow_pickle=allow_pickle,
+        )
+
+        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
+        if is_dora_scale_present:
+            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
+            logger.warning(warn_msg)
+            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
+
+        # conversion.
+        non_diffusers = any(k.startswith("diffusion_model.") for k in state_dict)
+        if non_diffusers:
+            state_dict = _convert_non_diffusers_lumina2_lora_to_diffusers(state_dict)
+
+        return state_dict
+
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
+    def load_lora_weights(
+        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
+        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
+        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
+        dict is loaded into `self.transformer`.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+            kwargs (`dict`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # if a dict is passed, copy it instead of modifying it inplace
+        if isinstance(pretrained_model_name_or_path_or_dict, dict):
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        self.load_lora_into_transformer(
+            state_dict,
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            adapter_name=adapter_name,
+            _pipeline=self,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->Lumina2Transformer2DModel
+    def load_lora_into_transformer(
+        cls, state_dict, transformer, adapter_name=None, _pipeline=None, low_cpu_mem_usage=False, hotswap: bool = False
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `transformer`.
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            transformer (`Lumina2Transformer2DModel`):
+                The Transformer model to load the LoRA layers into.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+            hotswap : (`bool`, *optional*)
+                Defaults to `False`. Whether to substitute an existing (LoRA) adapter with the newly loaded adapter
+                in-place. This means that, instead of loading an additional adapter, this will take the existing
+                adapter weights and replace them with the weights of the new adapter. This can be faster and more
+                memory efficient. However, the main advantage of hotswapping is that when the model is compiled with
+                torch.compile, loading the new adapter does not require recompilation of the model. When using
+                hotswapping, the passed `adapter_name` should be the name of an already loaded adapter.
+
+                If the new adapter and the old adapter have different ranks and/or LoRA alphas (i.e. scaling), you need
+                to call an additional method before loading the adapter:
+
+                ```py
+                pipeline = ...  # load diffusers pipeline
+                max_rank = ...  # the highest rank among all LoRAs that you want to load
+                # call *before* compiling and loading the LoRA adapter
+                pipeline.enable_lora_hotswap(target_rank=max_rank)
+                pipeline.load_lora_weights(file_name)
+                # optionally compile the model now
+                ```
+
+                Note that hotswapping adapters of the text encoder is not yet supported. There are some further
+                limitations to this technique, which are documented here:
+                https://huggingface.co/docs/peft/main/en/package_reference/hotswap
+        """
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # Load the layers corresponding to transformer.
+        logger.info(f"Loading {cls.transformer_name}.")
+        transformer.load_lora_adapter(
+            state_dict,
+            network_alphas=None,
+            adapter_name=adapter_name,
+            _pipeline=_pipeline,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+            hotswap=hotswap,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the UNet and text encoder.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `transformer`.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+        """
+        state_dict = {}
+
+        if not transformer_lora_layers:
+            raise ValueError("You must pass `transformer_lora_layers`.")
+
+        if transformer_lora_layers:
+            state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))
+
+        # Save the model
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.fuse_lora
+    def fuse_lora(
+        self,
+        components: List[str] = ["transformer"],
+        lora_scale: float = 1.0,
+        safe_fusing: bool = False,
+        adapter_names: Optional[List[str]] = None,
+        **kwargs,
+    ):
+        r"""
+        Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
+            lora_scale (`float`, defaults to 1.0):
+                Controls how much to influence the outputs with the LoRA parameters.
+            safe_fusing (`bool`, defaults to `False`):
+                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
+            adapter_names (`List[str]`, *optional*):
+                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
+
+        Example:
+
+        ```py
+        from diffusers import DiffusionPipeline
+        import torch
+
+        pipeline = DiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ).to("cuda")
+        pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.fuse_lora(lora_scale=0.7)
+        ```
+        """
+        super().fuse_lora(
+            components=components,
+            lora_scale=lora_scale,
+            safe_fusing=safe_fusing,
+            adapter_names=adapter_names,
+            **kwargs,
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.SanaLoraLoaderMixin.unfuse_lora
+    def unfuse_lora(self, components: List[str] = ["transformer"], **kwargs):
+        r"""
+        Reverses the effect of
+        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
+            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
+        """
+        super().unfuse_lora(components=components, **kwargs)

omnigen2/pipelines/omnigen2/pipeline_omnigen2.py

@@ -0,0 +1,774 @@
+"""
+OmniGen2 Diffusion Pipeline
+
+Copyright 2025 BAAI, The OmniGen2 Team and The HuggingFace 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.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import math
+
+from PIL import Image
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from transformers import Qwen2_5_VLForConditionalGeneration
+
+from diffusers.models.autoencoders import AutoencoderKL
+from ...models.transformers import OmniGen2Transformer2DModel
+from ...models.transformers.repo import OmniGen2RotaryPosEmbed
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import (
+    is_torch_xla_available,
+    logging,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+
+from dataclasses import dataclass
+
+import PIL.Image
+
+from diffusers.utils import BaseOutput
+
+from omnigen2.pipelines.image_processor import OmniGen2ImageProcessor
+
+from omnigen2.utils.teacache_util import TeaCacheParams
+
+from ..lora_pipeline import OmniGen2LoraLoaderMixin
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+from ...cache_functions import cache_init 
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+@dataclass
+class FMPipelineOutput(BaseOutput):
+    """
+    Output class for OmniGen2 pipeline.
+
+    Args:
+        images (Union[List[PIL.Image.Image], np.ndarray]): 
+            List of denoised PIL images of length `batch_size` or numpy array of shape 
+            `(batch_size, height, width, num_channels)`. Contains the generated images.
+    """
+    images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class OmniGen2Pipeline(DiffusionPipeline, OmniGen2LoraLoaderMixin):
+    """
+    Pipeline for text-to-image generation using OmniGen2.
+
+    This pipeline implements a text-to-image generation model that uses:
+    - Qwen2.5-VL for text encoding
+    - A custom transformer architecture for image generation
+    - VAE for image encoding/decoding
+    - FlowMatchEulerDiscreteScheduler for noise scheduling
+
+    Args:
+        transformer (OmniGen2Transformer2DModel): The transformer model for image generation.
+        vae (AutoencoderKL): The VAE model for image encoding/decoding.
+        scheduler (FlowMatchEulerDiscreteScheduler): The scheduler for noise scheduling.
+        text_encoder (Qwen2_5_VLModel): The text encoder model.
+        tokenizer (Union[Qwen2Tokenizer, Qwen2TokenizerFast]): The tokenizer for text processing.
+    """
+
+    model_cpu_offload_seq = "mllm->transformer->vae"
+
+    def __init__(
+        self,
+        transformer: OmniGen2Transformer2DModel,
+        vae: AutoencoderKL,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        mllm: Qwen2_5_VLForConditionalGeneration,
+        processor,
+    ) -> None:
+        """
+        Initialize the OmniGen2 pipeline.
+
+        Args:
+            transformer: The transformer model for image generation.
+            vae: The VAE model for image encoding/decoding.
+            scheduler: The scheduler for noise scheduling.
+            text_encoder: The text encoder model.
+            tokenizer: The tokenizer for text processing.
+        """
+        super().__init__()
+
+        self.register_modules(
+            transformer=transformer,
+            vae=vae,
+            scheduler=scheduler,
+            mllm=mllm,
+            processor=processor
+        )
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.image_processor = OmniGen2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2, do_resize=True)
+        self.default_sample_size = 128
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        generator: Optional[torch.Generator],
+        latents: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        """
+        Prepare the initial latents for the diffusion process.
+
+        Args:
+            batch_size: The number of images to generate.
+            num_channels_latents: The number of channels in the latent space.
+            height: The height of the generated image.
+            width: The width of the generated image.
+            dtype: The data type of the latents.
+            device: The device to place the latents on.
+            generator: The random number generator to use.
+            latents: Optional pre-computed latents to use instead of random initialization.
+
+        Returns:
+            torch.FloatTensor: The prepared latents tensor.
+        """
+        height = int(height) // self.vae_scale_factor
+        width = int(width) // self.vae_scale_factor
+
+        shape = (batch_size, num_channels_latents, height, width)
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+        return latents
+
+    def encode_vae(self, img: torch.FloatTensor) -> torch.FloatTensor:
+        """
+        Encode an image into the VAE latent space.
+
+        Args:
+            img: The input image tensor to encode.
+
+        Returns:
+            torch.FloatTensor: The encoded latent representation.
+        """
+        z0 = self.vae.encode(img.to(dtype=self.vae.dtype)).latent_dist.sample()
+        if self.vae.config.shift_factor is not None:
+            z0 = z0 - self.vae.config.shift_factor
+        if self.vae.config.scaling_factor is not None:
+            z0 = z0 * self.vae.config.scaling_factor
+        z0 = z0.to(dtype=self.vae.dtype)
+        return z0
+
+    def prepare_image(
+        self,
+        images: Union[List[PIL.Image.Image], PIL.Image.Image],
+        batch_size: int,
+        num_images_per_prompt: int,
+        max_pixels: int,
+        max_side_length: int,
+        device: torch.device,
+        dtype: torch.dtype,
+    ) -> List[Optional[torch.FloatTensor]]:
+        """
+        Prepare input images for processing by encoding them into the VAE latent space.
+
+        Args:
+            images: Single image or list of images to process.
+            batch_size: The number of images to generate per prompt.
+            num_images_per_prompt: The number of images to generate for each prompt.
+            device: The device to place the encoded latents on.
+            dtype: The data type of the encoded latents.
+
+        Returns:
+            List[Optional[torch.FloatTensor]]: List of encoded latent representations for each image.
+        """
+        if batch_size == 1:
+            images = [images]
+        latents = []
+        for i, img in enumerate(images):
+            if img is not None and len(img) > 0:
+                ref_latents = []
+                for j, img_j in enumerate(img):
+                    img_j = self.image_processor.preprocess(img_j, max_pixels=max_pixels, max_side_length=max_side_length)
+                    ref_latents.append(self.encode_vae(img_j.to(device=device)).squeeze(0))
+            else:
+                ref_latents = None
+            for _ in range(num_images_per_prompt):
+                latents.append(ref_latents)
+
+        return latents
+    
+    def _get_qwen2_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        max_sequence_length: int = 256,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Get prompt embeddings from the Qwen2 text encoder.
+
+        Args:
+            prompt: The prompt or list of prompts to encode.
+            device: The device to place the embeddings on. If None, uses the pipeline's device.
+            max_sequence_length: Maximum sequence length for tokenization.
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]: A tuple containing:
+                - The prompt embeddings tensor
+                - The attention mask tensor
+
+        Raises:
+            Warning: If the input text is truncated due to sequence length limitations.
+        """
+        device = device or self._execution_device
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        # text_inputs = self.processor.tokenizer(
+        #     prompt,
+        #     padding="max_length",
+        #     max_length=max_sequence_length,
+        #     truncation=True,
+        #     return_tensors="pt",
+        # )
+        text_inputs = self.processor.tokenizer(
+            prompt,
+            padding="longest",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids.to(device)
+        untruncated_ids = self.processor.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids.to(device)
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.processor.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because Gemma can only handle sequences up to"
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_attention_mask = text_inputs.attention_mask.to(device)
+        prompt_embeds = self.mllm(
+            text_input_ids,
+            attention_mask=prompt_attention_mask,
+            output_hidden_states=True,
+        ).hidden_states[-1]
+
+        if self.mllm is not None:
+            dtype = self.mllm.dtype
+        elif self.transformer is not None:
+            dtype = self.transformer.dtype
+        else:
+            dtype = None
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        return prompt_embeds, prompt_attention_mask
+    
+    def _apply_chat_template(self, prompt: str):
+        prompt = [
+            {
+                "role": "system",
+                "content": "You are a helpful assistant that generates high-quality images based on user instructions.",
+            },
+            {"role": "user", "content": prompt},
+        ]
+        prompt = self.processor.tokenizer.apply_chat_template(prompt, tokenize=False, add_generation_prompt=False)
+        return prompt
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 256,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds`
+                instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For
+                Lumina-T2I, this should be "".
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                whether to use classifier free guidance or not
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                number of images that should be generated per prompt
+            device: (`torch.device`, *optional*):
+                torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. For Lumina-T2I, it's should be the embeddings of the "" string.
+            max_sequence_length (`int`, defaults to `256`):
+                Maximum sequence length to use for the prompt.
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt = [self._apply_chat_template(_prompt) for _prompt in prompt]
+
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_qwen2_prompt_embeds(
+                prompt=prompt,
+                device=device,
+                max_sequence_length=max_sequence_length
+            )
+
+        batch_size, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
+        prompt_attention_mask = prompt_attention_mask.view(batch_size * num_images_per_prompt, -1)
+
+        # Get negative embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt if negative_prompt is not None else ""
+
+            # Normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt = [self._apply_chat_template(_negative_prompt) for _negative_prompt in negative_prompt]
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                negative_prompt = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_qwen2_prompt_embeds(
+                prompt=negative_prompt,
+                device=device,
+                max_sequence_length=max_sequence_length,
+            )
+
+            batch_size, seq_len, _ = negative_prompt_embeds.shape
+            # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.view(
+                batch_size * num_images_per_prompt, -1
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+    
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+    
+    @property
+    def text_guidance_scale(self):
+        return self._text_guidance_scale
+    
+    @property
+    def image_guidance_scale(self):
+        return self._image_guidance_scale
+    
+    @property
+    def cfg_range(self):
+        return self._cfg_range
+    
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.LongTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.LongTensor] = None,
+        max_sequence_length: Optional[int] = None,
+        callback_on_step_end_tensor_inputs: Optional[List[str]] = None,
+        input_images: Optional[List[PIL.Image.Image]] = None,
+        num_images_per_prompt: int = 1,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        max_pixels: int = 1024 * 1024,
+        max_input_image_side_length: int = 1024,
+        align_res: bool = True,
+        num_inference_steps: int = 28,
+        text_guidance_scale: float = 4.0,
+        image_guidance_scale: float = 1.0,
+        cfg_range: Tuple[float, float] = (0.0, 1.0),
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        timesteps: List[int] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        verbose: bool = False,
+        step_func=None,
+    ):
+
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        self._text_guidance_scale = text_guidance_scale
+        self._image_guidance_scale = image_guidance_scale
+        self._cfg_range = cfg_range
+        self._attention_kwargs = attention_kwargs
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt,
+            self.text_guidance_scale > 1.0,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+        )
+
+        dtype = self.vae.dtype
+        # 3. Prepare control image
+        ref_latents = self.prepare_image(
+            images=input_images,
+            batch_size=batch_size,
+            num_images_per_prompt=num_images_per_prompt,
+            max_pixels=max_pixels,
+            max_side_length=max_input_image_side_length,
+            device=device,
+            dtype=dtype,
+        )
+
+        if input_images is None:
+            input_images = []
+        
+        if len(input_images) == 1 and align_res:
+            width, height = ref_latents[0][0].shape[-1] * self.vae_scale_factor, ref_latents[0][0].shape[-2] * self.vae_scale_factor
+            ori_width, ori_height = width, height
+        else:
+            ori_width, ori_height = width, height
+
+            cur_pixels = height * width
+            ratio = (max_pixels / cur_pixels) ** 0.5
+            ratio = min(ratio, 1.0)
+
+            height, width = int(height * ratio) // 16 * 16, int(width * ratio) // 16 * 16
+        
+        if len(input_images) == 0:
+            self._image_guidance_scale = 1
+
+        # 4. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            latent_channels,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        freqs_cis = OmniGen2RotaryPosEmbed.get_freqs_cis(
+            self.transformer.config.axes_dim_rope,
+            self.transformer.config.axes_lens,
+            theta=10000,
+        )
+        
+        image = self.processing(
+            latents=latents,
+            ref_latents=ref_latents,
+            prompt_embeds=prompt_embeds,
+            freqs_cis=freqs_cis,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            num_inference_steps=num_inference_steps,
+            timesteps=timesteps,
+            device=device,
+            dtype=dtype,
+            verbose=verbose,
+            step_func=step_func,
+        )
+
+        image = F.interpolate(image, size=(ori_height, ori_width), mode='bilinear')
+
+        image = self.image_processor.postprocess(image, output_type=output_type)
+        
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return image
+        else:
+            return FMPipelineOutput(images=image)
+
+    def processing(
+        self,
+        latents,
+        ref_latents,
+        prompt_embeds,
+        freqs_cis,
+        negative_prompt_embeds,
+        prompt_attention_mask,
+        negative_prompt_attention_mask,
+        num_inference_steps,
+        timesteps,
+        device,
+        dtype,
+        verbose,
+        step_func=None
+    ):
+        batch_size = latents.shape[0]
+
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            num_tokens=latents.shape[-2] * latents.shape[-1]
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        enable_taylorseer = getattr(self, "enable_taylorseer", False)
+        if enable_taylorseer:
+            model_pred_cache_dic, model_pred_current = cache_init(self, num_inference_steps)
+            model_pred_ref_cache_dic, model_pred_ref_current = cache_init(self, num_inference_steps)
+            model_pred_uncond_cache_dic, model_pred_uncond_current = cache_init(self, num_inference_steps)
+            self.transformer.enable_taylorseer = True
+        elif self.transformer.enable_teacache:
+            # Use different TeaCacheParams for different conditions
+            teacache_params = TeaCacheParams()
+            teacache_params_uncond = TeaCacheParams()
+            teacache_params_ref = TeaCacheParams()
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if enable_taylorseer:
+                    self.transformer.cache_dic = model_pred_cache_dic
+                    self.transformer.current = model_pred_current
+                elif self.transformer.enable_teacache:
+                    teacache_params.is_first_or_last_step = i == 0 or i == len(timesteps) - 1
+                    self.transformer.teacache_params = teacache_params
+
+                model_pred = self.predict(
+                    t=t,
+                    latents=latents,
+                    prompt_embeds=prompt_embeds,
+                    freqs_cis=freqs_cis,
+                    prompt_attention_mask=prompt_attention_mask,
+                    ref_image_hidden_states=ref_latents,
+                )
+                text_guidance_scale = self.text_guidance_scale if self.cfg_range[0] <= i / len(timesteps) <= self.cfg_range[1] else 1.0
+                image_guidance_scale = self.image_guidance_scale if self.cfg_range[0] <= i / len(timesteps) <= self.cfg_range[1] else 1.0
+                
+                if text_guidance_scale > 1.0 and image_guidance_scale > 1.0:
+                    if enable_taylorseer:
+                        self.transformer.cache_dic = model_pred_ref_cache_dic
+                        self.transformer.current = model_pred_ref_current
+                    elif self.transformer.enable_teacache:
+                        teacache_params_ref.is_first_or_last_step = i == 0 or i == len(timesteps) - 1
+                        self.transformer.teacache_params = teacache_params_ref
+
+                    model_pred_ref = self.predict(
+                        t=t,
+                        latents=latents,
+                        prompt_embeds=negative_prompt_embeds,
+                        freqs_cis=freqs_cis,
+                        prompt_attention_mask=negative_prompt_attention_mask,
+                        ref_image_hidden_states=ref_latents,
+                    )
+
+                    if enable_taylorseer:
+                        self.transformer.cache_dic = model_pred_uncond_cache_dic
+                        self.transformer.current = model_pred_uncond_current
+                    elif self.transformer.enable_teacache:
+                        teacache_params_uncond.is_first_or_last_step = i == 0 or i == len(timesteps) - 1
+                        self.transformer.teacache_params = teacache_params_uncond
+
+                    model_pred_uncond = self.predict(
+                        t=t,
+                        latents=latents,
+                        prompt_embeds=negative_prompt_embeds,
+                        freqs_cis=freqs_cis,
+                        prompt_attention_mask=negative_prompt_attention_mask,
+                        ref_image_hidden_states=None,
+                    )
+
+                    model_pred = model_pred_uncond + image_guidance_scale * (model_pred_ref - model_pred_uncond) + \
+                        text_guidance_scale * (model_pred - model_pred_ref)
+                elif text_guidance_scale > 1.0:
+                    if enable_taylorseer:
+                        self.transformer.cache_dic = model_pred_uncond_cache_dic
+                        self.transformer.current = model_pred_uncond_current
+                    elif self.transformer.enable_teacache:
+                        teacache_params_uncond.is_first_or_last_step = i == 0 or i == len(timesteps) - 1
+                        self.transformer.teacache_params = teacache_params_uncond
+
+                    model_pred_uncond = self.predict(
+                        t=t,
+                        latents=latents,
+                        prompt_embeds=negative_prompt_embeds,
+                        freqs_cis=freqs_cis,
+                        prompt_attention_mask=negative_prompt_attention_mask,
+                        ref_image_hidden_states=None,
+                    )
+                    model_pred = model_pred_uncond + text_guidance_scale * (model_pred - model_pred_uncond)
+
+                latents = self.scheduler.step(model_pred, t, latents, return_dict=False)[0]
+
+                latents = latents.to(dtype=dtype)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                
+                if step_func is not None:
+                    step_func(i, self._num_timesteps)
+
+        if enable_taylorseer:
+            del model_pred_cache_dic, model_pred_ref_cache_dic, model_pred_uncond_cache_dic
+            del model_pred_current, model_pred_ref_current, model_pred_uncond_current
+
+        latents = latents.to(dtype=dtype)
+        if self.vae.config.scaling_factor is not None:
+            latents = latents / self.vae.config.scaling_factor
+        if self.vae.config.shift_factor is not None:
+            latents = latents + self.vae.config.shift_factor
+        image = self.vae.decode(latents, return_dict=False)[0]
+        
+        return image
+
+    def predict(
+        self,
+        t,
+        latents,
+        prompt_embeds,
+        freqs_cis,
+        prompt_attention_mask,
+        ref_image_hidden_states,
+    ):
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+        batch_size, num_channels_latents, height, width = latents.shape
+        
+        optional_kwargs = {}
+        if 'ref_image_hidden_states' in set(inspect.signature(self.transformer.forward).parameters.keys()):
+            optional_kwargs['ref_image_hidden_states'] = ref_image_hidden_states
+        
+        model_pred = self.transformer(
+            latents,
+            timestep,
+            prompt_embeds,
+            freqs_cis,
+            prompt_attention_mask,
+            **optional_kwargs
+        )
+        return model_pred

omnigen2/pipelines/omnigen2/pipeline_omnigen2_chat.py

@@ -0,0 +1,830 @@
+"""
+OmniGen2 Diffusion Pipeline
+
+Copyright 2025 BAAI, The OmniGen2 Team and The HuggingFace 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.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import math
+
+from PIL import Image
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from transformers import Qwen2_5_VLForConditionalGeneration
+
+from diffusers.models.autoencoders import AutoencoderKL
+from ...models.transformers import OmniGen2Transformer2DModel
+from ...models.transformers.repo import OmniGen2RotaryPosEmbed
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import (
+    is_torch_xla_available,
+    logging,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+
+from dataclasses import dataclass
+
+import PIL.Image
+
+from diffusers.utils import BaseOutput
+
+from omnigen2.pipelines.image_processor import OmniGen2ImageProcessor
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+@dataclass
+class OmniGen2PipelineOutput(BaseOutput):
+    """
+    Output class for OmniGen2 pipeline.
+
+    Args:
+        images (Union[List[PIL.Image.Image], np.ndarray]): 
+            List of denoised PIL images of length `batch_size` or numpy array of shape 
+            `(batch_size, height, width, num_channels)`. Contains the generated images.
+    """
+    text: str
+    images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class OmniGen2ChatPipeline(DiffusionPipeline):
+    """
+    Pipeline for text-to-image generation using OmniGen2.
+
+    This pipeline implements a text-to-image generation model that uses:
+    - Qwen2.5-VL for text encoding
+    - A custom transformer architecture for image generation
+    - VAE for image encoding/decoding
+    - FlowMatchEulerDiscreteScheduler for noise scheduling
+
+    Args:
+        transformer (OmniGen2Transformer2DModel): The transformer model for image generation.
+        vae (AutoencoderKL): The VAE model for image encoding/decoding.
+        scheduler (FlowMatchEulerDiscreteScheduler): The scheduler for noise scheduling.
+        text_encoder (Qwen2_5_VLModel): The text encoder model.
+        tokenizer (Union[Qwen2Tokenizer, Qwen2TokenizerFast]): The tokenizer for text processing.
+    """
+
+    model_cpu_offload_seq = "mllm->transformer->vae"
+    def __init__(
+        self,
+        transformer: OmniGen2Transformer2DModel,
+        vae: AutoencoderKL,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        mllm: Qwen2_5_VLForConditionalGeneration,
+        processor,
+    ) -> None:
+        """
+        Initialize the OmniGen2 pipeline.
+
+        Args:
+            transformer: The transformer model for image generation.
+            vae: The VAE model for image encoding/decoding.
+            scheduler: The scheduler for noise scheduling.
+            text_encoder: The text encoder model.
+            tokenizer: The tokenizer for text processing.
+        """
+        super().__init__()
+
+        self.register_modules(
+            transformer=transformer,
+            vae=vae,
+            scheduler=scheduler,
+            mllm=mllm,
+            processor=processor
+        )
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+        self.image_processor = OmniGen2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2, do_resize=True)
+        self.default_sample_size = 128
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        generator: Optional[torch.Generator],
+        latents: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        """
+        Prepare the initial latents for the diffusion process.
+
+        Args:
+            batch_size: The number of images to generate.
+            num_channels_latents: The number of channels in the latent space.
+            height: The height of the generated image.
+            width: The width of the generated image.
+            dtype: The data type of the latents.
+            device: The device to place the latents on.
+            generator: The random number generator to use.
+            latents: Optional pre-computed latents to use instead of random initialization.
+
+        Returns:
+            torch.FloatTensor: The prepared latents tensor.
+        """
+        height = int(height) // self.vae_scale_factor
+        width = int(width) // self.vae_scale_factor
+
+        shape = (batch_size, num_channels_latents, height, width)
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+        return latents
+
+    def encode_vae(self, img: torch.FloatTensor) -> torch.FloatTensor:
+        """
+        Encode an image into the VAE latent space.
+
+        Args:
+            img: The input image tensor to encode.
+
+        Returns:
+            torch.FloatTensor: The encoded latent representation.
+        """
+        z0 = self.vae.encode(img.to(dtype=self.vae.dtype)).latent_dist.sample()
+        if self.vae.config.shift_factor is not None:
+            z0 = z0 - self.vae.config.shift_factor
+        if self.vae.config.scaling_factor is not None:
+            z0 = z0 * self.vae.config.scaling_factor
+        z0 = z0.to(dtype=self.vae.dtype)
+        return z0
+
+    def prepare_image(
+        self,
+        images: Union[List[PIL.Image.Image], PIL.Image.Image],
+        batch_size: int,
+        num_images_per_prompt: int,
+        max_pixels: int,
+        max_side_length: int,
+        device: torch.device,
+        dtype: torch.dtype,
+    ) -> List[Optional[torch.FloatTensor]]:
+        """
+        Prepare input images for processing by encoding them into the VAE latent space.
+
+        Args:
+            images: Single image or list of images to process.
+            batch_size: The number of images to generate per prompt.
+            num_images_per_prompt: The number of images to generate for each prompt.
+            device: The device to place the encoded latents on.
+            dtype: The data type of the encoded latents.
+
+        Returns:
+            List[Optional[torch.FloatTensor]]: List of encoded latent representations for each image.
+        """
+        if batch_size == 1:
+            images = [images]
+        latents = []
+        for i, img in enumerate(images):
+            if img is not None and len(img) > 0:
+                ref_latents = []
+                for j, img_j in enumerate(img):
+                    img_j = self.image_processor.preprocess(img_j, max_pixels=max_pixels, max_side_length=max_side_length)
+                    ref_latents.append(self.encode_vae(img_j.to(device=device)).squeeze(0))
+            else:
+                ref_latents = None
+            for _ in range(num_images_per_prompt):
+                latents.append(ref_latents)
+
+        return latents
+    
+    def _apply_chat_template(self, prompt: str, images: List = None):
+        if images is not None:
+            prompt = "".join(
+                [
+                    f"<img{i}>: <|vision_start|><|image_pad|><|vision_end|>"
+                    for i in range(1, len(images) + 1)
+                ]
+            ) + prompt
+        prompt = f"<|im_start|>system\nYou are a helpful assistant that generates high-quality images based on user instructions.<|im_end|>\n<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"
+        return prompt
+    
+    def _get_qwen2_prompt_embeds(
+        self, 
+        prompt: Union[str, List[str]],
+        input_images = None, 
+        device: Optional[torch.device] = None,
+        use_only_text_hidden_states: bool = True,
+        ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Get prompt embeddings from the Qwen2 text encoder.
+
+        Args:
+            prompt: The prompt or list of prompts to encode.
+            device: The device to place the embeddings on. If None, uses the pipeline's device.
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]: A tuple containing:
+                - The prompt embeddings tensor
+                - The attention mask tensor
+
+        Raises:
+            Warning: If the input text is truncated due to sequence length limitations.
+        """
+        device = device or self._execution_device
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        inputs = self.processor(
+            text=prompt,
+            images=input_images,
+            videos=None,
+            padding=True,
+            return_tensors="pt",
+        )
+        inputs = inputs.to(device)
+
+        prompt_embeds = self.mllm(
+            **inputs,
+            output_hidden_states=True,
+        ).hidden_states[-1]
+
+        text_input_ids = inputs.input_ids
+        text_mask = inputs.attention_mask
+        if use_only_text_hidden_states:
+            mask = text_input_ids != self.mllm.config.image_token_id
+            mask = mask & text_mask
+            mask = mask.bool()
+            
+            text_l = mask.sum(dim=-1)
+            max_l = text_l.max()
+            text_batch_size = prompt_embeds.size(0)
+            new_prompt_embeds = torch.zeros((text_batch_size, max_l, prompt_embeds.size(-1)), device=prompt_embeds.device, dtype=prompt_embeds.dtype)
+            new_text_mask = torch.zeros((text_batch_size, max_l), dtype=text_mask.dtype, device=text_mask.device)
+            for i in range(text_batch_size):
+                new_prompt_embeds[i, :text_l[i]] = prompt_embeds[i, mask[i]]
+                new_text_mask[i, :text_l[i]] = 1
+
+            prompt_embeds = new_prompt_embeds
+            text_mask = new_text_mask
+
+        prompt_embeds = prompt_embeds.to(dtype=self.mllm.dtype, device=device)
+        return prompt_embeds, text_mask
+
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        input_images: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 256,
+        use_text_encoder_penultimate_layer_feats: bool = False
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds`
+                instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For
+                Lumina-T2I, this should be "".
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                whether to use classifier free guidance or not
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                number of images that should be generated per prompt
+            device: (`torch.device`, *optional*):
+                torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. For Lumina-T2I, it's should be the embeddings of the "" string.
+            max_sequence_length (`int`, defaults to `256`):
+                Maximum sequence length to use for the prompt.
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_qwen2_prompt_embeds(
+                prompt=prompt,
+                input_images=input_images,
+                device=device,
+            )
+
+        batch_size, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
+        prompt_attention_mask = prompt_attention_mask.view(batch_size * num_images_per_prompt, -1)
+
+        # Get negative embeddings for classifier free guidance
+        negative_prompt_embeds, negative_prompt_attention_mask = None, None
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt if negative_prompt is not None else ""
+
+            # Normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt = [self._apply_chat_template(_negative_prompt) for _negative_prompt in negative_prompt]
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                negative_prompt = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_qwen2_prompt_embeds(
+                prompt=negative_prompt,
+                device=device,
+            )
+
+            batch_size, seq_len, _ = negative_prompt_embeds.shape
+            # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.view(
+                batch_size * num_images_per_prompt, -1
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+    
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def text_guidance_scale(self):
+        return self._text_guidance_scale
+    
+    @property
+    def image_guidance_scale(self):
+        return self._image_guidance_scale
+
+    @property
+    def cfg_range(self):
+        return self._cfg_range
+    
+    def prepare_inputs_for_text_generation(self, prompts, input_images, device):
+        if isinstance(prompts, str):
+            prompts = [prompts]
+
+        ori_padding_side = self.processor.tokenizer.padding_side
+        self.processor.tokenizer.padding_side = "left"
+        inputs = self.processor(
+            text=prompts,
+            images=input_images,
+            videos=None,
+            padding=True,
+            return_tensors="pt",
+        ).to(device)
+        self.processor.tokenizer.padding_side = ori_padding_side
+        return inputs
+
+    def generate_text(self, prompt, input_images):
+        inputs = self.prepare_inputs_for_text_generation(
+            prompt, input_images, self.mllm.device
+        )
+        generated_ids = self.mllm.generate(
+            **inputs,
+            tokenizer=self.processor.tokenizer,
+            max_new_tokens=256,
+            stop_strings=["<|im_end|>", "<|img|>", "<|endoftext|>"],
+        )  # stop_words=[151643, 151645, 151665]
+        generated_ids_trimmed = [
+            out_ids[len(in_ids) :]
+            for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
+        ]
+        output_texts = self.processor.batch_decode(
+            generated_ids_trimmed,
+            # skip_special_tokens=True,
+            skip_special_tokens=False,
+            clean_up_tokenization_spaces=False,
+        )
+        return output_texts
+    
+    def generate_image(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.LongTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.LongTensor] = None,
+        use_text_encoder_penultimate_layer_feats: bool = False,
+        max_sequence_length: Optional[int] = None,
+        callback_on_step_end_tensor_inputs: Optional[List[str]] = None,
+        input_images: Optional[List[PIL.Image.Image]] = None,
+        num_images_per_prompt: int = 1,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        max_pixels: int = 1024 * 1024,
+        max_input_image_side_length: int = 1024,
+        align_res: bool = True,
+        num_inference_steps: int = 28,
+        text_guidance_scale: float = 4.0,
+        image_guidance_scale: float = 1.0,
+        cfg_range: Tuple[float, float] = (0.0, 1.0),
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        timesteps: List[int] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        verbose: bool = False,
+        step_func=None,
+    ):
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        self._text_guidance_scale = text_guidance_scale
+        self._image_guidance_scale = image_guidance_scale
+        self._cfg_range = cfg_range
+        self._attention_kwargs = attention_kwargs
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input promptb
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt,
+            input_images,
+            self.text_guidance_scale > 1.0,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            use_text_encoder_penultimate_layer_feats=use_text_encoder_penultimate_layer_feats
+        )
+
+        dtype = self.vae.dtype
+        # 3. Prepare control image
+        ref_latents = self.prepare_image(
+            images=input_images,
+            batch_size=batch_size,
+            num_images_per_prompt=num_images_per_prompt,
+            max_pixels=max_pixels,
+            max_side_length=max_input_image_side_length,
+            device=device,
+            dtype=dtype,
+        )
+
+        if input_images is None:
+            input_images = []
+        
+        if len(input_images) == 1 and align_res:
+            width, height = ref_latents[0][0].shape[-1] * self.vae_scale_factor, ref_latents[0][0].shape[-2] * self.vae_scale_factor
+            ori_width, ori_height = width, height
+        else:
+            ori_width, ori_height = width, height
+
+            cur_pixels = height * width
+            ratio = (max_pixels / cur_pixels) ** 0.5
+            ratio = min(ratio, 1.0)
+
+            height, width = int(height * ratio) // 16 * 16, int(width * ratio) // 16 * 16
+        
+        if len(input_images) == 0:
+            self._image_guidance_scale = 1
+
+        # 4. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            latent_channels,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        freqs_cis = OmniGen2RotaryPosEmbed.get_freqs_cis(
+            self.transformer.config.axes_dim_rope,
+            self.transformer.config.axes_lens,
+            theta=10000,
+        )
+        
+        image = self.processing(
+            latents=latents,
+            ref_latents=ref_latents,
+            prompt_embeds=prompt_embeds,
+            freqs_cis=freqs_cis,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            num_inference_steps=num_inference_steps,
+            timesteps=timesteps,
+            device=device,
+            dtype=dtype,
+            verbose=verbose,
+            step_func=step_func,
+        )
+
+        image = F.interpolate(image, size=(ori_height, ori_width), mode='bilinear')
+
+        image = self.image_processor.postprocess(image, output_type=output_type)
+        
+        # Offload all models
+        self.maybe_free_model_hooks()
+        return image
+        
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.LongTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.LongTensor] = None,
+        use_text_encoder_penultimate_layer_feats: bool = False,
+        max_sequence_length: Optional[int] = None,
+        callback_on_step_end_tensor_inputs: Optional[List[str]] = None,
+        input_images: Optional[List[PIL.Image.Image]] = None,
+        num_images_per_prompt: int = 1,
+        height: Optional[int] = 1024,
+        width: Optional[int] = 1024,
+        max_pixels: Optional[int] = 1024 * 1024,
+        max_input_image_side_length: int = 1024,
+        align_res: bool = True,
+        num_inference_steps: int = 28,
+        text_guidance_scale: float = 4.0,
+        image_guidance_scale: float = 1.0,
+        cfg_range: Tuple[float, float] = (0.0, 1.0),
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        timesteps: List[int] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        verbose: bool = False,
+        step_func=None,
+    ):
+        assert isinstance(prompt, str), "prompt must be a string since chat mode only support one prompt per turn"
+
+        # input_images = self.preprocess_images(input_images, max_input_image_size)
+        prompt = self._apply_chat_template(prompt, input_images)
+        generated_text = self.generate_text(prompt, input_images)[0]
+
+        images = None
+        if generated_text.startswith("<|img|>"):
+            #TODO: reuse the hidden state when generate text instead of re-generating
+            prompt = prompt + generated_text.split("<|img|>")[0]
+            images = self.generate_image(
+                prompt=prompt,
+                negative_prompt=negative_prompt,
+                use_text_encoder_penultimate_layer_feats=use_text_encoder_penultimate_layer_feats,
+                max_sequence_length=max_sequence_length,
+                input_images=input_images,
+                num_images_per_prompt=num_images_per_prompt,
+                height=height,
+                width=width,
+                max_pixels=max_pixels,
+                max_input_image_side_length=max_input_image_side_length,
+                align_res=align_res,
+                num_inference_steps=num_inference_steps,
+                text_guidance_scale=text_guidance_scale,
+                image_guidance_scale=image_guidance_scale,
+                cfg_range=cfg_range,
+                timesteps=timesteps,
+                generator=generator,
+                latents=latents,
+                return_dict=False,
+                verbose=verbose,
+                step_func=step_func,
+            )
+
+        generated_text = generated_text.replace("<|im_end|>", "")
+        if not return_dict:
+            return generated_text, images
+        else:
+            return OmniGen2PipelineOutput(text=generated_text, images=images)
+    
+    def processing(
+        self,
+        latents,
+        ref_latents,
+        prompt_embeds,
+        freqs_cis,
+        negative_prompt_embeds,
+        prompt_attention_mask,
+        negative_prompt_attention_mask,
+        num_inference_steps,
+        timesteps,
+        device,
+        dtype,
+        verbose,
+        step_func=None
+    ):
+        batch_size = latents.shape[0]
+
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            num_tokens=latents.shape[-2] * latents.shape[-1]
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+        
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                model_pred = self.predict(
+                    t=t,
+                    latents=latents,
+                    prompt_embeds=prompt_embeds,
+                    freqs_cis=freqs_cis,
+                    prompt_attention_mask=prompt_attention_mask,
+                    ref_image_hidden_states=ref_latents,
+                )
+                
+                text_guidance_scale = self.text_guidance_scale if self.cfg_range[0] <= i / len(timesteps) <= self.cfg_range[1] else 1.0
+                image_guidance_scale = self.image_guidance_scale if self.cfg_range[0] <= i / len(timesteps) <= self.cfg_range[1] else 1.0
+                if text_guidance_scale > 1.0 and image_guidance_scale > 1.0:
+                    model_pred_ref = self.predict(
+                        t=t,
+                        latents=latents,
+                        prompt_embeds=negative_prompt_embeds,
+                        freqs_cis=freqs_cis,
+                        prompt_attention_mask=negative_prompt_attention_mask,
+                        ref_image_hidden_states=ref_latents,
+                    )
+
+                    if image_guidance_scale != 1:
+                        model_pred_uncond = self.predict(
+                            t=t,
+                            latents=latents,
+                            prompt_embeds=negative_prompt_embeds,
+                            freqs_cis=freqs_cis,
+                            prompt_attention_mask=negative_prompt_attention_mask,
+                            ref_image_hidden_states=None,
+                        )
+                    else:
+                        model_pred_uncond = torch.zeros_like(model_pred)
+
+                    model_pred = model_pred_uncond + image_guidance_scale * (model_pred_ref - model_pred_uncond) + \
+                    text_guidance_scale * (model_pred - model_pred_ref)
+                elif text_guidance_scale > 1.0:
+                    model_pred_uncond = self.predict(
+                        t=t,
+                        latents=latents,
+                        prompt_embeds=negative_prompt_embeds,
+                        freqs_cis=freqs_cis,
+                        prompt_attention_mask=negative_prompt_attention_mask,
+                        ref_image_hidden_states=None,
+                    )
+                    model_pred = model_pred_uncond + text_guidance_scale * (model_pred - model_pred_uncond)
+
+                latents = self.scheduler.step(model_pred, t, latents, return_dict=False)[0]
+
+                latents = latents.to(dtype=dtype)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                
+                if step_func is not None:
+                    step_func(i, self._num_timesteps)
+
+        latents = latents.to(dtype=dtype)
+        if self.vae.config.scaling_factor is not None:
+            latents = latents / self.vae.config.scaling_factor
+        if self.vae.config.shift_factor is not None:
+            latents = latents + self.vae.config.shift_factor
+        image = self.vae.decode(latents, return_dict=False)[0]
+        
+        return image
+
+    def predict(
+        self,
+        t,
+        latents,
+        prompt_embeds,
+        freqs_cis,
+        prompt_attention_mask,
+        ref_image_hidden_states,
+    ):
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+        batch_size, num_channels_latents, height, width = latents.shape
+        
+        optional_kwargs = {}
+        if 'ref_image_hidden_states' in set(inspect.signature(self.transformer.forward).parameters.keys()):
+            optional_kwargs['ref_image_hidden_states'] = ref_image_hidden_states
+
+        model_pred = self.transformer(
+            latents,
+            timestep,
+            prompt_embeds,
+            freqs_cis,
+            prompt_attention_mask,
+            **optional_kwargs
+        )
+        return model_pred

omnigen2/pipelines/pipeline_utils.py

@@ -0,0 +1,62 @@
+import torch
+
+
+def get_pipeline_embeds(pipeline, prompt, negative_prompt, device):
+    """ Get pipeline embeds for prompts bigger than the maxlength of the pipe
+    :param pipeline:
+    :param prompt:
+    :param negative_prompt:
+    :param device:
+    :return:
+    """
+    max_length = pipeline.tokenizer.model_max_length
+
+    # simple way to determine length of tokens
+    # count_prompt = len(prompt.split(" "))
+    # count_negative_prompt = len(negative_prompt.split(" "))
+
+    # create the tensor based on which prompt is longer
+    # if count_prompt >= count_negative_prompt:
+    input_ids = pipeline.tokenizer(prompt, return_tensors="pt", truncation=False, padding='longest').input_ids.to(device)
+    # input_ids = pipeline.tokenizer(prompt, padding="max_length",
+    #             max_length=pipeline.tokenizer.model_max_length,
+    #             truncation=True,
+    #             return_tensors="pt",).input_ids.to(device)
+    shape_max_length = input_ids.shape[-1]
+
+    if negative_prompt is not None:
+        negative_ids = pipeline.tokenizer(negative_prompt, truncation=True, padding="max_length",
+                                        max_length=shape_max_length, return_tensors="pt").input_ids.to(device)
+
+    # else:
+    #     negative_ids = pipeline.tokenizer(negative_prompt, return_tensors="pt", truncation=False).input_ids.to(device)
+    #     shape_max_length = negative_ids.shape[-1]
+    #     input_ids = pipeline.tokenizer(prompt, return_tensors="pt", truncation=False, padding="max_length",
+    #                                    max_length=shape_max_length).input_ids.to(device)
+
+    concat_embeds = []
+    neg_embeds = []
+    for i in range(0, shape_max_length, max_length):
+        if hasattr(pipeline.text_encoder.config, "use_attention_mask") and pipeline.text_encoder.config.use_attention_mask:
+            attention_mask = input_ids[:, i: i + max_length].attention_mask.to(device)
+        else:
+            attention_mask = None
+        concat_embeds.append(pipeline.text_encoder(input_ids[:, i: i + max_length],
+                                                   attention_mask=attention_mask)[0])
+        
+        if negative_prompt is not None:
+            if hasattr(pipeline.text_encoder.config, "use_attention_mask") and pipeline.text_encoder.config.use_attention_mask:
+                attention_mask = negative_ids[:, i: i + max_length].attention_mask.to(device)
+            else:
+                attention_mask = None
+            neg_embeds.append(pipeline.text_encoder(negative_ids[:, i: i + max_length],
+                                                    attention_mask=attention_mask)[0])
+
+    concat_embeds = torch.cat(concat_embeds, dim=1)
+
+    if negative_prompt is not None:
+        neg_embeds = torch.cat(neg_embeds, dim=1)
+    else:
+        neg_embeds = None
+
+    return concat_embeds, neg_embeds

omnigen2/schedulers/scheduling_dpmsolver_multistep.py

@@ -0,0 +1,1052 @@
+# Copyright 2024 TSAIL Team and The HuggingFace 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import deprecate, is_scipy_available
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+if is_scipy_available():
+    import scipy.stats
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+    (1-beta) over time from t = [0,1].
+
+    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+    to that part of the diffusion process.
+
+
+    Args:
+        num_diffusion_timesteps (`int`): the number of betas to produce.
+        max_beta (`float`): the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
+
+    Returns:
+        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+    """
+    if alpha_transform_type == "cosine":
+
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
+
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+    return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+    """
+    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+    Args:
+        betas (`torch.Tensor`):
+            the betas that the scheduler is being initialized with.
+
+    Returns:
+        `torch.Tensor`: rescaled betas with zero terminal SNR
+    """
+    # Convert betas to alphas_bar_sqrt
+    alphas = 1.0 - betas
+    alphas_cumprod = torch.cumprod(alphas, dim=0)
+    alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
+    alphas = torch.cat([alphas_bar[0:1], alphas])
+    betas = 1 - alphas
+
+    return betas
+
+
+class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
+    """
+    `DPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+    methods the library implements for all schedulers such as loading and saving.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        beta_start (`float`, defaults to 0.0001):
+            The starting `beta` value of inference.
+        beta_end (`float`, defaults to 0.02):
+            The final `beta` value.
+        beta_schedule (`str`, defaults to `"linear"`):
+            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+        trained_betas (`np.ndarray`, *optional*):
+            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+        solver_order (`int`, defaults to 2):
+            The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
+            sampling, and `solver_order=3` for unconditional sampling.
+        prediction_type (`str`, defaults to `epsilon`, *optional*):
+            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+            `sample` (directly predicts the noisy sample), `v_prediction` (see section 2.4 of [Imagen
+            Video](https://imagen.research.google/video/paper.pdf) paper), or `flow_prediction`.
+        thresholding (`bool`, defaults to `False`):
+            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+            as Stable Diffusion.
+        dynamic_thresholding_ratio (`float`, defaults to 0.995):
+            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+        sample_max_value (`float`, defaults to 1.0):
+            The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
+            `algorithm_type="dpmsolver++"`.
+        algorithm_type (`str`, defaults to `dpmsolver++`):
+            Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, `sde-dpmsolver` or `sde-dpmsolver++`. The
+            `dpmsolver` type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927)
+            paper, and the `dpmsolver++` type implements the algorithms in the
+            [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or
+            `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion.
+        solver_type (`str`, defaults to `midpoint`):
+            Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
+            sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
+        lower_order_final (`bool`, defaults to `True`):
+            Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
+            stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
+        euler_at_final (`bool`, defaults to `False`):
+            Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail
+            richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
+            steps, but sometimes may result in blurring.
+        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+            Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+            the sigmas are determined according to a sequence of noise levels {σi}.
+        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
+            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
+        use_beta_sigmas (`bool`, *optional*, defaults to `False`):
+            Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta
+            Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information.
+        use_lu_lambdas (`bool`, *optional*, defaults to `False`):
+            Whether to use the uniform-logSNR for step sizes proposed by Lu's DPM-Solver in the noise schedule during
+            the sampling process. If `True`, the sigmas and time steps are determined according to a sequence of
+            `lambda(t)`.
+        use_flow_sigmas (`bool`, *optional*, defaults to `False`):
+            Whether to use flow sigmas for step sizes in the noise schedule during the sampling process.
+        flow_shift (`float`, *optional*, defaults to 1.0):
+            The shift value for the timestep schedule for flow matching.
+        final_sigmas_type (`str`, defaults to `"zero"`):
+            The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final
+            sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
+        lambda_min_clipped (`float`, defaults to `-inf`):
+            Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
+            cosine (`squaredcos_cap_v2`) noise schedule.
+        variance_type (`str`, *optional*):
+            Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
+            contains the predicted Gaussian variance.
+        timestep_spacing (`str`, defaults to `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        steps_offset (`int`, defaults to 0):
+            An offset added to the inference steps, as required by some model families.
+        rescale_betas_zero_snr (`bool`, defaults to `False`):
+            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+            dark samples instead of limiting it to samples with medium brightness. Loosely related to
+            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+    """
+
+    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        beta_start: float = 0.0001,
+        beta_end: float = 0.02,
+        beta_schedule: str = "linear",
+        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+        solver_order: int = 2,
+        prediction_type: str = "epsilon",
+        thresholding: bool = False,
+        dynamic_thresholding_ratio: float = 0.995,
+        sample_max_value: float = 1.0,
+        algorithm_type: str = "dpmsolver++",
+        solver_type: str = "midpoint",
+        lower_order_final: bool = True,
+        euler_at_final: bool = False,
+        final_sigmas_type: str = 'zero',
+        dynamic_time_shift: bool = True
+    ):
+        if algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
+            deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
+            deprecate("algorithm_types dpmsolver and sde-dpmsolver", "1.0.0", deprecation_message)
+
+        if trained_betas is not None:
+            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+        elif beta_schedule == "linear":
+            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+        elif beta_schedule == "scaled_linear":
+            # this schedule is very specific to the latent diffusion model.
+            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+        elif beta_schedule == "squaredcos_cap_v2":
+            # Glide cosine schedule
+            self.betas = betas_for_alpha_bar(num_train_timesteps)
+        else:
+            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
+        self.alphas = 1.0 - self.betas
+        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+        # Currently we only support VP-type noise schedule
+        self.alpha_t = torch.sqrt(self.alphas_cumprod)
+        self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
+        self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
+        self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
+
+        # standard deviation of the initial noise distribution
+        self.init_noise_sigma = 1.0
+
+        # settings for DPM-Solver
+        if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"]:
+            if algorithm_type == "deis":
+                self.register_to_config(algorithm_type="dpmsolver++")
+            else:
+                raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
+
+        if solver_type not in ["midpoint", "heun"]:
+            if solver_type in ["logrho", "bh1", "bh2"]:
+                self.register_to_config(solver_type="midpoint")
+            else:
+                raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}")
+
+        # if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero":
+        #     raise ValueError(
+        #         f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead."
+        #     )
+
+        # setable values
+        self.num_inference_steps = None
+        timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
+        self.timesteps = torch.from_numpy(timesteps)
+        self.model_outputs = [None] * solver_order
+        self.lower_order_nums = 0
+        self._step_index = None
+        self._begin_index = None
+        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
+
+    @property
+    def step_index(self):
+        """
+        The index counter for current timestep. It will increase 1 after each scheduler step.
+        """
+        return self._step_index
+
+    @property
+    def begin_index(self):
+        """
+        The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+        """
+        return self._begin_index
+
+    def set_begin_index(self, begin_index: int = 0):
+        """
+        Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+        Args:
+            begin_index (`int`):
+                The begin index for the scheduler.
+        """
+        self._begin_index = begin_index
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int = None,
+        device: Union[str, torch.device] = None,
+        timesteps: Optional[List[int]] = None,
+        num_tokens: Optional[int] = None
+    ):
+        if timesteps is None:
+            self.num_inference_steps = num_inference_steps
+            timesteps = np.linspace(0, 1, num_inference_steps + 1, dtype=np.float32)[:-1]
+            if self.config.dynamic_time_shift and num_tokens is not None:
+                m = np.sqrt(num_tokens) / 40 # when input resolution is 320 * 320, m = 1, when input resolution is 1024 * 1024, m = 3.2
+                timesteps = timesteps / (m - m * timesteps + timesteps)
+
+        timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device)
+        sigmas = torch.cat([1 - timesteps, torch.zeros(1, device=timesteps.device)])
+
+        self.sigmas = sigmas
+        self.timesteps = timesteps
+
+        self.num_inference_steps = len(timesteps)
+
+        self.model_outputs = [
+            None,
+        ] * self.config.solver_order
+        self.lower_order_nums = 0
+
+        # add an index counter for schedulers that allow duplicated timesteps
+        self._step_index = None
+        self._begin_index = None
+        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+    def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor:
+        """
+        "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+        prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+        s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+        pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+        photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+        https://arxiv.org/abs/2205.11487
+        """
+        dtype = sample.dtype
+        batch_size, channels, *remaining_dims = sample.shape
+
+        if dtype not in (torch.float32, torch.float64):
+            sample = sample.float()  # upcast for quantile calculation, and clamp not implemented for cpu half
+
+        # Flatten sample for doing quantile calculation along each image
+        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+        abs_sample = sample.abs()  # "a certain percentile absolute pixel value"
+
+        s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+        s = torch.clamp(
+            s, min=1, max=self.config.sample_max_value
+        )  # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+        s = s.unsqueeze(1)  # (batch_size, 1) because clamp will broadcast along dim=0
+        sample = torch.clamp(sample, -s, s) / s  # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+        sample = sample.reshape(batch_size, channels, *remaining_dims)
+        sample = sample.to(dtype)
+
+        return sample
+
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+    def _sigma_to_t(self, sigma, log_sigmas):
+        # get log sigma
+        log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+        # get distribution
+        dists = log_sigma - log_sigmas[:, np.newaxis]
+
+        # get sigmas range
+        low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+        high_idx = low_idx + 1
+
+        low = log_sigmas[low_idx]
+        high = log_sigmas[high_idx]
+
+        # interpolate sigmas
+        w = (low - log_sigma) / (low - high)
+        w = np.clip(w, 0, 1)
+
+        # transform interpolation to time range
+        t = (1 - w) * low_idx + w * high_idx
+        t = t.reshape(sigma.shape)
+        return t
+
+    def _sigma_to_alpha_sigma_t(self, sigma):
+        alpha_t = 1 - sigma
+        sigma_t = sigma
+
+        return alpha_t, sigma_t
+
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
+        """Constructs the noise schedule of Karras et al. (2022)."""
+
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+        rho = 7.0  # 7.0 is the value used in the paper
+        ramp = np.linspace(0, 1, num_inference_steps)
+        min_inv_rho = sigma_min ** (1 / rho)
+        max_inv_rho = sigma_max ** (1 / rho)
+        sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+        return sigmas
+
+    def _convert_to_lu(self, in_lambdas: torch.Tensor, num_inference_steps) -> torch.Tensor:
+        """Constructs the noise schedule of Lu et al. (2022)."""
+
+        lambda_min: float = in_lambdas[-1].item()
+        lambda_max: float = in_lambdas[0].item()
+
+        rho = 1.0  # 1.0 is the value used in the paper
+        ramp = np.linspace(0, 1, num_inference_steps)
+        min_inv_rho = lambda_min ** (1 / rho)
+        max_inv_rho = lambda_max ** (1 / rho)
+        lambdas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+        return lambdas
+
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential
+    def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
+        """Constructs an exponential noise schedule."""
+
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+        sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps))
+        return sigmas
+
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta
+    def _convert_to_beta(
+        self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6
+    ) -> torch.Tensor:
+        """From "Beta Sampling is All You Need" [arXiv:2407.12173] (Lee et. al, 2024)"""
+
+        # Hack to make sure that other schedulers which copy this function don't break
+        # TODO: Add this logic to the other schedulers
+        if hasattr(self.config, "sigma_min"):
+            sigma_min = self.config.sigma_min
+        else:
+            sigma_min = None
+
+        if hasattr(self.config, "sigma_max"):
+            sigma_max = self.config.sigma_max
+        else:
+            sigma_max = None
+
+        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+        sigmas = np.array(
+            [
+                sigma_min + (ppf * (sigma_max - sigma_min))
+                for ppf in [
+                    scipy.stats.beta.ppf(timestep, alpha, beta)
+                    for timestep in 1 - np.linspace(0, 1, num_inference_steps)
+                ]
+            ]
+        )
+        return sigmas
+
+    def convert_model_output(
+        self,
+        model_output: torch.Tensor,
+        *args,
+        sample: torch.Tensor = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """
+        Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
+        designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
+        integral of the data prediction model.
+
+        <Tip>
+
+        The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise
+        prediction and data prediction models.
+
+        </Tip>
+
+        Args:
+            model_output (`torch.Tensor`):
+                The direct output from the learned diffusion model.
+            sample (`torch.Tensor`):
+                A current instance of a sample created by the diffusion process.
+
+        Returns:
+            `torch.Tensor`:
+                The converted model output.
+        """
+        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+        if sample is None:
+            if len(args) > 1:
+                sample = args[1]
+            else:
+                raise ValueError("missing `sample` as a required keyward argument")
+        if timestep is not None:
+            deprecate(
+                "timesteps",
+                "1.0.0",
+                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+            )
+
+        # DPM-Solver++ needs to solve an integral of the data prediction model.
+        if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]:
+            if self.config.prediction_type == "epsilon":
+                # DPM-Solver and DPM-Solver++ only need the "mean" output.
+                if self.config.variance_type in ["learned", "learned_range"]:
+                    model_output = model_output[:, :3]
+                sigma = self.sigmas[self.step_index]
+                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+                x0_pred = (sample - sigma_t * model_output) / alpha_t
+            elif self.config.prediction_type == "sample":
+                x0_pred = model_output
+            elif self.config.prediction_type == "v_prediction":
+                sigma = self.sigmas[self.step_index]
+                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+                x0_pred = alpha_t * sample - sigma_t * model_output
+            elif self.config.prediction_type == "flow_prediction":
+                sigma_t = self.sigmas[self.step_index]
+                x0_pred = sample + sigma_t * model_output
+            else:
+                raise ValueError(
+                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+                    "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler."
+                )
+
+            if self.config.thresholding:
+                x0_pred = self._threshold_sample(x0_pred)
+
+            return x0_pred
+
+        # DPM-Solver needs to solve an integral of the noise prediction model.
+        elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
+            if self.config.prediction_type == "epsilon":
+                # DPM-Solver and DPM-Solver++ only need the "mean" output.
+                if self.config.variance_type in ["learned", "learned_range"]:
+                    epsilon = model_output[:, :3]
+                else:
+                    epsilon = model_output
+            elif self.config.prediction_type == "sample":
+                sigma = self.sigmas[self.step_index]
+                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+                epsilon = (sample - alpha_t * model_output) / sigma_t
+            elif self.config.prediction_type == "v_prediction":
+                sigma = self.sigmas[self.step_index]
+                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+                epsilon = alpha_t * model_output + sigma_t * sample
+            else:
+                raise ValueError(
+                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+                    " `v_prediction` for the DPMSolverMultistepScheduler."
+                )
+
+            if self.config.thresholding:
+                sigma = self.sigmas[self.step_index]
+                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+                x0_pred = (sample - sigma_t * epsilon) / alpha_t
+                x0_pred = self._threshold_sample(x0_pred)
+                epsilon = (sample - alpha_t * x0_pred) / sigma_t
+
+            return epsilon
+
+    def dpm_solver_first_order_update(
+        self,
+        model_output: torch.Tensor,
+        *args,
+        sample: torch.Tensor = None,
+        noise: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """
+        One step for the first-order DPMSolver (equivalent to DDIM).
+
+        Args:
+            model_output (`torch.Tensor`):
+                The direct output from the learned diffusion model.
+            sample (`torch.Tensor`):
+                A current instance of a sample created by the diffusion process.
+
+        Returns:
+            `torch.Tensor`:
+                The sample tensor at the previous timestep.
+        """
+        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+        if sample is None:
+            if len(args) > 2:
+                sample = args[2]
+            else:
+                raise ValueError(" missing `sample` as a required keyward argument")
+        if timestep is not None:
+            deprecate(
+                "timesteps",
+                "1.0.0",
+                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+            )
+
+        if prev_timestep is not None:
+            deprecate(
+                "prev_timestep",
+                "1.0.0",
+                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+            )
+
+        sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
+        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+        alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
+        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+        lambda_s = torch.log(alpha_s) - torch.log(sigma_s)
+
+        h = lambda_t - lambda_s
+        if self.config.algorithm_type == "dpmsolver++":
+            x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
+        elif self.config.algorithm_type == "dpmsolver":
+            x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
+        elif self.config.algorithm_type == "sde-dpmsolver++":
+            assert noise is not None
+            x_t = (
+                (sigma_t / sigma_s * torch.exp(-h)) * sample
+                + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
+                + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+            )
+        elif self.config.algorithm_type == "sde-dpmsolver":
+            assert noise is not None
+            x_t = (
+                (alpha_t / alpha_s) * sample
+                - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output
+                + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+            )
+        return x_t
+
+    def multistep_dpm_solver_second_order_update(
+        self,
+        model_output_list: List[torch.Tensor],
+        *args,
+        sample: torch.Tensor = None,
+        noise: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """
+        One step for the second-order multistep DPMSolver.
+
+        Args:
+            model_output_list (`List[torch.Tensor]`):
+                The direct outputs from learned diffusion model at current and latter timesteps.
+            sample (`torch.Tensor`):
+                A current instance of a sample created by the diffusion process.
+
+        Returns:
+            `torch.Tensor`:
+                The sample tensor at the previous timestep.
+        """
+        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+        if sample is None:
+            if len(args) > 2:
+                sample = args[2]
+            else:
+                raise ValueError(" missing `sample` as a required keyward argument")
+        if timestep_list is not None:
+            deprecate(
+                "timestep_list",
+                "1.0.0",
+                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+            )
+
+        if prev_timestep is not None:
+            deprecate(
+                "prev_timestep",
+                "1.0.0",
+                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+            )
+
+        sigma_t, sigma_s0, sigma_s1 = (
+            self.sigmas[self.step_index + 1],
+            self.sigmas[self.step_index],
+            self.sigmas[self.step_index - 1],
+        )
+
+        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+
+        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+
+        m0, m1 = model_output_list[-1], model_output_list[-2]
+
+        h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
+        r0 = h_0 / h
+        D0, D1 = m0, (1.0 / r0) * (m0 - m1)
+        if self.config.algorithm_type == "dpmsolver++":
+            # See https://arxiv.org/abs/2211.01095 for detailed derivations
+            if self.config.solver_type == "midpoint":
+                x_t = (
+                    (sigma_t / sigma_s0) * sample
+                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+                    - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
+                )
+            elif self.config.solver_type == "heun":
+                x_t = (
+                    (sigma_t / sigma_s0) * sample
+                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+                )
+        elif self.config.algorithm_type == "dpmsolver":
+            # See https://arxiv.org/abs/2206.00927 for detailed derivations
+            if self.config.solver_type == "midpoint":
+                x_t = (
+                    (alpha_t / alpha_s0) * sample
+                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
+                    - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
+                )
+            elif self.config.solver_type == "heun":
+                x_t = (
+                    (alpha_t / alpha_s0) * sample
+                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
+                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+                )
+        elif self.config.algorithm_type == "sde-dpmsolver++":
+            assert noise is not None
+            if self.config.solver_type == "midpoint":
+                x_t = (
+                    (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
+                    + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
+                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+                )
+            elif self.config.solver_type == "heun":
+                x_t = (
+                    (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
+                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
+                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+                )
+        elif self.config.algorithm_type == "sde-dpmsolver":
+            assert noise is not None
+            if self.config.solver_type == "midpoint":
+                x_t = (
+                    (alpha_t / alpha_s0) * sample
+                    - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
+                    - (sigma_t * (torch.exp(h) - 1.0)) * D1
+                    + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+                )
+            elif self.config.solver_type == "heun":
+                x_t = (
+                    (alpha_t / alpha_s0) * sample
+                    - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
+                    - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+                    + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+                )
+        return x_t
+
+    def multistep_dpm_solver_third_order_update(
+        self,
+        model_output_list: List[torch.Tensor],
+        *args,
+        sample: torch.Tensor = None,
+        noise: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """
+        One step for the third-order multistep DPMSolver.
+
+        Args:
+            model_output_list (`List[torch.Tensor]`):
+                The direct outputs from learned diffusion model at current and latter timesteps.
+            sample (`torch.Tensor`):
+                A current instance of a sample created by diffusion process.
+
+        Returns:
+            `torch.Tensor`:
+                The sample tensor at the previous timestep.
+        """
+
+        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+        if sample is None:
+            if len(args) > 2:
+                sample = args[2]
+            else:
+                raise ValueError(" missing`sample` as a required keyward argument")
+        if timestep_list is not None:
+            deprecate(
+                "timestep_list",
+                "1.0.0",
+                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+            )
+
+        if prev_timestep is not None:
+            deprecate(
+                "prev_timestep",
+                "1.0.0",
+                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+            )
+
+        sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
+            self.sigmas[self.step_index + 1],
+            self.sigmas[self.step_index],
+            self.sigmas[self.step_index - 1],
+            self.sigmas[self.step_index - 2],
+        )
+
+        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+        alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)
+
+        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+        lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)
+
+        m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
+
+        h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2
+        r0, r1 = h_0 / h, h_1 / h
+        D0 = m0
+        D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2)
+        D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
+        D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
+        if self.config.algorithm_type == "dpmsolver++":
+            # See https://arxiv.org/abs/2206.00927 for detailed derivations
+            x_t = (
+                (sigma_t / sigma_s0) * sample
+                - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+                + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+                - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
+            )
+        elif self.config.algorithm_type == "dpmsolver":
+            # See https://arxiv.org/abs/2206.00927 for detailed derivations
+            x_t = (
+                (alpha_t / alpha_s0) * sample
+                - (sigma_t * (torch.exp(h) - 1.0)) * D0
+                - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+                - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
+            )
+        elif self.config.algorithm_type == "sde-dpmsolver++":
+            assert noise is not None
+            x_t = (
+                (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+                + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0
+                + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
+                + (alpha_t * ((1.0 - torch.exp(-2.0 * h) - 2.0 * h) / (2.0 * h) ** 2 - 0.5)) * D2
+                + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+            )
+        return x_t
+
+    def index_for_timestep(self, timestep, schedule_timesteps=None):
+        if schedule_timesteps is None:
+            schedule_timesteps = self.timesteps
+
+        index_candidates = (schedule_timesteps == timestep).nonzero()
+
+        if len(index_candidates) == 0:
+            step_index = len(self.timesteps) - 1
+        # The sigma index that is taken for the **very** first `step`
+        # is always the second index (or the last index if there is only 1)
+        # This way we can ensure we don't accidentally skip a sigma in
+        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+        elif len(index_candidates) > 1:
+            step_index = index_candidates[1].item()
+        else:
+            step_index = index_candidates[0].item()
+
+        return step_index
+
+    def _init_step_index(self, timestep):
+        """
+        Initialize the step_index counter for the scheduler.
+        """
+
+        if self.begin_index is None:
+            if isinstance(timestep, torch.Tensor):
+                timestep = timestep.to(self.timesteps.device)
+            self._step_index = self.index_for_timestep(timestep)
+        else:
+            self._step_index = self._begin_index
+
+    def step(
+        self,
+        model_output: torch.Tensor,
+        timestep: Union[int, torch.Tensor],
+        sample: torch.Tensor,
+        generator=None,
+        variance_noise: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[SchedulerOutput, Tuple]:
+        """
+        Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+        the multistep DPMSolver.
+
+        Args:
+            model_output (`torch.Tensor`):
+                The direct output from learned diffusion model.
+            timestep (`int`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.Tensor`):
+                A current instance of a sample created by the diffusion process.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            variance_noise (`torch.Tensor`):
+                Alternative to generating noise with `generator` by directly providing the noise for the variance
+                itself. Useful for methods such as [`LEdits++`].
+            return_dict (`bool`):
+                Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+        Returns:
+            [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+                tuple is returned where the first element is the sample tensor.
+
+        """
+        if self.num_inference_steps is None:
+            raise ValueError(
+                "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+            )
+
+        if self.step_index is None:
+            self._init_step_index(timestep)
+
+        # Improve numerical stability for small number of steps
+        lower_order_final = (self.step_index == len(self.timesteps) - 1) and (
+            self.config.euler_at_final
+            or (self.config.lower_order_final and len(self.timesteps) < 15)
+            or self.config.final_sigmas_type == "zero"
+        )
+        lower_order_second = (
+            (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
+        )
+
+        model_output = self.convert_model_output(model_output, sample=sample)
+        for i in range(self.config.solver_order - 1):
+            self.model_outputs[i] = self.model_outputs[i + 1]
+        self.model_outputs[-1] = model_output
+
+        # Upcast to avoid precision issues when computing prev_sample
+        sample = sample.to(torch.float32)
+        if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"] and variance_noise is None:
+            noise = randn_tensor(
+                model_output.shape, generator=generator, device=model_output.device, dtype=torch.float32
+            )
+        elif self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]:
+            noise = variance_noise.to(device=model_output.device, dtype=torch.float32)
+        else:
+            noise = None
+
+        if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
+            prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise)
+        elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
+            prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise)
+        else:
+            prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample, noise=noise)
+
+        if self.lower_order_nums < self.config.solver_order:
+            self.lower_order_nums += 1
+
+        # Cast sample back to expected dtype
+        prev_sample = prev_sample.to(model_output.dtype)
+
+        # upon completion increase step index by one
+        self._step_index += 1
+
+        if not return_dict:
+            return (prev_sample,)
+
+        return SchedulerOutput(prev_sample=prev_sample)
+
+    def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        """
+        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+        current timestep.
+
+        Args:
+            sample (`torch.Tensor`):
+                The input sample.
+
+        Returns:
+            `torch.Tensor`:
+                A scaled input sample.
+        """
+        return sample
+
+    def add_noise(
+        self,
+        original_samples: torch.Tensor,
+        noise: torch.Tensor,
+        timesteps: torch.IntTensor,
+    ) -> torch.Tensor:
+        # Make sure sigmas and timesteps have the same device and dtype as original_samples
+        sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+        if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+            # mps does not support float64
+            schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+        else:
+            schedule_timesteps = self.timesteps.to(original_samples.device)
+            timesteps = timesteps.to(original_samples.device)
+
+        # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index
+        if self.begin_index is None:
+            step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+        elif self.step_index is not None:
+            # add_noise is called after first denoising step (for inpainting)
+            step_indices = [self.step_index] * timesteps.shape[0]
+        else:
+            # add noise is called before first denoising step to create initial latent(img2img)
+            step_indices = [self.begin_index] * timesteps.shape[0]
+
+        sigma = sigmas[step_indices].flatten()
+        while len(sigma.shape) < len(original_samples.shape):
+            sigma = sigma.unsqueeze(-1)
+
+        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+        noisy_samples = alpha_t * original_samples + sigma_t * noise
+        return noisy_samples
+
+    def __len__(self):
+        return self.config.num_train_timesteps

omnigen2/schedulers/scheduling_flow_match_euler_discrete.py

@@ -0,0 +1,229 @@
+# Copyright 2024 Stability AI, Katherine Crowson and The HuggingFace 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import BaseOutput, logging
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class FlowMatchEulerDiscreteSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+            denoising loop.
+    """
+
+    prev_sample: torch.FloatTensor
+
+
+class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
+    """
+    Euler scheduler.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+    methods the library implements for all schedulers such as loading and saving.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        timestep_spacing (`str`, defaults to `"linspace"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        shift (`float`, defaults to 1.0):
+            The shift value for the timestep schedule.
+    """
+
+    _compatibles = []
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        dynamic_time_shift: bool = True
+    ):
+        timesteps = torch.linspace(0, 1, num_train_timesteps + 1, dtype=torch.float32)[:-1]
+
+        self.timesteps = timesteps
+
+        self._step_index = None
+        self._begin_index = None
+
+    @property
+    def step_index(self):
+        """
+        The index counter for current timestep. It will increase 1 after each scheduler step.
+        """
+        return self._step_index
+
+    @property
+    def begin_index(self):
+        """
+        The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+        """
+        return self._begin_index
+
+    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+    def set_begin_index(self, begin_index: int = 0):
+        """
+        Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+        Args:
+            begin_index (`int`):
+                The begin index for the scheduler.
+        """
+        self._begin_index = begin_index
+
+    def index_for_timestep(self, timestep, schedule_timesteps=None):
+        if schedule_timesteps is None:
+            schedule_timesteps = self._timesteps
+
+        indices = (schedule_timesteps == timestep).nonzero()
+
+        # The sigma index that is taken for the **very** first `step`
+        # is always the second index (or the last index if there is only 1)
+        # This way we can ensure we don't accidentally skip a sigma in
+        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+        pos = 1 if len(indices) > 1 else 0
+
+        return indices[pos].item()
+    
+    # def time_shift(self, mu: float, sigma: float, t: torch.Tensor):
+    #     return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int = None,
+        device: Union[str, torch.device] = None,
+        timesteps: Optional[List[float]] = None,
+        num_tokens: Optional[int] = None
+    ):
+        """
+        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+        Args:
+            num_inference_steps (`int`):
+                The number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, *optional*):
+                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        """
+
+        if timesteps is None:
+            self.num_inference_steps = num_inference_steps
+            timesteps = np.linspace(0, 1, num_inference_steps + 1, dtype=np.float32)[:-1]
+            if self.config.dynamic_time_shift and num_tokens is not None:
+                m = np.sqrt(num_tokens) / 40 # when input resolution is 320 * 320, m = 1, when input resolution is 1024 * 1024, m = 3.2
+                timesteps = timesteps / (m - m * timesteps + timesteps)
+
+        timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device)
+        _timesteps = torch.cat([timesteps, torch.ones(1, device=timesteps.device)])
+        
+        self.timesteps = timesteps
+        self._timesteps = _timesteps
+        self._step_index = None
+        self._begin_index = None
+
+    def _init_step_index(self, timestep):
+        if self.begin_index is None:
+            if isinstance(timestep, torch.Tensor):
+                timestep = timestep.to(self.timesteps.device)
+            self._step_index = self.index_for_timestep(timestep)
+        else:
+            self._step_index = self._begin_index
+
+    def step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: Union[float, torch.FloatTensor],
+        sample: torch.FloatTensor,
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+    ) -> Union[FlowMatchEulerDiscreteSchedulerOutput, Tuple]:
+        """
+        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+        process from the learned model outputs (most often the predicted noise).
+
+        Args:
+            model_output (`torch.FloatTensor`):
+                The direct output from learned diffusion model.
+            timestep (`float`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.FloatTensor`):
+                A current instance of a sample created by the diffusion process.
+            s_churn (`float`):
+            s_tmin  (`float`):
+            s_tmax  (`float`):
+            s_noise (`float`, defaults to 1.0):
+                Scaling factor for noise added to the sample.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            return_dict (`bool`):
+                Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or
+                tuple.
+
+        Returns:
+            [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is
+                returned, otherwise a tuple is returned where the first element is the sample tensor.
+        """
+
+        if (
+            isinstance(timestep, int)
+            or isinstance(timestep, torch.IntTensor)
+            or isinstance(timestep, torch.LongTensor)
+        ):
+            raise ValueError(
+                (
+                    "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+                    " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
+                    " one of the `scheduler.timesteps` as a timestep."
+                ),
+            )
+
+        if self.step_index is None:
+            self._init_step_index(timestep)
+        # Upcast to avoid precision issues when computing prev_sample
+        sample = sample.to(torch.float32)
+        t = self._timesteps[self.step_index]
+        t_next = self._timesteps[self.step_index + 1]
+
+        prev_sample = sample + (t_next - t) * model_output
+
+        # Cast sample back to model compatible dtype
+        prev_sample = prev_sample.to(model_output.dtype)
+
+        # upon completion increase step index by one
+        self._step_index += 1
+
+        if not return_dict:
+            return (prev_sample,)
+
+        return FlowMatchEulerDiscreteSchedulerOutput(prev_sample=prev_sample)
+
+    def __len__(self):
+        return self.config.num_train_timesteps

omnigen2/taylorseer_utils/__init__.py

@@ -0,0 +1,51 @@
+# Copied from https://github.com/Shenyi-Z/TaylorSeer/blob/main/TaylorSeers-xDiT/taylorseer_flux/taylorseer_utils/__init__.py
+
+from typing import Dict 
+import torch
+import math
+
+def derivative_approximation(cache_dic: Dict, current: Dict, feature: torch.Tensor):
+    """
+    Compute derivative approximation.
+    
+    :param cache_dic: Cache dictionary
+    :param current: Information of the current step
+    """
+    difference_distance = current['activated_steps'][-1] - current['activated_steps'][-2]
+
+    updated_taylor_factors = {}
+    updated_taylor_factors[0] = feature
+
+    for i in range(cache_dic['max_order']):
+        if (cache_dic['cache'][-1][current['stream']][current['layer']][current['module']].get(i, None) is not None) and (current['step'] > cache_dic['first_enhance'] - 2):
+            updated_taylor_factors[i + 1] = (updated_taylor_factors[i] - cache_dic['cache'][-1][current['stream']][current['layer']][current['module']][i]) / difference_distance
+        else:
+            break
+    
+    cache_dic['cache'][-1][current['stream']][current['layer']][current['module']] = updated_taylor_factors
+
+def taylor_formula(cache_dic: Dict, current: Dict) -> torch.Tensor: 
+    """
+    Compute Taylor expansion error.
+    
+    :param cache_dic: Cache dictionary
+    :param current: Information of the current step
+    """
+    x = current['step'] - current['activated_steps'][-1]
+    #x = current['t'] - current['activated_times'][-1]
+    output = 0
+
+    for i in range(len(cache_dic['cache'][-1][current['stream']][current['layer']][current['module']])):
+        output += (1 / math.factorial(i)) * cache_dic['cache'][-1][current['stream']][current['layer']][current['module']][i] * (x ** i)
+    
+    return output
+
+def taylor_cache_init(cache_dic: Dict, current: Dict):
+    """
+    Initialize Taylor cache and allocate storage for different-order derivatives in the Taylor cache.
+    
+    :param cache_dic: Cache dictionary
+    :param current: Information of the current step
+    """
+    if (current['step'] == 0) and (cache_dic['taylor_cache']):
+        cache_dic['cache'][-1][current['stream']][current['layer']][current['module']] = {}

omnigen2/training_utils.py

@@ -0,0 +1,645 @@
+import contextlib
+import copy
+import gc
+import math
+import random
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.models import UNet2DConditionModel
+from diffusers.schedulers import SchedulerMixin
+from diffusers.utils import (
+    convert_state_dict_to_diffusers,
+    convert_state_dict_to_peft,
+    deprecate,
+    is_peft_available,
+    is_torch_npu_available,
+    is_torchvision_available,
+    is_transformers_available,
+)
+
+
+if is_transformers_available():
+    import transformers
+
+    if transformers.integrations.deepspeed.is_deepspeed_zero3_enabled():
+        import deepspeed
+
+if is_peft_available():
+    from peft import set_peft_model_state_dict
+
+if is_torchvision_available():
+    from torchvision import transforms
+
+if is_torch_npu_available():
+    import torch_npu  # noqa: F401
+
+
+def set_seed(seed: int):
+    """
+    Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`.
+
+    Args:
+        seed (`int`): The seed to set.
+
+    Returns:
+        `None`
+    """
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if is_torch_npu_available():
+        torch.npu.manual_seed_all(seed)
+    else:
+        torch.cuda.manual_seed_all(seed)
+        # ^^ safe to call this function even if cuda is not available
+
+
+def compute_snr(noise_scheduler, timesteps):
+    """
+    Computes SNR as per
+    https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
+    for the given timesteps using the provided noise scheduler.
+
+    Args:
+        noise_scheduler (`NoiseScheduler`):
+            An object containing the noise schedule parameters, specifically `alphas_cumprod`, which is used to compute
+            the SNR values.
+        timesteps (`torch.Tensor`):
+            A tensor of timesteps for which the SNR is computed.
+
+    Returns:
+        `torch.Tensor`: A tensor containing the computed SNR values for each timestep.
+    """
+    alphas_cumprod = noise_scheduler.alphas_cumprod
+    sqrt_alphas_cumprod = alphas_cumprod**0.5
+    sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
+
+    # Expand the tensors.
+    # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
+    sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
+    while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
+        sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
+    alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
+
+    sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
+    while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
+        sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
+    sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
+
+    # Compute SNR.
+    snr = (alpha / sigma) ** 2
+    return snr
+
+
+def resolve_interpolation_mode(interpolation_type: str):
+    """
+    Maps a string describing an interpolation function to the corresponding torchvision `InterpolationMode` enum. The
+    full list of supported enums is documented at
+    https://pytorch.org/vision/0.9/transforms.html#torchvision.transforms.functional.InterpolationMode.
+
+    Args:
+        interpolation_type (`str`):
+            A string describing an interpolation method. Currently, `bilinear`, `bicubic`, `box`, `nearest`,
+            `nearest_exact`, `hamming`, and `lanczos` are supported, corresponding to the supported interpolation modes
+            in torchvision.
+
+    Returns:
+        `torchvision.transforms.InterpolationMode`: an `InterpolationMode` enum used by torchvision's `resize`
+        transform.
+    """
+    if not is_torchvision_available():
+        raise ImportError(
+            "Please make sure to install `torchvision` to be able to use the `resolve_interpolation_mode()` function."
+        )
+
+    if interpolation_type == "bilinear":
+        interpolation_mode = transforms.InterpolationMode.BILINEAR
+    elif interpolation_type == "bicubic":
+        interpolation_mode = transforms.InterpolationMode.BICUBIC
+    elif interpolation_type == "box":
+        interpolation_mode = transforms.InterpolationMode.BOX
+    elif interpolation_type == "nearest":
+        interpolation_mode = transforms.InterpolationMode.NEAREST
+    elif interpolation_type == "nearest_exact":
+        interpolation_mode = transforms.InterpolationMode.NEAREST_EXACT
+    elif interpolation_type == "hamming":
+        interpolation_mode = transforms.InterpolationMode.HAMMING
+    elif interpolation_type == "lanczos":
+        interpolation_mode = transforms.InterpolationMode.LANCZOS
+    else:
+        raise ValueError(
+            f"The given interpolation mode {interpolation_type} is not supported. Currently supported interpolation"
+            f" modes are `bilinear`, `bicubic`, `box`, `nearest`, `nearest_exact`, `hamming`, and `lanczos`."
+        )
+
+    return interpolation_mode
+
+
+def compute_dream_and_update_latents(
+    unet: UNet2DConditionModel,
+    noise_scheduler: SchedulerMixin,
+    timesteps: torch.Tensor,
+    noise: torch.Tensor,
+    noisy_latents: torch.Tensor,
+    target: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    dream_detail_preservation: float = 1.0,
+) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+    """
+    Implements "DREAM (Diffusion Rectification and Estimation-Adaptive Models)" from http://arxiv.org/abs/2312.00210.
+    DREAM helps align training with sampling to help training be more efficient and accurate at the cost of an extra
+    forward step without gradients.
+
+    Args:
+        `unet`: The state unet to use to make a prediction.
+        `noise_scheduler`: The noise scheduler used to add noise for the given timestep.
+        `timesteps`: The timesteps for the noise_scheduler to user.
+        `noise`: A tensor of noise in the shape of noisy_latents.
+        `noisy_latents`: Previously noise latents from the training loop.
+        `target`: The ground-truth tensor to predict after eps is removed.
+        `encoder_hidden_states`: Text embeddings from the text model.
+        `dream_detail_preservation`: A float value that indicates detail preservation level.
+          See reference.
+
+    Returns:
+        `tuple[torch.Tensor, torch.Tensor]`: Adjusted noisy_latents and target.
+    """
+    alphas_cumprod = noise_scheduler.alphas_cumprod.to(timesteps.device)[timesteps, None, None, None]
+    sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
+
+    # The paper uses lambda = sqrt(1 - alpha) ** p, with p = 1 in their experiments.
+    dream_lambda = sqrt_one_minus_alphas_cumprod**dream_detail_preservation
+
+    pred = None
+    with torch.no_grad():
+        pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
+
+    _noisy_latents, _target = (None, None)
+    if noise_scheduler.config.prediction_type == "epsilon":
+        predicted_noise = pred
+        delta_noise = (noise - predicted_noise).detach()
+        delta_noise.mul_(dream_lambda)
+        _noisy_latents = noisy_latents.add(sqrt_one_minus_alphas_cumprod * delta_noise)
+        _target = target.add(delta_noise)
+    elif noise_scheduler.config.prediction_type == "v_prediction":
+        raise NotImplementedError("DREAM has not been implemented for v-prediction")
+    else:
+        raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
+
+    return _noisy_latents, _target
+
+
+def unet_lora_state_dict(unet: UNet2DConditionModel) -> Dict[str, torch.Tensor]:
+    r"""
+    Returns:
+        A state dict containing just the LoRA parameters.
+    """
+    lora_state_dict = {}
+
+    for name, module in unet.named_modules():
+        if hasattr(module, "set_lora_layer"):
+            lora_layer = getattr(module, "lora_layer")
+            if lora_layer is not None:
+                current_lora_layer_sd = lora_layer.state_dict()
+                for lora_layer_matrix_name, lora_param in current_lora_layer_sd.items():
+                    # The matrix name can either be "down" or "up".
+                    lora_state_dict[f"{name}.lora.{lora_layer_matrix_name}"] = lora_param
+
+    return lora_state_dict
+
+
+def cast_training_params(model: Union[torch.nn.Module, List[torch.nn.Module]], dtype=torch.float32):
+    """
+    Casts the training parameters of the model to the specified data type.
+
+    Args:
+        model: The PyTorch model whose parameters will be cast.
+        dtype: The data type to which the model parameters will be cast.
+    """
+    if not isinstance(model, list):
+        model = [model]
+    for m in model:
+        for param in m.parameters():
+            # only upcast trainable parameters into fp32
+            if param.requires_grad:
+                param.data = param.to(dtype)
+
+
+def _set_state_dict_into_text_encoder(
+    lora_state_dict: Dict[str, torch.Tensor], prefix: str, text_encoder: torch.nn.Module
+):
+    """
+    Sets the `lora_state_dict` into `text_encoder` coming from `transformers`.
+
+    Args:
+        lora_state_dict: The state dictionary to be set.
+        prefix: String identifier to retrieve the portion of the state dict that belongs to `text_encoder`.
+        text_encoder: Where the `lora_state_dict` is to be set.
+    """
+
+    text_encoder_state_dict = {
+        f"{k.replace(prefix, '')}": v for k, v in lora_state_dict.items() if k.startswith(prefix)
+    }
+    text_encoder_state_dict = convert_state_dict_to_peft(convert_state_dict_to_diffusers(text_encoder_state_dict))
+    set_peft_model_state_dict(text_encoder, text_encoder_state_dict, adapter_name="default")
+
+
+def compute_density_for_timestep_sampling(
+    weighting_scheme: str,
+    batch_size: int,
+    logit_mean: float = None,
+    logit_std: float = None,
+    mode_scale: float = None,
+    device: Union[torch.device, str] = "cpu",
+    generator: Optional[torch.Generator] = None,
+):
+    """
+    Compute the density for sampling the timesteps when doing SD3 training.
+
+    Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528.
+
+    SD3 paper reference: https://arxiv.org/abs/2403.03206v1.
+    """
+    if weighting_scheme == "logit_normal":
+        u = torch.normal(mean=logit_mean, std=logit_std, size=(batch_size,), device=device, generator=generator)
+        u = torch.nn.functional.sigmoid(u)
+    elif weighting_scheme == "mode":
+        u = torch.rand(size=(batch_size,), device=device, generator=generator)
+        u = 1 - u - mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u)
+    else:
+        u = torch.rand(size=(batch_size,), device=device, generator=generator)
+    return u
+
+
+def compute_loss_weighting_for_sd3(weighting_scheme: str, sigmas=None):
+    """
+    Computes loss weighting scheme for SD3 training.
+
+    Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528.
+
+    SD3 paper reference: https://arxiv.org/abs/2403.03206v1.
+    """
+    if weighting_scheme == "sigma_sqrt":
+        weighting = (sigmas**-2.0).float()
+    elif weighting_scheme == "cosmap":
+        bot = 1 - 2 * sigmas + 2 * sigmas**2
+        weighting = 2 / (math.pi * bot)
+    else:
+        weighting = torch.ones_like(sigmas)
+    return weighting
+
+
+def free_memory():
+    """
+    Runs garbage collection. Then clears the cache of the available accelerator.
+    """
+    gc.collect()
+
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    elif torch.backends.mps.is_available():
+        torch.mps.empty_cache()
+    elif is_torch_npu_available():
+        torch_npu.npu.empty_cache()
+    elif hasattr(torch, "xpu") and torch.xpu.is_available():
+        torch.xpu.empty_cache()
+
+
+# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
+class EMAModel:
+    """
+    Exponential Moving Average of models weights
+    """
+
+    def __init__(
+        self,
+        parameters: Iterable[torch.nn.Parameter],
+        decay: float = 0.9999,
+        min_decay: float = 0.0,
+        update_after_step: int = 0,
+        use_ema_warmup: bool = False,
+        inv_gamma: Union[float, int] = 1.0,
+        power: Union[float, int] = 2 / 3,
+        foreach: bool = False,
+        model_cls: Optional[Any] = None,
+        model_config: Dict[str, Any] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            parameters (Iterable[torch.nn.Parameter]): The parameters to track.
+            decay (float): The decay factor for the exponential moving average.
+            min_decay (float): The minimum decay factor for the exponential moving average.
+            update_after_step (int): The number of steps to wait before starting to update the EMA weights.
+            use_ema_warmup (bool): Whether to use EMA warmup.
+            inv_gamma (float):
+                Inverse multiplicative factor of EMA warmup. Default: 1. Only used if `use_ema_warmup` is True.
+            power (float): Exponential factor of EMA warmup. Default: 2/3. Only used if `use_ema_warmup` is True.
+            foreach (bool): Use torch._foreach functions for updating shadow parameters. Should be faster.
+            device (Optional[Union[str, torch.device]]): The device to store the EMA weights on. If None, the EMA
+                        weights will be stored on CPU.
+
+        @crowsonkb's notes on EMA Warmup:
+            If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are good values for models you plan
+            to train for a million or more steps (reaches decay factor 0.999 at 31.6K steps, 0.9999 at 1M steps),
+            gamma=1, power=3/4 for models you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999
+            at 215.4k steps).
+        """
+
+        if isinstance(parameters, torch.nn.Module):
+            deprecation_message = (
+                "Passing a `torch.nn.Module` to `ExponentialMovingAverage` is deprecated. "
+                "Please pass the parameters of the module instead."
+            )
+            deprecate(
+                "passing a `torch.nn.Module` to `ExponentialMovingAverage`",
+                "1.0.0",
+                deprecation_message,
+                standard_warn=False,
+            )
+            parameters = parameters.parameters()
+
+            # set use_ema_warmup to True if a torch.nn.Module is passed for backwards compatibility
+            use_ema_warmup = True
+
+        if kwargs.get("max_value", None) is not None:
+            deprecation_message = "The `max_value` argument is deprecated. Please use `decay` instead."
+            deprecate("max_value", "1.0.0", deprecation_message, standard_warn=False)
+            decay = kwargs["max_value"]
+
+        if kwargs.get("min_value", None) is not None:
+            deprecation_message = "The `min_value` argument is deprecated. Please use `min_decay` instead."
+            deprecate("min_value", "1.0.0", deprecation_message, standard_warn=False)
+            min_decay = kwargs["min_value"]
+
+        parameters = list(parameters)
+        self.shadow_params = [p.clone().detach() for p in parameters]
+
+        if kwargs.get("device", None) is not None:
+            deprecation_message = "The `device` argument is deprecated. Please use `to` instead."
+            deprecate("device", "1.0.0", deprecation_message, standard_warn=False)
+            self.to(device=kwargs["device"])
+
+        self.temp_stored_params = None
+
+        self.decay = decay
+        self.min_decay = min_decay
+        self.update_after_step = update_after_step
+        self.use_ema_warmup = use_ema_warmup
+        self.inv_gamma = inv_gamma
+        self.power = power
+        self.optimization_step = 0
+        self.cur_decay_value = None  # set in `step()`
+        self.foreach = foreach
+
+        self.model_cls = model_cls
+        self.model_config = model_config
+
+    @classmethod
+    def from_pretrained(cls, path, model_cls, foreach=False) -> "EMAModel":
+        _, ema_kwargs = model_cls.from_config(path, return_unused_kwargs=True)
+        model = model_cls.from_pretrained(path)
+
+        ema_model = cls(model.parameters(), model_cls=model_cls, model_config=model.config, foreach=foreach)
+
+        ema_model.load_state_dict(ema_kwargs)
+        return ema_model
+
+    def save_pretrained(self, path):
+        if self.model_cls is None:
+            raise ValueError("`save_pretrained` can only be used if `model_cls` was defined at __init__.")
+
+        if self.model_config is None:
+            raise ValueError("`save_pretrained` can only be used if `model_config` was defined at __init__.")
+
+        model = self.model_cls.from_config(self.model_config)
+        state_dict = self.state_dict()
+        state_dict.pop("shadow_params", None)
+
+        model.register_to_config(**state_dict)
+        self.copy_to(model.parameters())
+        model.save_pretrained(path)
+
+    def get_decay(self, optimization_step: int) -> float:
+        """
+        Compute the decay factor for the exponential moving average.
+        """
+        step = max(0, optimization_step - self.update_after_step - 1)
+
+        if step <= 0:
+            return 0.0
+
+        if self.use_ema_warmup:
+            cur_decay_value = 1 - (1 + step / self.inv_gamma) ** -self.power
+        else:
+            cur_decay_value = (1 + step) / (10 + step)
+
+        cur_decay_value = min(cur_decay_value, self.decay)
+        # make sure decay is not smaller than min_decay
+        cur_decay_value = max(cur_decay_value, self.min_decay)
+        return cur_decay_value
+
+    @torch.no_grad()
+    def step(self, parameters: Iterable[torch.nn.Parameter]):
+        if isinstance(parameters, torch.nn.Module):
+            deprecation_message = (
+                "Passing a `torch.nn.Module` to `ExponentialMovingAverage.step` is deprecated. "
+                "Please pass the parameters of the module instead."
+            )
+            deprecate(
+                "passing a `torch.nn.Module` to `ExponentialMovingAverage.step`",
+                "1.0.0",
+                deprecation_message,
+                standard_warn=False,
+            )
+            parameters = parameters.parameters()
+
+        parameters = list(parameters)
+
+        self.optimization_step += 1
+
+        # Compute the decay factor for the exponential moving average.
+        decay = self.get_decay(self.optimization_step)
+        self.cur_decay_value = decay
+        one_minus_decay = 1 - decay
+
+        context_manager = contextlib.nullcontext()
+
+        if self.foreach:
+            if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled():
+                context_manager = deepspeed.zero.GatheredParameters(parameters, modifier_rank=None)
+
+            with context_manager:
+                params_grad = [param for param in parameters if param.requires_grad]
+                s_params_grad = [
+                    s_param for s_param, param in zip(self.shadow_params, parameters) if param.requires_grad
+                ]
+
+                if len(params_grad) < len(parameters):
+                    torch._foreach_copy_(
+                        [s_param for s_param, param in zip(self.shadow_params, parameters) if not param.requires_grad],
+                        [param for param in parameters if not param.requires_grad],
+                        non_blocking=True,
+                    )
+
+                torch._foreach_sub_(
+                    s_params_grad, torch._foreach_sub(s_params_grad, params_grad), alpha=one_minus_decay
+                )
+
+        else:
+            for s_param, param in zip(self.shadow_params, parameters):
+                if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled():
+                    context_manager = deepspeed.zero.GatheredParameters(param, modifier_rank=None)
+
+                with context_manager:
+                    if param.requires_grad:
+                        # print(f"{s_param.shape=} {param.shape=}")
+                        s_param.sub_(one_minus_decay * (s_param - param))
+                    else:
+                        s_param.copy_(param)
+
+    def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None:
+        """
+        Copy current averaged parameters into given collection of parameters.
+
+        Args:
+            parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+                updated with the stored moving averages. If `None`, the parameters with which this
+                `ExponentialMovingAverage` was initialized will be used.
+        """
+        parameters = list(parameters)
+        if self.foreach:
+            torch._foreach_copy_(
+                [param.data for param in parameters],
+                [s_param.to(param.device).data for s_param, param in zip(self.shadow_params, parameters)],
+            )
+        else:
+            for s_param, param in zip(self.shadow_params, parameters):
+                param.data.copy_(s_param.to(param.device).data)
+
+    def pin_memory(self) -> None:
+        r"""
+        Move internal buffers of the ExponentialMovingAverage to pinned memory. Useful for non-blocking transfers for
+        offloading EMA params to the host.
+        """
+
+        self.shadow_params = [p.pin_memory() for p in self.shadow_params]
+
+    def to(self, device=None, dtype=None, non_blocking=False) -> None:
+        r"""
+        Move internal buffers of the ExponentialMovingAverage to `device`.
+
+        Args:
+            device: like `device` argument to `torch.Tensor.to`
+        """
+        # .to() on the tensors handles None correctly
+        self.shadow_params = [
+            p.to(device=device, dtype=dtype, non_blocking=non_blocking)
+            if p.is_floating_point()
+            else p.to(device=device, non_blocking=non_blocking)
+            for p in self.shadow_params
+        ]
+
+    def state_dict(self) -> dict:
+        r"""
+        Returns the state of the ExponentialMovingAverage as a dict. This method is used by accelerate during
+        checkpointing to save the ema state dict.
+        """
+        # Following PyTorch conventions, references to tensors are returned:
+        # "returns a reference to the state and not its copy!" -
+        # https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict
+        return {
+            "decay": self.decay,
+            "min_decay": self.min_decay,
+            "optimization_step": self.optimization_step,
+            "update_after_step": self.update_after_step,
+            "use_ema_warmup": self.use_ema_warmup,
+            "inv_gamma": self.inv_gamma,
+            "power": self.power,
+            "shadow_params": self.shadow_params,
+        }
+
+    def store(self, parameters: Iterable[torch.nn.Parameter]) -> None:
+        r"""
+        Saves the current parameters for restoring later.
+
+        Args:
+            parameters: Iterable of `torch.nn.Parameter`. The parameters to be temporarily stored.
+        """
+        self.temp_stored_params = [param.detach().cpu().clone() for param in parameters]
+
+    def restore(self, parameters: Iterable[torch.nn.Parameter]) -> None:
+        r"""
+        Restore the parameters stored with the `store` method. Useful to validate the model with EMA parameters
+        without: affecting the original optimization process. Store the parameters before the `copy_to()` method. After
+        validation (or model saving), use this to restore the former parameters.
+
+        Args:
+            parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+                updated with the stored parameters. If `None`, the parameters with which this
+                `ExponentialMovingAverage` was initialized will be used.
+        """
+
+        if self.temp_stored_params is None:
+            raise RuntimeError("This ExponentialMovingAverage has no `store()`ed weights to `restore()`")
+        if self.foreach:
+            torch._foreach_copy_(
+                [param.data for param in parameters], [c_param.data for c_param in self.temp_stored_params]
+            )
+        else:
+            for c_param, param in zip(self.temp_stored_params, parameters):
+                param.data.copy_(c_param.data)
+
+        # Better memory-wise.
+        self.temp_stored_params = None
+
+    def load_state_dict(self, state_dict: dict) -> None:
+        r"""
+        Loads the ExponentialMovingAverage state. This method is used by accelerate during checkpointing to save the
+        ema state dict.
+
+        Args:
+            state_dict (dict): EMA state. Should be an object returned
+                from a call to :meth:`state_dict`.
+        """
+        # deepcopy, to be consistent with module API
+        state_dict = copy.deepcopy(state_dict)
+
+        self.decay = state_dict.get("decay", self.decay)
+        if self.decay < 0.0 or self.decay > 1.0:
+            raise ValueError("Decay must be between 0 and 1")
+
+        self.min_decay = state_dict.get("min_decay", self.min_decay)
+        if not isinstance(self.min_decay, float):
+            raise ValueError("Invalid min_decay")
+
+        self.optimization_step = state_dict.get("optimization_step", self.optimization_step)
+        if not isinstance(self.optimization_step, int):
+            raise ValueError("Invalid optimization_step")
+
+        self.update_after_step = state_dict.get("update_after_step", self.update_after_step)
+        if not isinstance(self.update_after_step, int):
+            raise ValueError("Invalid update_after_step")
+
+        self.use_ema_warmup = state_dict.get("use_ema_warmup", self.use_ema_warmup)
+        if not isinstance(self.use_ema_warmup, bool):
+            raise ValueError("Invalid use_ema_warmup")
+
+        self.inv_gamma = state_dict.get("inv_gamma", self.inv_gamma)
+        if not isinstance(self.inv_gamma, (float, int)):
+            raise ValueError("Invalid inv_gamma")
+
+        self.power = state_dict.get("power", self.power)
+        if not isinstance(self.power, (float, int)):
+            raise ValueError("Invalid power")
+
+        shadow_params = state_dict.get("shadow_params", None)
+        if shadow_params is not None:
+            self.shadow_params = shadow_params
+            if not isinstance(self.shadow_params, list):
+                raise ValueError("shadow_params must be a list")
+            if not all(isinstance(p, torch.Tensor) for p in self.shadow_params):
+                raise ValueError("shadow_params must all be Tensors")

omnigen2/transport/__init__.py

@@ -0,0 +1,74 @@
+from .transport import ModelType, PathType, Sampler, Transport, WeightType
+
+
+def create_transport(
+    path_type="Linear",
+    prediction="velocity",
+    loss_weight=None,
+    train_eps=None,
+    sample_eps=None,
+    snr_type="uniform",
+    do_shift=True,
+    seq_len=1024,  # corresponding to 512x512
+    dynamic_time_shift: bool = False,
+    time_shift_version: str = "v1",
+):
+    """function for creating Transport object
+    **Note**: model prediction defaults to velocity
+    Args:
+    - path_type: type of path to use; default to linear
+    - learn_score: set model prediction to score
+    - learn_noise: set model prediction to noise
+    - velocity_weighted: weight loss by velocity weight
+    - likelihood_weighted: weight loss by likelihood weight
+    - train_eps: small epsilon for avoiding instability during training
+    - sample_eps: small epsilon for avoiding instability during sampling
+    """
+
+    if prediction == "noise":
+        model_type = ModelType.NOISE
+    elif prediction == "score":
+        model_type = ModelType.SCORE
+    else:
+        model_type = ModelType.VELOCITY
+
+    if loss_weight == "velocity":
+        loss_type = WeightType.VELOCITY
+    elif loss_weight == "likelihood":
+        loss_type = WeightType.LIKELIHOOD
+    else:
+        loss_type = WeightType.NONE
+
+    path_choice = {
+        "Linear": PathType.LINEAR,
+        "GVP": PathType.GVP,
+        "VP": PathType.VP,
+    }
+
+    path_type = path_choice[path_type]
+
+    if path_type in [PathType.VP]:
+        train_eps = 1e-5 if train_eps is None else train_eps
+        sample_eps = 1e-3 if train_eps is None else sample_eps
+    elif path_type in [PathType.GVP, PathType.LINEAR] and model_type != ModelType.VELOCITY:
+        train_eps = 1e-3 if train_eps is None else train_eps
+        sample_eps = 1e-3 if train_eps is None else sample_eps
+    else:  # velocity & [GVP, LINEAR] is stable everywhere
+        train_eps = 0
+        sample_eps = 0
+
+    # create flow state
+    state = Transport(
+        model_type=model_type,
+        path_type=path_type,
+        loss_type=loss_type,
+        train_eps=train_eps,
+        sample_eps=sample_eps,
+        snr_type=snr_type,
+        do_shift=do_shift,
+        seq_len=seq_len,
+        dynamic_time_shift=dynamic_time_shift,
+        time_shift_version=time_shift_version,
+    )
+
+    return state

omnigen2/transport/dpm_solver.py

@@ -0,0 +1,1386 @@
+# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+# This file is modified from https://github.com/PixArt-alpha/PixArt-sigma
+import os
+
+import torch
+from tqdm import tqdm
+
+
+class NoiseScheduleFlow:
+    def __init__(
+        self,
+        schedule="discrete_flow",
+    ):
+        """Create a wrapper class for the forward SDE (EDM type)."""
+        self.T = 1
+        self.t0 = 0.001
+        self.schedule = schedule  # ['continuous', 'discrete_flow']
+        self.total_N = 1000
+
+    def marginal_log_mean_coeff(self, t):
+        """
+        Compute log(alpha_t) of a given continuous-time label t in [0, T].
+        """
+        return torch.log(self.marginal_alpha(t))
+
+    def marginal_alpha(self, t):
+        """
+        Compute alpha_t of a given continuous-time label t in [0, T].
+        """
+        return 1 - t
+
+    @staticmethod
+    def marginal_std(t):
+        """
+        Compute sigma_t of a given continuous-time label t in [0, T].
+        """
+        return t
+
+    def marginal_lambda(self, t):
+        """
+        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
+        """
+        log_mean_coeff = self.marginal_log_mean_coeff(t)
+        log_std = torch.log(self.marginal_std(t))
+        return log_mean_coeff - log_std
+
+    @staticmethod
+    def inverse_lambda(lamb):
+        """
+        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
+        """
+        return torch.exp(-lamb)
+
+
+def model_wrapper(
+    model,
+    noise_schedule,
+    model_type="noise",
+    model_kwargs={},
+    guidance_type="uncond",
+    condition=None,
+    unconditional_condition=None,
+    guidance_scale=1.0,
+    interval_guidance=[0, 1.0],
+    classifier_fn=None,
+    classifier_kwargs={},
+):
+    """Create a wrapper function for the noise prediction model.
+
+    DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
+    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
+
+    We support four types of the diffusion model by setting `model_type`:
+
+        1. "noise": noise prediction model. (Trained by predicting noise).
+
+        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
+
+        3. "v": velocity prediction model. (Trained by predicting the velocity).
+            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
+
+            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
+                arXiv preprint arXiv:2202.00512 (2022).
+            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
+                arXiv preprint arXiv:2210.02303 (2022).
+
+        4. "score": marginal score function. (Trained by denoising score matching).
+            Note that the score function and the noise prediction model follows a simple relationship:
+            ```
+                noise(x_t, t) = -sigma_t * score(x_t, t)
+            ```
+
+    We support three types of guided sampling by DPMs by setting `guidance_type`:
+        1. "uncond": unconditional sampling by DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+
+        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+
+            The input `classifier_fn` has the following format:
+            ``
+                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
+            ``
+
+            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
+                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
+
+        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
+            ``
+            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
+
+            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
+                arXiv preprint arXiv:2207.12598 (2022).
+
+
+    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
+    or continuous-time labels (i.e. epsilon to T).
+
+    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
+    ``
+        def model_fn(x, t_continuous) -> noise:
+            t_input = get_model_input_time(t_continuous)
+            return noise_pred(model, x, t_input, **model_kwargs)
+    ``
+    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
+
+    ===============================================================
+
+    Args:
+        model: A diffusion model with the corresponding format described above.
+        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+        model_type: A `str`. The parameterization type of the diffusion model.
+                    "noise" or "x_start" or "v" or "score".
+        model_kwargs: A `dict`. A dict for the other inputs of the model function.
+        guidance_type: A `str`. The type of the guidance for sampling.
+                    "uncond" or "classifier" or "classifier-free".
+        condition: A pytorch tensor. The condition for the guided sampling.
+                    Only used for "classifier" or "classifier-free" guidance type.
+        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
+                    Only used for "classifier-free" guidance type.
+        guidance_scale: A `float`. The scale for the guided sampling.
+        classifier_fn: A classifier function. Only used for the classifier guidance.
+        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
+    Returns:
+        A noise prediction model that accepts the noised data and the continuous time as the inputs.
+    """
+
+    def get_model_input_time(t_continuous):
+        """
+        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
+        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
+        For continuous-time DPMs, we just use `t_continuous`.
+        """
+        if noise_schedule.schedule == "discrete":
+            return (t_continuous - 1.0 / noise_schedule.total_N) * noise_schedule.total_N
+        elif noise_schedule.schedule == "discrete_flow":
+            return t_continuous * noise_schedule.total_N
+        else:
+            return t_continuous
+
+    def noise_pred_fn(x, t_continuous, cond=None):
+        t_input = get_model_input_time(t_continuous)
+        if cond is None:
+            output = model(x, t_input, **model_kwargs)
+        else:
+            output = model(x, t_input, cond, **model_kwargs)
+        if model_type == "noise":
+            return output
+        elif model_type == "x_start":
+            alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
+            return (x - expand_dims(alpha_t, x.dim()) * output) / expand_dims(sigma_t, x.dim())
+        elif model_type == "v":
+            alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
+            return expand_dims(alpha_t, x.dim()) * output + expand_dims(sigma_t, x.dim()) * x
+        elif model_type == "score":
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            return -expand_dims(sigma_t, x.dim()) * output
+        elif model_type == "flow":
+            _, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
+            try:
+                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output + x
+            except:
+                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output[0] + x
+            return noise
+
+    def cond_grad_fn(x, t_input):
+        """
+        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
+        """
+        with torch.enable_grad():
+            x_in = x.detach().requires_grad_(True)
+            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
+            return torch.autograd.grad(log_prob.sum(), x_in)[0]
+
+    def model_fn(x, t_continuous):
+        """
+        The noise predicition model function that is used for DPM-Solver.
+        """
+        guidance_tp = guidance_type
+        if guidance_tp == "uncond":
+            return noise_pred_fn(x, t_continuous)
+        elif guidance_tp == "classifier":
+            assert classifier_fn is not None
+            t_input = get_model_input_time(t_continuous)
+            cond_grad = cond_grad_fn(x, t_input)
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            noise = noise_pred_fn(x, t_continuous)
+            return noise - guidance_scale * expand_dims(sigma_t, x.dim()) * cond_grad
+        elif guidance_tp == "classifier-free":
+            if (
+                guidance_scale == 1.0
+                or unconditional_condition is None
+                or not (interval_guidance[0] < t_continuous[0] < interval_guidance[1])
+            ):
+                return noise_pred_fn(x, t_continuous, cond=condition)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t_continuous] * 2)
+                c_in = torch.cat([unconditional_condition, condition])
+                try:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
+                except:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(2)
+                return noise_uncond + guidance_scale * (noise - noise_uncond)
+
+    assert model_type in ["noise", "x_start", "v", "score", "flow"]
+    assert guidance_type in [
+        "uncond",
+        "classifier",
+        "classifier-free",
+    ]
+    return model_fn
+
+
+class DPM_Solver:
+    def __init__(
+        self,
+        model_fn,
+        noise_schedule,
+        algorithm_type="dpmsolver++",
+        correcting_x0_fn=None,
+        correcting_xt_fn=None,
+        thresholding_max_val=1.0,
+        dynamic_thresholding_ratio=0.995,
+    ):
+        """Construct a DPM-Solver.
+
+        We support both DPM-Solver (`algorithm_type="dpmsolver"`) and DPM-Solver++ (`algorithm_type="dpmsolver++"`).
+
+        We also support the "dynamic thresholding" method in Imagen[1]. For pixel-space diffusion models, you
+        can set both `algorithm_type="dpmsolver++"` and `correcting_x0_fn="dynamic_thresholding"` to use the
+        dynamic thresholding. The "dynamic thresholding" can greatly improve the sample quality for pixel-space
+        DPMs with large guidance scales. Note that the thresholding method is **unsuitable** for latent-space
+        DPMs (such as stable-diffusion).
+
+        To support advanced algorithms in image-to-image applications, we also support corrector functions for
+        both x0 and xt.
+
+        Args:
+            model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
+                ``
+                def model_fn(x, t_continuous):
+                    return noise
+                ``
+                The shape of `x` is `(batch_size, **shape)`, and the shape of `t_continuous` is `(batch_size,)`.
+            noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+            algorithm_type: A `str`. Either "dpmsolver" or "dpmsolver++".
+            correcting_x0_fn: A `str` or a function with the following format:
+                ```
+                def correcting_x0_fn(x0, t):
+                    x0_new = ...
+                    return x0_new
+                ```
+                This function is to correct the outputs of the data prediction model at each sampling step. e.g.,
+                ```
+                x0_pred = data_pred_model(xt, t)
+                if correcting_x0_fn is not None:
+                    x0_pred = correcting_x0_fn(x0_pred, t)
+                xt_1 = update(x0_pred, xt, t)
+                ```
+                If `correcting_x0_fn="dynamic_thresholding"`, we use the dynamic thresholding proposed in Imagen[1].
+            correcting_xt_fn: A function with the following format:
+                ```
+                def correcting_xt_fn(xt, t, step):
+                    x_new = ...
+                    return x_new
+                ```
+                This function is to correct the intermediate samples xt at each sampling step. e.g.,
+                ```
+                xt = ...
+                xt = correcting_xt_fn(xt, t, step)
+                ```
+            thresholding_max_val: A `float`. The max value for thresholding.
+                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
+            dynamic_thresholding_ratio: A `float`. The ratio for dynamic thresholding (see Imagen[1] for details).
+                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
+
+        [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour,
+            Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models
+            with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
+        """
+        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
+        self.noise_schedule = noise_schedule
+        assert algorithm_type in ["dpmsolver", "dpmsolver++"]
+        self.algorithm_type = algorithm_type
+        if correcting_x0_fn == "dynamic_thresholding":
+            self.correcting_x0_fn = self.dynamic_thresholding_fn
+        else:
+            self.correcting_x0_fn = correcting_x0_fn
+        self.correcting_xt_fn = correcting_xt_fn
+        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
+        self.thresholding_max_val = thresholding_max_val
+        self.register_progress_bar()
+
+    def register_progress_bar(self, progress_fn=None):
+        """
+        Register a progress bar callback function
+
+        Args:
+            progress_fn: Callback function that takes current step and total steps as parameters
+        """
+        self.progress_fn = progress_fn if progress_fn is not None else lambda step, total: None
+
+    def update_progress(self, step, total_steps):
+        """
+        Update sampling progress
+
+        Args:
+            step: Current step number
+            total_steps: Total number of steps
+        """
+        if hasattr(self, "progress_fn"):
+            try:
+                self.progress_fn(step / total_steps, desc=f"Generating {step}/{total_steps}")
+            except:
+                self.progress_fn(step, total_steps)
+
+        else:
+            # If no progress_fn registered, use default empty function
+            pass
+
+    def dynamic_thresholding_fn(self, x0, t):
+        """
+        The dynamic thresholding method.
+        """
+        dims = x0.dim()
+        p = self.dynamic_thresholding_ratio
+        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
+        s = expand_dims(torch.maximum(s, self.thresholding_max_val * torch.ones_like(s).to(s.device)), dims)
+        x0 = torch.clamp(x0, -s, s) / s
+        return x0
+
+    def noise_prediction_fn(self, x, t):
+        """
+        Return the noise prediction model.
+        """
+        return self.model(x, t)
+
+    def data_prediction_fn(self, x, t):
+        """
+        Return the data prediction model (with corrector).
+        """
+        noise = self.noise_prediction_fn(x, t)
+        alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t)
+        x0 = (x - sigma_t * noise) / alpha_t
+        if self.correcting_x0_fn is not None:
+            x0 = self.correcting_x0_fn(x0, t)
+        return x0
+
+    def model_fn(self, x, t):
+        """
+        Convert the model to the noise prediction model or the data prediction model.
+        """
+        if self.algorithm_type == "dpmsolver++":
+            return self.data_prediction_fn(x, t)
+        else:
+            return self.noise_prediction_fn(x, t)
+
+    def get_time_steps(self, skip_type, t_T, t_0, N, device, shift=1.0):
+        """Compute the intermediate time steps for sampling.
+
+        Args:
+            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
+                - 'logSNR': uniform logSNR for the time steps.
+                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
+                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            N: A `int`. The total number of the spacing of the time steps.
+            device: A torch device.
+        Returns:
+            A pytorch tensor of the time steps, with the shape (N + 1,).
+        """
+        if skip_type == "logSNR":
+            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
+            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
+            logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device)
+            return self.noise_schedule.inverse_lambda(logSNR_steps)
+        elif skip_type == "time_uniform":
+            return torch.linspace(t_T, t_0, N + 1).to(device)
+        elif skip_type == "time_quadratic":
+            t_order = 2
+            t = torch.linspace(t_T ** (1.0 / t_order), t_0 ** (1.0 / t_order), N + 1).pow(t_order).to(device)
+            return t
+        elif skip_type == "time_uniform_flow":
+            betas = torch.linspace(t_T, t_0, N + 1).to(device)
+            sigmas = 1.0 - betas
+            sigmas = (shift * sigmas / (1 + (shift - 1) * sigmas)).flip(dims=[0])
+            return sigmas
+        else:
+            raise ValueError(
+                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'"
+            )
+
+    def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device):
+        """
+        Get the order of each step for sampling by the singlestep DPM-Solver.
+
+        We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
+        Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
+            - If order == 1:
+                We take `steps` of DPM-Solver-1 (i.e. DDIM).
+            - If order == 2:
+                - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
+                - If steps % 2 == 0, we use K steps of DPM-Solver-2.
+                - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
+            - If order == 3:
+                - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
+                - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
+                - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
+                - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
+
+        ============================================
+        Args:
+            order: A `int`. The max order for the solver (2 or 3).
+            steps: A `int`. The total number of function evaluations (NFE).
+            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
+                - 'logSNR': uniform logSNR for the time steps.
+                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
+                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            device: A torch device.
+        Returns:
+            orders: A list of the solver order of each step.
+        """
+        if order == 3:
+            K = steps // 3 + 1
+            if steps % 3 == 0:
+                orders = [3,] * (
+                    K - 2
+                ) + [2, 1]
+            elif steps % 3 == 1:
+                orders = [3,] * (
+                    K - 1
+                ) + [1]
+            else:
+                orders = [3,] * (
+                    K - 1
+                ) + [2]
+        elif order == 2:
+            if steps % 2 == 0:
+                K = steps // 2
+                orders = [
+                    2,
+                ] * K
+            else:
+                K = steps // 2 + 1
+                orders = [2,] * (
+                    K - 1
+                ) + [1]
+        elif order == 1:
+            K = 1
+            orders = [
+                1,
+            ] * steps
+        else:
+            raise ValueError("'order' must be '1' or '2' or '3'.")
+        if skip_type == "logSNR":
+            # To reproduce the results in DPM-Solver paper
+            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
+        else:
+            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[
+                torch.cumsum(
+                    torch.tensor(
+                        [
+                            0,
+                        ]
+                        + orders
+                    ),
+                    0,
+                ).to(device)
+            ]
+        return timesteps_outer, orders
+
+    def denoise_to_zero_fn(self, x, s):
+        """
+        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
+        """
+        return self.data_prediction_fn(x, s)
+
+    def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
+        """
+        DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        ns = self.noise_schedule
+        dims = x.dim()
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t)
+        sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_t = sigma_t / sigma_s * x - alpha_t * phi_1 * model_s
+            if return_intermediate:
+                return x_t, {"model_s": model_s}
+            else:
+                return x_t
+        else:
+            phi_1 = torch.expm1(h)
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_t = torch.exp(log_alpha_t - log_alpha_s) * x - (sigma_t * phi_1) * model_s
+            if return_intermediate:
+                return x_t, {"model_s": model_s}
+            else:
+                return x_t
+
+    def singlestep_dpm_solver_second_update(
+        self, x, s, t, r1=0.5, model_s=None, return_intermediate=False, solver_type="dpmsolver"
+    ):
+        """
+        Singlestep solver DPM-Solver-2 from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            r1: A `float`. The hyperparameter of the second-order solver.
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}")
+        if r1 is None:
+            r1 = 0.5
+        ns = self.noise_schedule
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        lambda_s1 = lambda_s + r1 * h
+        s1 = ns.inverse_lambda(lambda_s1)
+        log_alpha_s, log_alpha_s1, log_alpha_t = (
+            ns.marginal_log_mean_coeff(s),
+            ns.marginal_log_mean_coeff(s1),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t)
+        alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_11 = torch.expm1(-r1 * h)
+            phi_1 = torch.expm1(-h)
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
+            model_s1 = self.model_fn(x_s1, s1)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    - (0.5 / r1) * (alpha_t * phi_1) * (model_s1 - model_s)
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (1.0 / r1) * (alpha_t * (phi_1 / h + 1.0)) * (model_s1 - model_s)
+                )
+        else:
+            phi_11 = torch.expm1(r1 * h)
+            phi_1 = torch.expm1(h)
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_s1 = torch.exp(log_alpha_s1 - log_alpha_s) * x - (sigma_s1 * phi_11) * model_s
+            model_s1 = self.model_fn(x_s1, s1)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    torch.exp(log_alpha_t - log_alpha_s) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (0.5 / r1) * (sigma_t * phi_1) * (model_s1 - model_s)
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    torch.exp(log_alpha_t - log_alpha_s) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (1.0 / r1) * (sigma_t * (phi_1 / h - 1.0)) * (model_s1 - model_s)
+                )
+        if return_intermediate:
+            return x_t, {"model_s": model_s, "model_s1": model_s1}
+        else:
+            return x_t
+
+    def singlestep_dpm_solver_third_update(
+        self,
+        x,
+        s,
+        t,
+        r1=1.0 / 3.0,
+        r2=2.0 / 3.0,
+        model_s=None,
+        model_s1=None,
+        return_intermediate=False,
+        solver_type="dpmsolver",
+    ):
+        """
+        Singlestep solver DPM-Solver-3 from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            r1: A `float`. The hyperparameter of the third-order solver.
+            r2: A `float`. The hyperparameter of the third-order solver.
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
+                If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}")
+        if r1 is None:
+            r1 = 1.0 / 3.0
+        if r2 is None:
+            r2 = 2.0 / 3.0
+        ns = self.noise_schedule
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        lambda_s1 = lambda_s + r1 * h
+        lambda_s2 = lambda_s + r2 * h
+        s1 = ns.inverse_lambda(lambda_s1)
+        s2 = ns.inverse_lambda(lambda_s2)
+        log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = (
+            ns.marginal_log_mean_coeff(s),
+            ns.marginal_log_mean_coeff(s1),
+            ns.marginal_log_mean_coeff(s2),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_s, sigma_s1, sigma_s2, sigma_t = (
+            ns.marginal_std(s),
+            ns.marginal_std(s1),
+            ns.marginal_std(s2),
+            ns.marginal_std(t),
+        )
+        alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t)
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_11 = torch.expm1(-r1 * h)
+            phi_12 = torch.expm1(-r2 * h)
+            phi_1 = torch.expm1(-h)
+            phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.0
+            phi_2 = phi_1 / h + 1.0
+            phi_3 = phi_2 / h - 0.5
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            if model_s1 is None:
+                x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
+                model_s1 = self.model_fn(x_s1, s1)
+            x_s2 = (
+                (sigma_s2 / sigma_s) * x
+                - (alpha_s2 * phi_12) * model_s
+                + r2 / r1 * (alpha_s2 * phi_22) * (model_s1 - model_s)
+            )
+            model_s2 = self.model_fn(x_s2, s2)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (1.0 / r2) * (alpha_t * phi_2) * (model_s2 - model_s)
+                )
+            elif solver_type == "taylor":
+                D1_0 = (1.0 / r1) * (model_s1 - model_s)
+                D1_1 = (1.0 / r2) * (model_s2 - model_s)
+                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
+                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (alpha_t * phi_2) * D1
+                    - (alpha_t * phi_3) * D2
+                )
+        else:
+            phi_11 = torch.expm1(r1 * h)
+            phi_12 = torch.expm1(r2 * h)
+            phi_1 = torch.expm1(h)
+            phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.0
+            phi_2 = phi_1 / h - 1.0
+            phi_3 = phi_2 / h - 0.5
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            if model_s1 is None:
+                x_s1 = (torch.exp(log_alpha_s1 - log_alpha_s)) * x - (sigma_s1 * phi_11) * model_s
+                model_s1 = self.model_fn(x_s1, s1)
+            x_s2 = (
+                (torch.exp(log_alpha_s2 - log_alpha_s)) * x
+                - (sigma_s2 * phi_12) * model_s
+                - r2 / r1 * (sigma_s2 * phi_22) * (model_s1 - model_s)
+            )
+            model_s2 = self.model_fn(x_s2, s2)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_s)) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (1.0 / r2) * (sigma_t * phi_2) * (model_s2 - model_s)
+                )
+            elif solver_type == "taylor":
+                D1_0 = (1.0 / r1) * (model_s1 - model_s)
+                D1_1 = (1.0 / r2) * (model_s2 - model_s)
+                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
+                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_s)) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (sigma_t * phi_2) * D1
+                    - (sigma_t * phi_3) * D2
+                )
+
+        if return_intermediate:
+            return x_t, {"model_s": model_s, "model_s1": model_s1, "model_s2": model_s2}
+        else:
+            return x_t
+
+    def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"):
+        """
+        Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}")
+        ns = self.noise_schedule
+        model_prev_1, model_prev_0 = model_prev_list[-2], model_prev_list[-1]
+        t_prev_1, t_prev_0 = t_prev_list[-2], t_prev_list[-1]
+        lambda_prev_1, lambda_prev_0, lambda_t = (
+            ns.marginal_lambda(t_prev_1),
+            ns.marginal_lambda(t_prev_0),
+            ns.marginal_lambda(t),
+        )
+        log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
+        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        h_0 = lambda_prev_0 - lambda_prev_1
+        h = lambda_t - lambda_prev_0
+        r0 = h_0 / h
+        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            if solver_type == "dpmsolver":
+                x_t = (sigma_t / sigma_prev_0) * x - (alpha_t * phi_1) * model_prev_0 - 0.5 * (alpha_t * phi_1) * D1_0
+            elif solver_type == "taylor":
+                x_t = (
+                    (sigma_t / sigma_prev_0) * x
+                    - (alpha_t * phi_1) * model_prev_0
+                    + (alpha_t * (phi_1 / h + 1.0)) * D1_0
+                )
+        else:
+            phi_1 = torch.expm1(h)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                    - (sigma_t * phi_1) * model_prev_0
+                    - 0.5 * (sigma_t * phi_1) * D1_0
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                    - (sigma_t * phi_1) * model_prev_0
+                    - (sigma_t * (phi_1 / h - 1.0)) * D1_0
+                )
+        return x_t
+
+    def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"):
+        """
+        Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        ns = self.noise_schedule
+        model_prev_2, model_prev_1, model_prev_0 = model_prev_list
+        t_prev_2, t_prev_1, t_prev_0 = t_prev_list
+        lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = (
+            ns.marginal_lambda(t_prev_2),
+            ns.marginal_lambda(t_prev_1),
+            ns.marginal_lambda(t_prev_0),
+            ns.marginal_lambda(t),
+        )
+        log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
+        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        h_1 = lambda_prev_1 - lambda_prev_2
+        h_0 = lambda_prev_0 - lambda_prev_1
+        h = lambda_t - lambda_prev_0
+        r0, r1 = h_0 / h, h_1 / h
+        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
+        D1_1 = (1.0 / r1) * (model_prev_1 - model_prev_2)
+        D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
+        D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            phi_2 = phi_1 / h + 1.0
+            phi_3 = phi_2 / h - 0.5
+            x_t = (
+                (sigma_t / sigma_prev_0) * x
+                - (alpha_t * phi_1) * model_prev_0
+                + (alpha_t * phi_2) * D1
+                - (alpha_t * phi_3) * D2
+            )
+        else:
+            phi_1 = torch.expm1(h)
+            phi_2 = phi_1 / h - 1.0
+            phi_3 = phi_2 / h - 0.5
+            x_t = (
+                (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                - (sigma_t * phi_1) * model_prev_0
+                - (sigma_t * phi_2) * D1
+                - (sigma_t * phi_3) * D2
+            )
+        return x_t
+
+    def singlestep_dpm_solver_update(
+        self, x, s, t, order, return_intermediate=False, solver_type="dpmsolver", r1=None, r2=None
+    ):
+        """
+        Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+            r1: A `float`. The hyperparameter of the second-order or third-order solver.
+            r2: A `float`. The hyperparameter of the third-order solver.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if order == 1:
+            return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate)
+        elif order == 2:
+            return self.singlestep_dpm_solver_second_update(
+                x, s, t, return_intermediate=return_intermediate, solver_type=solver_type, r1=r1
+            )
+        elif order == 3:
+            return self.singlestep_dpm_solver_third_update(
+                x, s, t, return_intermediate=return_intermediate, solver_type=solver_type, r1=r1, r2=r2
+            )
+        else:
+            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
+
+    def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type="dpmsolver"):
+        """
+        Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if order == 1:
+            return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1])
+        elif order == 2:
+            return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
+        elif order == 3:
+            return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
+        else:
+            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
+
+    def dpm_solver_adaptive(
+        self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5, solver_type="dpmsolver"
+    ):
+        """
+        The adaptive step size solver based on singlestep DPM-Solver.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `t_T`.
+            order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            h_init: A `float`. The initial step size (for logSNR).
+            atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
+            rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
+            theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
+            t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
+                current time and `t_0` is less than `t_err`. The default setting is 1e-5.
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_0: A pytorch tensor. The approximated solution at time `t_0`.
+
+        [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
+        """
+        ns = self.noise_schedule
+        s = t_T * torch.ones((1,)).to(x)
+        lambda_s = ns.marginal_lambda(s)
+        lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
+        h = h_init * torch.ones_like(s).to(x)
+        x_prev = x
+        nfe = 0
+        if order == 2:
+            r1 = 0.5
+            lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True)
+            higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(
+                x, s, t, r1=r1, solver_type=solver_type, **kwargs
+            )
+        elif order == 3:
+            r1, r2 = 1.0 / 3.0, 2.0 / 3.0
+            lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(
+                x, s, t, r1=r1, return_intermediate=True, solver_type=solver_type
+            )
+            higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(
+                x, s, t, r1=r1, r2=r2, solver_type=solver_type, **kwargs
+            )
+        else:
+            raise ValueError(f"For adaptive step size solver, order must be 2 or 3, got {order}")
+        while torch.abs(s - t_0).mean() > t_err:
+            t = ns.inverse_lambda(lambda_s + h)
+            x_lower, lower_noise_kwargs = lower_update(x, s, t)
+            x_higher = higher_update(x, s, t, **lower_noise_kwargs)
+            delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)))
+            norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True))
+            E = norm_fn((x_higher - x_lower) / delta).max()
+            if torch.all(E <= 1.0):
+                x = x_higher
+                s = t
+                x_prev = x_lower
+                lambda_s = ns.marginal_lambda(s)
+            h = torch.min(theta * h * torch.float_power(E, -1.0 / order).float(), lambda_0 - lambda_s)
+            nfe += order
+        print("adaptive solver nfe", nfe)
+        return x
+
+    def add_noise(self, x, t, noise=None):
+        """
+        Compute the noised input xt = alpha_t * x + sigma_t * noise.
+
+        Args:
+            x: A `torch.Tensor` with shape `(batch_size, *shape)`.
+            t: A `torch.Tensor` with shape `(t_size,)`.
+        Returns:
+            xt with shape `(t_size, batch_size, *shape)`.
+        """
+        alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t)
+        if noise is None:
+            noise = torch.randn((t.shape[0], *x.shape), device=x.device)
+        x = x.reshape((-1, *x.shape))
+        xt = expand_dims(alpha_t, x.dim()) * x + expand_dims(sigma_t, x.dim()) * noise
+        if t.shape[0] == 1:
+            return xt.squeeze(0)
+        else:
+            return xt
+
+    def inverse(
+        self,
+        x,
+        steps=20,
+        t_start=None,
+        t_end=None,
+        order=2,
+        skip_type="time_uniform",
+        method="multistep",
+        lower_order_final=True,
+        denoise_to_zero=False,
+        solver_type="dpmsolver",
+        atol=0.0078,
+        rtol=0.05,
+        return_intermediate=False,
+    ):
+        """
+        Inverse the sample `x` from time `t_start` to `t_end` by DPM-Solver.
+        For discrete-time DPMs, we use `t_start=1/N`, where `N` is the total time steps during training.
+        """
+        t_0 = 1.0 / self.noise_schedule.total_N if t_start is None else t_start
+        t_T = self.noise_schedule.T if t_end is None else t_end
+        assert (
+            t_0 > 0 and t_T > 0
+        ), "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        return self.sample(
+            x,
+            steps=steps,
+            t_start=t_0,
+            t_end=t_T,
+            order=order,
+            skip_type=skip_type,
+            method=method,
+            lower_order_final=lower_order_final,
+            denoise_to_zero=denoise_to_zero,
+            solver_type=solver_type,
+            atol=atol,
+            rtol=rtol,
+            return_intermediate=return_intermediate,
+        )
+
+    def sample(
+        self,
+        x,
+        steps=20,
+        t_start=None,
+        t_end=None,
+        order=2,
+        skip_type="time_uniform",
+        method="multistep",
+        lower_order_final=True,
+        denoise_to_zero=False,
+        solver_type="dpmsolver",
+        atol=0.0078,
+        rtol=0.05,
+        return_intermediate=False,
+        flow_shift=1.0,
+    ):
+        """
+        Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
+
+        =====================================================
+
+        We support the following algorithms for both noise prediction model and data prediction model:
+            - 'singlestep':
+                Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
+                We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
+                The total number of function evaluations (NFE) == `steps`.
+                Given a fixed NFE == `steps`, the sampling procedure is:
+                    - If `order` == 1:
+                        - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
+                    - If `order` == 2:
+                        - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
+                        - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
+                        - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
+                    - If `order` == 3:
+                        - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
+                        - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
+                        - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
+                        - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
+            - 'multistep':
+                Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
+                We initialize the first `order` values by lower order multistep solvers.
+                Given a fixed NFE == `steps`, the sampling procedure is:
+                    Denote K = steps.
+                    - If `order` == 1:
+                        - We use K steps of DPM-Solver-1 (i.e. DDIM).
+                    - If `order` == 2:
+                        - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
+                    - If `order` == 3:
+                        - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
+            - 'singlestep_fixed':
+                Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
+                We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
+            - 'adaptive':
+                Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
+                We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
+                You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
+                (NFE) and the sample quality.
+                    - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
+                    - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
+
+        =====================================================
+
+        Some advices for choosing the algorithm:
+            - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
+                Use singlestep DPM-Solver or DPM-Solver++ ("DPM-Solver-fast" in the paper) with `order = 3`.
+                e.g., DPM-Solver:
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
+                            skip_type='time_uniform', method='singlestep')
+                e.g., DPM-Solver++:
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
+                            skip_type='time_uniform', method='singlestep')
+            - For **guided sampling with large guidance scale** by DPMs:
+                Use multistep DPM-Solver with `algorithm_type="dpmsolver++"` and `order = 2`.
+                e.g.
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
+                            skip_type='time_uniform', method='multistep')
+
+        We support three types of `skip_type`:
+            - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
+            - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
+            - 'time_quadratic': quadratic time for the time steps.
+
+        =====================================================
+        Args:
+            x: A pytorch tensor. The initial value at time `t_start`
+                e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
+            steps: A `int`. The total number of function evaluations (NFE).
+            t_start: A `float`. The starting time of the sampling.
+                If `T` is None, we use self.noise_schedule.T (default is 1.0).
+            t_end: A `float`. The ending time of the sampling.
+                If `t_end` is None, we use 1. / self.noise_schedule.total_N.
+                e.g. if total_N == 1000, we have `t_end` == 1e-3.
+                For discrete-time DPMs:
+                    - We recommend `t_end` == 1. / self.noise_schedule.total_N.
+                For continuous-time DPMs:
+                    - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
+            order: A `int`. The order of DPM-Solver.
+            skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
+            method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
+            denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
+                Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
+
+                This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
+                score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
+                for diffusion models sampling by diffusion SDEs for low-resolutional images
+                (such as CIFAR-10). However, we observed that such trick does not matter for
+                high-resolutional images. As it needs an additional NFE, we do not recommend
+                it for high-resolutional images.
+            lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
+                Only valid for `method=multistep` and `steps < 15`. We empirically find that
+                this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
+                (especially for steps <= 10). So we recommend to set it to be `True`.
+            solver_type: A `str`. The taylor expansion type for the solver. `dpmsolver` or `taylor`. We recommend `dpmsolver`.
+            atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
+            rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
+            return_intermediate: A `bool`. Whether to save the xt at each step.
+                When set to `True`, method returns a tuple (x0, intermediates); when set to False, method returns only x0.
+        Returns:
+            x_end: A pytorch tensor. The approximated solution at time `t_end`.
+
+        """
+        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
+        t_T = self.noise_schedule.T if t_start is None else t_start
+        assert (
+            t_0 > 0 and t_T > 0
+        ), "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        if return_intermediate:
+            assert method in [
+                "multistep",
+                "singlestep",
+                "singlestep_fixed",
+            ], "Cannot use adaptive solver when saving intermediate values"
+        if self.correcting_xt_fn is not None:
+            assert method in [
+                "multistep",
+                "singlestep",
+                "singlestep_fixed",
+            ], "Cannot use adaptive solver when correcting_xt_fn is not None"
+        device = x.device
+        intermediates = []
+        with torch.no_grad():
+            if method == "adaptive":
+                x = self.dpm_solver_adaptive(
+                    x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol, solver_type=solver_type
+                )
+            elif method == "multistep":
+                assert steps >= order
+                timesteps = self.get_time_steps(
+                    skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device, shift=flow_shift
+                )
+                assert timesteps.shape[0] - 1 == steps
+                # Init the initial values.
+                step = 0
+                t = timesteps[step]
+                t_prev_list = [t]
+                model_prev_list = [self.model_fn(x, t)]
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+                self.update_progress(step + 1, len(timesteps))
+                # Init the first `order` values by lower order multistep DPM-Solver.
+                for step in range(1, order):
+                    t = timesteps[step]
+                    x = self.multistep_dpm_solver_update(
+                        x, model_prev_list, t_prev_list, t, step, solver_type=solver_type
+                    )
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    t_prev_list.append(t)
+                    model_prev_list.append(self.model_fn(x, t))
+                    # update progress bar
+                    self.update_progress(step + 1, len(timesteps))
+                # Compute the remaining values by `order`-th order multistep DPM-Solver.
+                for step in tqdm(range(order, steps + 1), disable=os.getenv("DPM_TQDM", "False") == "True"):
+                    t = timesteps[step]
+                    # We only use lower order for steps < 10
+                    # if lower_order_final and steps < 10:
+                    if lower_order_final:  # recommended by Shuchen Xue
+                        step_order = min(order, steps + 1 - step)
+                    else:
+                        step_order = order
+                    x = self.multistep_dpm_solver_update(
+                        x, model_prev_list, t_prev_list, t, step_order, solver_type=solver_type
+                    )
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    for i in range(order - 1):
+                        t_prev_list[i] = t_prev_list[i + 1]
+                        model_prev_list[i] = model_prev_list[i + 1]
+                    t_prev_list[-1] = t
+                    # We do not need to evaluate the final model value.
+                    if step < steps:
+                        model_prev_list[-1] = self.model_fn(x, t)
+                    # update progress bar
+                    self.update_progress(step + 1, len(timesteps))
+            elif method in ["singlestep", "singlestep_fixed"]:
+                if method == "singlestep":
+                    timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(
+                        steps=steps, order=order, skip_type=skip_type, t_T=t_T, t_0=t_0, device=device
+                    )
+                elif method == "singlestep_fixed":
+                    K = steps // order
+                    orders = [
+                        order,
+                    ] * K
+                    timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device)
+                for step, order in enumerate(orders):
+                    s, t = timesteps_outer[step], timesteps_outer[step + 1]
+                    timesteps_inner = self.get_time_steps(
+                        skip_type=skip_type, t_T=s.item(), t_0=t.item(), N=order, device=device
+                    )
+                    lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
+                    h = lambda_inner[-1] - lambda_inner[0]
+                    r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
+                    r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
+                    x = self.singlestep_dpm_solver_update(x, s, t, order, solver_type=solver_type, r1=r1, r2=r2)
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    self.update_progress(step + 1, len(timesteps_outer))
+            else:
+                raise ValueError(f"Got wrong method {method}")
+            if denoise_to_zero:
+                t = torch.ones((1,)).to(device) * t_0
+                x = self.denoise_to_zero_fn(x, t)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step + 1)
+                if return_intermediate:
+                    intermediates.append(x)
+        if return_intermediate:
+            return x, intermediates
+        else:
+            return x
+
+    
+#############################################################
+# other utility functions
+#############################################################
+
+
+def interpolate_fn(x, xp, yp):
+    """
+    A piecewise linear function y = f(x), using xp and yp as keypoints.
+    We implement f(x) in a differentiable way (i.e. applicable for autograd).
+    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
+
+    Args:
+        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
+        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
+        yp: PyTorch tensor with shape [C, K].
+    Returns:
+        The function values f(x), with shape [N, C].
+    """
+    N, K = x.shape[0], xp.shape[1]
+    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
+    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
+    x_idx = torch.argmin(x_indices, dim=2)
+    cand_start_idx = x_idx - 1
+    start_idx = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(1, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1)
+    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
+    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
+    start_idx2 = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(0, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
+    start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2)
+    end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2)
+    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
+    return cand
+
+
+def expand_dims(v, dims):
+    """
+    Expand the tensor `v` to the dim `dims`.
+
+    Args:
+        `v`: a PyTorch tensor with shape [N].
+        `dim`: a `int`.
+    Returns:
+        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
+    """
+    return v[(...,) + (None,) * (dims - 1)]

omnigen2/transport/integrators.py

@@ -0,0 +1,122 @@
+import torch as th
+from torchdiffeq import odeint
+from .utils import time_shift, get_lin_function
+
+class sde:
+    """SDE solver class"""
+
+    def __init__(
+        self,
+        drift,
+        diffusion,
+        *,
+        t0,
+        t1,
+        num_steps,
+        sampler_type,
+    ):
+        assert t0 < t1, "SDE sampler has to be in forward time"
+
+        self.num_timesteps = num_steps
+        self.t = th.linspace(t0, t1, num_steps)
+        self.dt = self.t[1] - self.t[0]
+        self.drift = drift
+        self.diffusion = diffusion
+        self.sampler_type = sampler_type
+
+    def __Euler_Maruyama_step(self, x, mean_x, t, model, **model_kwargs):
+        w_cur = th.randn(x.size()).to(x)
+        t = th.ones(x.size(0)).to(x) * t
+        dw = w_cur * th.sqrt(self.dt)
+        drift = self.drift(x, t, model, **model_kwargs)
+        diffusion = self.diffusion(x, t)
+        mean_x = x + drift * self.dt
+        x = mean_x + th.sqrt(2 * diffusion) * dw
+        return x, mean_x
+
+    def __Heun_step(self, x, _, t, model, **model_kwargs):
+        w_cur = th.randn(x.size()).to(x)
+        dw = w_cur * th.sqrt(self.dt)
+        t_cur = th.ones(x.size(0)).to(x) * t
+        diffusion = self.diffusion(x, t_cur)
+        xhat = x + th.sqrt(2 * diffusion) * dw
+        K1 = self.drift(xhat, t_cur, model, **model_kwargs)
+        xp = xhat + self.dt * K1
+        K2 = self.drift(xp, t_cur + self.dt, model, **model_kwargs)
+        return (
+            xhat + 0.5 * self.dt * (K1 + K2),
+            xhat,
+        )  # at last time point we do not perform the heun step
+
+    def __forward_fn(self):
+        """TODO: generalize here by adding all private functions ending with steps to it"""
+        sampler_dict = {
+            "Euler": self.__Euler_Maruyama_step,
+            "Heun": self.__Heun_step,
+        }
+
+        try:
+            sampler = sampler_dict[self.sampler_type]
+        except:
+            raise NotImplementedError("Smapler type not implemented.")
+
+        return sampler
+
+    def sample(self, init, model, **model_kwargs):
+        """forward loop of sde"""
+        x = init
+        mean_x = init
+        samples = []
+        sampler = self.__forward_fn()
+        for ti in self.t[:-1]:
+            with th.no_grad():
+                x, mean_x = sampler(x, mean_x, ti, model, **model_kwargs)
+                samples.append(x)
+
+        return samples
+
+
+class ode:
+    """ODE solver class"""
+
+    def __init__(
+        self,
+        drift,
+        *,
+        t0,
+        t1,
+        sampler_type,
+        num_steps,
+        atol,
+        rtol,
+        do_shift=False,
+        time_shifting_factor=None,
+    ):
+        assert t0 < t1, "ODE sampler has to be in forward time"
+
+        self.drift = drift
+        self.do_shift = do_shift
+        self.t = th.linspace(t0, t1, num_steps)
+        if time_shifting_factor:
+            self.t = self.t / (self.t + time_shifting_factor - time_shifting_factor * self.t)
+        self.atol = atol
+        self.rtol = rtol
+        self.sampler_type = sampler_type
+
+    def sample(self, x, model, **model_kwargs):
+        x = x.float()
+        device = x[0].device if isinstance(x, tuple) else x.device
+
+        def _fn(t, x):
+            t = th.ones(x[0].size(0)).to(device) * t if isinstance(x, tuple) else th.ones(x.size(0)).to(device) * t
+            model_output = self.drift(x, t, model, **model_kwargs).float()
+            return model_output
+        
+        t = self.t.to(device)
+        if self.do_shift:
+            mu = get_lin_function(y1=0.5, y2=1.15)(x.shape[1])
+            t = time_shift(mu, 1.0, t)
+        atol = [self.atol] * len(x) if isinstance(x, tuple) else [self.atol]
+        rtol = [self.rtol] * len(x) if isinstance(x, tuple) else [self.rtol]
+        samples = odeint(_fn, x, t, method=self.sampler_type, atol=atol, rtol=rtol)
+        return samples

omnigen2/transport/path.py

@@ -0,0 +1,201 @@
+import numpy as np
+import torch as th
+
+
+def expand_t_like_x(t, x):
+    """Function to reshape time t to broadcastable dimension of x
+    Args:
+      t: [batch_dim,], time vector
+      x: [batch_dim,...], data point
+    """
+    dims = [1] * len(x[0].size())
+    t = t.view(t.size(0), *dims)
+    return t
+
+
+#################### Coupling Plans ####################
+
+
+class ICPlan:
+    """Linear Coupling Plan"""
+
+    def __init__(self, sigma=0.0):
+        self.sigma = sigma
+
+    def compute_alpha_t(self, t):
+        """Compute the data coefficient along the path"""
+        return t, 1
+
+    def compute_sigma_t(self, t):
+        """Compute the noise coefficient along the path"""
+        return 1 - t, -1
+
+    def compute_d_alpha_alpha_ratio_t(self, t):
+        """Compute the ratio between d_alpha and alpha"""
+        return 1 / t
+
+    def compute_drift(self, x, t):
+        """We always output sde according to score parametrization;"""
+        t = expand_t_like_x(t, x)
+        alpha_ratio = self.compute_d_alpha_alpha_ratio_t(t)
+        sigma_t, d_sigma_t = self.compute_sigma_t(t)
+        drift = alpha_ratio * x
+        diffusion = alpha_ratio * (sigma_t**2) - sigma_t * d_sigma_t
+
+        return -drift, diffusion
+
+    def compute_diffusion(self, x, t, form="constant", norm=1.0):
+        """Compute the diffusion term of the SDE
+        Args:
+          x: [batch_dim, ...], data point
+          t: [batch_dim,], time vector
+          form: str, form of the diffusion term
+          norm: float, norm of the diffusion term
+        """
+        t = expand_t_like_x(t, x)
+        choices = {
+            "constant": norm,
+            "SBDM": norm * self.compute_drift(x, t)[1],
+            "sigma": norm * self.compute_sigma_t(t)[0],
+            "linear": norm * (1 - t),
+            "decreasing": 0.25 * (norm * th.cos(np.pi * t) + 1) ** 2,
+            "inccreasing-decreasing": norm * th.sin(np.pi * t) ** 2,
+        }
+
+        try:
+            diffusion = choices[form]
+        except KeyError:
+            raise NotImplementedError(f"Diffusion form {form} not implemented")
+
+        return diffusion
+
+    def get_score_from_velocity(self, velocity, x, t):
+        """Wrapper function: transfrom velocity prediction model to score
+        Args:
+            velocity: [batch_dim, ...] shaped tensor; velocity model output
+            x: [batch_dim, ...] shaped tensor; x_t data point
+            t: [batch_dim,] time tensor
+        """
+        t = expand_t_like_x(t, x)
+        alpha_t, d_alpha_t = self.compute_alpha_t(t)
+        sigma_t, d_sigma_t = self.compute_sigma_t(t)
+        mean = x
+        reverse_alpha_ratio = alpha_t / d_alpha_t
+        var = sigma_t**2 - reverse_alpha_ratio * d_sigma_t * sigma_t
+        score = (reverse_alpha_ratio * velocity - mean) / var
+        return score
+
+    def get_noise_from_velocity(self, velocity, x, t):
+        """Wrapper function: transfrom velocity prediction model to denoiser
+        Args:
+            velocity: [batch_dim, ...] shaped tensor; velocity model output
+            x: [batch_dim, ...] shaped tensor; x_t data point
+            t: [batch_dim,] time tensor
+        """
+        t = expand_t_like_x(t, x)
+        alpha_t, d_alpha_t = self.compute_alpha_t(t)
+        sigma_t, d_sigma_t = self.compute_sigma_t(t)
+        mean = x
+        reverse_alpha_ratio = alpha_t / d_alpha_t
+        var = reverse_alpha_ratio * d_sigma_t - sigma_t
+        noise = (reverse_alpha_ratio * velocity - mean) / var
+        return noise
+
+    def get_velocity_from_score(self, score, x, t):
+        """Wrapper function: transfrom score prediction model to velocity
+        Args:
+            score: [batch_dim, ...] shaped tensor; score model output
+            x: [batch_dim, ...] shaped tensor; x_t data point
+            t: [batch_dim,] time tensor
+        """
+        t = expand_t_like_x(t, x)
+        drift, var = self.compute_drift(x, t)
+        velocity = var * score - drift
+        return velocity
+
+    def compute_mu_t(self, t, x0, x1):
+        """Compute the mean of time-dependent density p_t"""
+        t = expand_t_like_x(t, x1)
+        alpha_t, _ = self.compute_alpha_t(t)
+        sigma_t, _ = self.compute_sigma_t(t)
+        if isinstance(x1, (list, tuple)):
+            return [alpha_t[i] * x1[i] + sigma_t[i] * x0[i] for i in range(len(x1))]
+        else:
+            return alpha_t * x1 + sigma_t * x0
+
+    def compute_xt(self, t, x0, x1):
+        """Sample xt from time-dependent density p_t; rng is required"""
+        xt = self.compute_mu_t(t, x0, x1)
+        return xt
+
+    def compute_ut(self, t, x0, x1, xt):
+        """Compute the vector field corresponding to p_t"""
+        t = expand_t_like_x(t, x1)
+        _, d_alpha_t = self.compute_alpha_t(t)
+        _, d_sigma_t = self.compute_sigma_t(t)
+        if isinstance(x1, (list, tuple)):
+            return [d_alpha_t * x1[i] + d_sigma_t * x0[i] for i in range(len(x1))]
+        else:
+            return d_alpha_t * x1 + d_sigma_t * x0
+
+    def plan(self, t, x0, x1):
+        xt = self.compute_xt(t, x0, x1)
+        ut = self.compute_ut(t, x0, x1, xt)
+        return t, xt, ut
+
+
+class VPCPlan(ICPlan):
+    """class for VP path flow matching"""
+
+    def __init__(self, sigma_min=0.1, sigma_max=20.0):
+        self.sigma_min = sigma_min
+        self.sigma_max = sigma_max
+        self.log_mean_coeff = (
+            lambda t: -0.25 * ((1 - t) ** 2) * (self.sigma_max - self.sigma_min) - 0.5 * (1 - t) * self.sigma_min
+        )
+        self.d_log_mean_coeff = lambda t: 0.5 * (1 - t) * (self.sigma_max - self.sigma_min) + 0.5 * self.sigma_min
+
+    def compute_alpha_t(self, t):
+        """Compute coefficient of x1"""
+        alpha_t = self.log_mean_coeff(t)
+        alpha_t = th.exp(alpha_t)
+        d_alpha_t = alpha_t * self.d_log_mean_coeff(t)
+        return alpha_t, d_alpha_t
+
+    def compute_sigma_t(self, t):
+        """Compute coefficient of x0"""
+        p_sigma_t = 2 * self.log_mean_coeff(t)
+        sigma_t = th.sqrt(1 - th.exp(p_sigma_t))
+        d_sigma_t = th.exp(p_sigma_t) * (2 * self.d_log_mean_coeff(t)) / (-2 * sigma_t)
+        return sigma_t, d_sigma_t
+
+    def compute_d_alpha_alpha_ratio_t(self, t):
+        """Special purposed function for computing numerical stabled d_alpha_t / alpha_t"""
+        return self.d_log_mean_coeff(t)
+
+    def compute_drift(self, x, t):
+        """Compute the drift term of the SDE"""
+        t = expand_t_like_x(t, x)
+        beta_t = self.sigma_min + (1 - t) * (self.sigma_max - self.sigma_min)
+        return -0.5 * beta_t * x, beta_t / 2
+
+
+class GVPCPlan(ICPlan):
+    def __init__(self, sigma=0.0):
+        super().__init__(sigma)
+
+    def compute_alpha_t(self, t):
+        """Compute coefficient of x1"""
+        alpha_t = th.sin(t * np.pi / 2)
+        d_alpha_t = np.pi / 2 * th.cos(t * np.pi / 2)
+        return alpha_t, d_alpha_t
+
+    def compute_sigma_t(self, t):
+        """Compute coefficient of x0"""
+        sigma_t = th.cos(t * np.pi / 2)
+        d_sigma_t = -np.pi / 2 * th.sin(t * np.pi / 2)
+        return sigma_t, d_sigma_t
+
+    def compute_d_alpha_alpha_ratio_t(self, t):
+        """Special purposed function for computing numerical stabled d_alpha_t / alpha_t"""
+        return np.pi / (2 * th.tan(t * np.pi / 2))

omnigen2/transport/transport.py

@@ -0,0 +1,545 @@
+import enum
+import math
+from typing import Callable, Optional
+
+import numpy as np
+import torch as th
+import random
+
+from . import path
+from .integrators import ode, sde
+from .utils import mean_flat, expand_dims
+from .dpm_solver import NoiseScheduleFlow, model_wrapper, DPM_Solver
+
+
+class ModelType(enum.Enum):
+    """
+    Which type of output the model predicts.
+    """
+
+    NOISE = enum.auto()  # the model predicts epsilon
+    SCORE = enum.auto()  # the model predicts \nabla \log p(x)
+    VELOCITY = enum.auto()  # the model predicts v(x)
+
+
+class PathType(enum.Enum):
+    """
+    Which type of path to use.
+    """
+
+    LINEAR = enum.auto()
+    GVP = enum.auto()
+    VP = enum.auto()
+
+
+class WeightType(enum.Enum):
+    """
+    Which type of weighting to use.
+    """
+
+    NONE = enum.auto()
+    VELOCITY = enum.auto()
+    LIKELIHOOD = enum.auto()
+
+
+class Transport:
+    def __init__(self, *, model_type, path_type, loss_type, train_eps, sample_eps, snr_type, do_shift, seq_len,
+                 dynamic_time_shift: bool = False,
+                 time_shift_version: str = "v1"):
+        path_options = {
+            PathType.LINEAR: path.ICPlan,
+            PathType.GVP: path.GVPCPlan,
+            PathType.VP: path.VPCPlan,
+        }
+
+        self.loss_type = loss_type
+        self.model_type = model_type
+        self.path_sampler = path_options[path_type]()
+        self.train_eps = train_eps
+        self.sample_eps = sample_eps
+
+        self.snr_type = snr_type
+        self.do_shift = do_shift
+        self.seq_len = seq_len
+        self.dynamic_time_shift = dynamic_time_shift
+        self.time_shift_version = time_shift_version
+    def prior_logp(self, z):
+        """
+        Standard multivariate normal prior
+        Assume z is batched
+        """
+        shape = th.tensor(z.size())
+        N = th.prod(shape[1:])
+        _fn = lambda x: -N / 2.0 * np.log(2 * np.pi) - th.sum(x**2) / 2.0
+        return th.vmap(_fn)(z)
+
+    def check_interval(
+        self,
+        train_eps,
+        sample_eps,
+        *,
+        diffusion_form="SBDM",
+        sde=False,
+        reverse=False,
+        eval=False,
+        last_step_size=0.0,
+    ):
+        t0 = 0
+        t1 = 1
+        eps = train_eps if not eval else sample_eps
+        if type(self.path_sampler) in [path.VPCPlan]:
+            t1 = 1 - eps if (not sde or last_step_size == 0) else 1 - last_step_size
+
+        elif (type(self.path_sampler) in [path.ICPlan, path.GVPCPlan]) and (
+            self.model_type != ModelType.VELOCITY or sde
+        ):  # avoid numerical issue by taking a first semi-implicit step
+            t0 = eps if (diffusion_form == "SBDM" and sde) or self.model_type != ModelType.VELOCITY else 0
+            t1 = 1 - eps if (not sde or last_step_size == 0) else 1 - last_step_size
+
+        if reverse:
+            t0, t1 = 1 - t0, 1 - t1
+
+        return t0, t1
+
+    def sample(self, x1, process_index, num_processes):
+        """Sampling x0 & t based on shape of x1 (if needed)
+        Args:
+          x1 - data point; [batch, *dim]
+        """
+        if isinstance(x1, (list, tuple)):
+            x0 = [th.randn_like(img_start) for img_start in x1]
+        else:
+            x0 = th.randn_like(x1)
+        t0, t1 = self.check_interval(self.train_eps, self.sample_eps)
+
+        if self.snr_type.startswith("uniform"):
+            assert t0 == 0.0 and t1 == 1.0, "not implemented."
+            if "_" in self.snr_type:
+                _, t0, t1 = self.snr_type.split("_")
+                t0, t1 = float(t0), float(t1)
+            t = th.rand((len(x1),)) * (t1 - t0) + t0
+        if self.snr_type == "stratified_uniform":
+            batch_size = len(x1)
+            n = batch_size * num_processes
+            offsets = th.arange(process_index, n, num_processes)
+            u = th.rand(size=(batch_size,))
+            t = ((offsets + u) / n)
+        elif self.snr_type == "lognorm":
+            u = th.normal(mean=0.0, std=1.0, size=(len(x1),))
+            t = 1 / (1 + th.exp(-u)) * (t1 - t0) + t0
+        elif self.snr_type == "zero":
+            t = th.rand((len(x1),)) 
+            for _ in range(len(x1)):
+                if random.random() < 1.0:
+                    t[_] = 0.0
+            # print(t)
+        else:
+            raise NotImplementedError("Not implemented snr_type %s" % self.snr_type)
+
+        if self.do_shift:
+            if self.dynamic_time_shift:
+                if self.time_shift_version == "v1":
+                    base_shift: float = 0.5
+                    max_shift: float = 1.15
+                    lin_func = self.get_lin_function(y1=base_shift, y2=max_shift)
+                        
+                    mu = th.tensor([lin_func((_x1.shape[-2] // 2) * (_x1.shape[-1] // 2)) for _x1 in x1], dtype=t.dtype, device=t.device).view_as(t)
+                    t = self.time_shift(mu, 1.0, t)
+                elif self.time_shift_version == "v2":
+                    tokens = th.tensor([(_x1.shape[-2] // 2) * (_x1.shape[-1] // 2) for _x1 in x1], dtype=t.dtype, device=t.device).view_as(t)
+                    t = self.time_shift_v2(tokens, t)
+            else:
+                if self.time_shift_version == "v1":
+                    base_shift: float = 0.5
+                    max_shift: float = 1.15
+                    mu = self.get_lin_function(y1=base_shift, y2=max_shift)(self.seq_len)
+                    t = self.time_shift(mu, 1.0, t)
+                elif self.time_shift_version == "v2":
+                    tokens = th.tensor([self.seq_len] * len(x1), dtype=t.dtype, device=t.device).view_as(t)
+                    t = self.time_shift_v2(tokens, t)
+        t = t.to(x1[0])
+        return t, x0, x1
+
+    def time_shift(self, mu: float, sigma: float, t: th.Tensor):
+        # the following implementation was original for t=0: clean / t=1: noise
+        # Since we adopt the reverse, the 1-t operations are needed
+        t = 1 - t
+        if isinstance(mu, th.Tensor):
+            t = th.exp(mu) / (th.exp(mu) + (1 / t - 1) ** sigma)
+        else:
+            t = math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+        t = 1 - t
+        return t
+    
+    def time_shift_v2(self, tokens: th.Tensor, t: th.Tensor):
+        # t = th.exp(mu) / (th.exp(mu) + (1 / t - 1) ** sigma)
+        m = th.sqrt(tokens) / 20
+        t = t / (m - m * t + t)
+        return t
+
+    def get_lin_function(
+        self, x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15
+    ) -> Callable[[float], float]:
+        m = (y2 - y1) / (x2 - x1)
+        b = y1 - m * x1
+        return lambda x: m * x + b
+
+    def training_losses(
+        self,
+        model,
+        x1,
+        model_kwargs=None,
+        process_index: Optional[int] = None,
+        num_processes: Optional[int] = None,
+        reduction: str = 'mean',
+    ):
+        """Loss for training the score model
+        Args:
+        - model: backbone model; could be score, noise, or velocity
+        - x1: datapoint
+        - model_kwargs: additional arguments for the model
+        """
+
+        terms = {}
+
+        if model_kwargs is None:
+            model_kwargs = {}
+        t, x0, x1 = self.sample(x1, process_index, num_processes)
+        t, xt, ut = self.path_sampler.plan(t, x0, x1)
+
+        terms = {}
+        terms['t'] = t
+        terms['xt'] = xt
+        
+        if "cond" in model_kwargs:
+            conds = model_kwargs.pop("cond")
+            xt = [th.cat([x, cond], dim=0) if cond is not None else x for x, cond in zip(xt, conds)]
+        model_output = model(xt, t, **model_kwargs)
+        B = len(x0)
+
+        terms['pred'] = model_output
+        if self.model_type == ModelType.VELOCITY:
+            if isinstance(x1, (list, tuple)):
+                assert len(model_output) == len(ut) == len(x1)
+                for i in range(B):
+                    assert (
+                        model_output[i].shape == ut[i].shape == x1[i].shape
+                    ), f"{model_output[i].shape} {ut[i].shape} {x1[i].shape}"
+                terms["task_loss"] = th.stack(
+                    [th.nn.functional.mse_loss(ut[i].float(), model_output[i].float(), reduction=reduction) for i in range(B)],
+                    dim=0,
+                )
+            else:
+                terms["task_loss"] = mean_flat(((model_output - ut) ** 2))
+        else:
+            raise NotImplementedError
+
+        terms["loss"] = terms["task_loss"]
+        terms["t"] = t
+        return terms
+
+    def get_drift(self):
+        """member function for obtaining the drift of the probability flow ODE"""
+
+        def score_ode(x, t, model, **model_kwargs):
+            drift_mean, drift_var = self.path_sampler.compute_drift(x, t)
+            model_output = model(x, t, **model_kwargs)
+            return -drift_mean + drift_var * model_output  # by change of variable
+
+        def noise_ode(x, t, model, **model_kwargs):
+            drift_mean, drift_var = self.path_sampler.compute_drift(x, t)
+            sigma_t, _ = self.path_sampler.compute_sigma_t(path.expand_t_like_x(t, x))
+            model_output = model(x, t, **model_kwargs)
+            score = model_output / -sigma_t
+            return -drift_mean + drift_var * score
+
+        def velocity_ode(x, t, model, **model_kwargs):
+            model_output = model(x, t, **model_kwargs)
+            return model_output
+
+        if self.model_type == ModelType.NOISE:
+            drift_fn = noise_ode
+        elif self.model_type == ModelType.SCORE:
+            drift_fn = score_ode
+        else:
+            drift_fn = velocity_ode
+
+        def body_fn(x, t, model, **model_kwargs):
+            model_output = drift_fn(x, t, model, **model_kwargs)
+            assert model_output.shape == x.shape, "Output shape from ODE solver must match input shape"
+            return model_output
+
+        return body_fn
+
+    def get_score(
+        self,
+    ):
+        """member function for obtaining score of
+        x_t = alpha_t * x + sigma_t * eps"""
+        if self.model_type == ModelType.NOISE:
+            score_fn = (
+                lambda x, t, model, **kwargs: model(x, t, **kwargs)
+                / -self.path_sampler.compute_sigma_t(path.expand_t_like_x(t, x))[0]
+            )
+        elif self.model_type == ModelType.SCORE:
+            score_fn = lambda x, t, model, **kwagrs: model(x, t, **kwagrs)
+        elif self.model_type == ModelType.VELOCITY:
+            score_fn = lambda x, t, model, **kwargs: self.path_sampler.get_score_from_velocity(
+                model(x, t, **kwargs), x, t
+            )
+        else:
+            raise NotImplementedError()
+
+        return score_fn
+
+
+class Sampler:
+    """Sampler class for the transport model"""
+
+    def __init__(
+        self,
+        transport,
+    ):
+        """Constructor for a general sampler; supporting different sampling methods
+        Args:
+        - transport: an tranport object specify model prediction & interpolant type
+        """
+
+        self.transport = transport
+        self.drift = self.transport.get_drift()
+        self.score = self.transport.get_score()
+
+    def __get_sde_diffusion_and_drift(
+        self,
+        *,
+        diffusion_form="SBDM",
+        diffusion_norm=1.0,
+    ):
+        def diffusion_fn(x, t):
+            diffusion = self.transport.path_sampler.compute_diffusion(x, t, form=diffusion_form, norm=diffusion_norm)
+            return diffusion
+
+        sde_drift = lambda x, t, model, **kwargs: self.drift(x, t, model, **kwargs) + diffusion_fn(x, t) * self.score(
+            x, t, model, **kwargs
+        )
+
+        sde_diffusion = diffusion_fn
+
+        return sde_drift, sde_diffusion
+
+    def __get_last_step(
+        self,
+        sde_drift,
+        *,
+        last_step,
+        last_step_size,
+    ):
+        """Get the last step function of the SDE solver"""
+
+        if last_step is None:
+            last_step_fn = lambda x, t, model, **model_kwargs: x
+        elif last_step == "Mean":
+            last_step_fn = (
+                lambda x, t, model, **model_kwargs: x + sde_drift(x, t, model, **model_kwargs) * last_step_size
+            )
+        elif last_step == "Tweedie":
+            alpha = self.transport.path_sampler.compute_alpha_t  # simple aliasing; the original name was too long
+            sigma = self.transport.path_sampler.compute_sigma_t
+            last_step_fn = lambda x, t, model, **model_kwargs: x / alpha(t)[0][0] + (sigma(t)[0][0] ** 2) / alpha(t)[0][
+                0
+            ] * self.score(x, t, model, **model_kwargs)
+        elif last_step == "Euler":
+            last_step_fn = (
+                lambda x, t, model, **model_kwargs: x + self.drift(x, t, model, **model_kwargs) * last_step_size
+            )
+        else:
+            raise NotImplementedError()
+
+        return last_step_fn
+
+    def sample_sde(
+        self,
+        *,
+        sampling_method="Euler",
+        diffusion_form="SBDM",
+        diffusion_norm=1.0,
+        last_step="Mean",
+        last_step_size=0.04,
+        num_steps=250,
+    ):
+        """returns a sampling function with given SDE settings
+        Args:
+        - sampling_method: type of sampler used in solving the SDE; default to be Euler-Maruyama
+        - diffusion_form: function form of diffusion coefficient; default to be matching SBDM
+        - diffusion_norm: function magnitude of diffusion coefficient; default to 1
+        - last_step: type of the last step; default to identity
+        - last_step_size: size of the last step; default to match the stride of 250 steps over [0,1]
+        - num_steps: total integration step of SDE
+        """
+
+        if last_step is None:
+            last_step_size = 0.0
+
+        sde_drift, sde_diffusion = self.__get_sde_diffusion_and_drift(
+            diffusion_form=diffusion_form,
+            diffusion_norm=diffusion_norm,
+        )
+
+        t0, t1 = self.transport.check_interval(
+            self.transport.train_eps,
+            self.transport.sample_eps,
+            diffusion_form=diffusion_form,
+            sde=True,
+            eval=True,
+            reverse=False,
+            last_step_size=last_step_size,
+        )
+
+        _sde = sde(
+            sde_drift,
+            sde_diffusion,
+            t0=t0,
+            t1=t1,
+            num_steps=num_steps,
+            sampler_type=sampling_method,
+        )
+
+        last_step_fn = self.__get_last_step(sde_drift, last_step=last_step, last_step_size=last_step_size)
+
+        def _sample(init, model, **model_kwargs):
+            xs = _sde.sample(init, model, **model_kwargs)
+            ts = th.ones(init.size(0), device=init.device) * t1
+            x = last_step_fn(xs[-1], ts, model, **model_kwargs)
+            xs.append(x)
+
+            assert len(xs) == num_steps, "Samples does not match the number of steps"
+
+            return xs
+
+        return _sample
+    
+    def sample_dpm(
+        self,
+        model,
+        model_kwargs=None,
+    ):
+
+        noise_schedule = NoiseScheduleFlow(schedule="discrete_flow")
+
+        def noise_pred_fn(x, t_continuous):
+            output = model(x, 1 - t_continuous, **model_kwargs)
+            _, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
+            try:
+                noise = x - (1 - expand_dims(sigma_t, x.dim()).to(x)) * output
+            except:
+                noise = x - (1 - expand_dims(sigma_t, x.dim()).to(x)) * output[0]
+            return noise
+
+        return DPM_Solver(noise_pred_fn, noise_schedule, algorithm_type="dpmsolver++").sample
+
+
+    def sample_ode(
+        self,
+        *,
+        sampling_method="dopri5",
+        num_steps=50,
+        atol=1e-6,
+        rtol=1e-3,
+        reverse=False,
+        do_shift=False,
+        time_shifting_factor=None, 
+    ):
+        """returns a sampling function with given ODE settings
+        Args:
+        - sampling_method: type of sampler used in solving the ODE; default to be Dopri5
+        - num_steps:
+            - fixed solver (Euler, Heun): the actual number of integration steps performed
+            - adaptive solver (Dopri5): the number of datapoints saved during integration; produced by interpolation
+        - atol: absolute error tolerance for the solver
+        - rtol: relative error tolerance for the solver
+        """
+
+        # for flux
+        drift = lambda x, t, model, **kwargs: self.drift(x, t, model, **kwargs)
+
+        t0, t1 = self.transport.check_interval(
+            self.transport.train_eps,
+            self.transport.sample_eps,
+            sde=False,
+            eval=True,
+            reverse=reverse,
+            last_step_size=0.0,
+        )
+
+        _ode = ode(
+            drift=drift,
+            t0=t0,
+            t1=t1,
+            sampler_type=sampling_method,
+            num_steps=num_steps,
+            atol=atol,
+            rtol=rtol,
+            do_shift=do_shift,
+            time_shifting_factor=time_shifting_factor,
+        )
+
+        return _ode.sample
+
+    def sample_ode_likelihood(
+        self,
+        *,
+        sampling_method="dopri5",
+        num_steps=50,
+        atol=1e-6,
+        rtol=1e-3,
+    ):
+        """returns a sampling function for calculating likelihood with given ODE settings
+        Args:
+        - sampling_method: type of sampler used in solving the ODE; default to be Dopri5
+        - num_steps:
+            - fixed solver (Euler, Heun): the actual number of integration steps performed
+            - adaptive solver (Dopri5): the number of datapoints saved during integration; produced by interpolation
+        - atol: absolute error tolerance for the solver
+        - rtol: relative error tolerance for the solver
+        """
+
+        def _likelihood_drift(x, t, model, **model_kwargs):
+            x, _ = x
+            eps = th.randint(2, x.size(), dtype=th.float, device=x.device) * 2 - 1
+            t = th.ones_like(t) * (1 - t)
+            with th.enable_grad():
+                x.requires_grad = True
+                grad = th.autograd.grad(th.sum(self.drift(x, t, model, **model_kwargs) * eps), x)[0]
+                logp_grad = th.sum(grad * eps, dim=tuple(range(1, len(x.size()))))
+                drift = self.drift(x, t, model, **model_kwargs)
+            return (-drift, logp_grad)
+
+        t0, t1 = self.transport.check_interval(
+            self.transport.train_eps,
+            self.transport.sample_eps,
+            sde=False,
+            eval=True,
+            reverse=False,
+            last_step_size=0.0,
+        )
+
+        _ode = ode(
+            drift=_likelihood_drift,
+            t0=t0,
+            t1=t1,
+            sampler_type=sampling_method,
+            num_steps=num_steps,
+            atol=atol,
+            rtol=rtol,
+        )
+
+        def _sample_fn(x, model, **model_kwargs):
+            init_logp = th.zeros(x.size(0)).to(x)
+            input = (x, init_logp)
+            drift, delta_logp = _ode.sample(input, model, **model_kwargs)
+            drift, delta_logp = drift[-1], delta_logp[-1]
+            prior_logp = self.transport.prior_logp(drift)
+            logp = prior_logp - delta_logp
+            return logp, drift
+
+        return _sample_fn

omnigen2/transport/utils.py

@@ -0,0 +1,56 @@
+import torch as th
+import math
+
+class EasyDict:
+    def __init__(self, sub_dict):
+        for k, v in sub_dict.items():
+            setattr(self, k, v)
+
+    def __getitem__(self, key):
+        return getattr(self, key)
+
+
+def mean_flat(x):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return th.mean(x, dim=list(range(1, len(x.size()))))
+
+
+def log_state(state):
+    result = []
+
+    sorted_state = dict(sorted(state.items()))
+    for key, value in sorted_state.items():
+        # Check if the value is an instance of a class
+        if "<object" in str(value) or "object at" in str(value):
+            result.append(f"{key}: [{value.__class__.__name__}]")
+        else:
+            result.append(f"{key}: {value}")
+
+    return "\n".join(result)
+
+def time_shift(mu: float, sigma: float, t: th.Tensor):
+    # the following implementation was original for t=0: clean / t=1: noise
+    # Since we adopt the reverse, the 1-t operations are needed
+    t = 1 - t
+    t = math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+    t = 1 - t
+    return t
+
+def get_lin_function(x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15):
+    m = (y2 - y1) / (x2 - x1)
+    b = y1 - m * x1
+    return lambda x: m * x + b
+
+def expand_dims(v, dims):
+    """
+    Expand the tensor `v` to the dim `dims`.
+
+    Args:
+        `v`: a PyTorch tensor with shape [N].
+        `dim`: a `int`.
+    Returns:
+        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
+    """
+    return v[(...,) + (None,) * (dims - 1)]

omnigen2/utils/img_util.py

@@ -0,0 +1,31 @@
+from typing import List
+
+from PIL import Image
+
+import torch
+from torchvision.transforms.functional import to_pil_image
+
+def resize_image(image, max_pixels, img_scale_num):
+    width, height = image.size
+    cur_pixels = height * width
+    ratio = (max_pixels / cur_pixels) ** 0.5
+    ratio = min(ratio, 1.0) # do not upscale input image
+
+    new_height, new_width = int(height * ratio) // img_scale_num * img_scale_num, int(width * ratio) // img_scale_num * img_scale_num
+
+    image = image.resize((new_width, new_height), resample=Image.BICUBIC)
+    return image
+
+def create_collage(images: List[torch.Tensor]) -> Image.Image:
+    """Create a horizontal collage from a list of images."""
+    max_height = max(img.shape[-2] for img in images)
+    total_width = sum(img.shape[-1] for img in images)
+    canvas = torch.zeros((3, max_height, total_width), device=images[0].device)
+    
+    current_x = 0
+    for img in images:
+        h, w = img.shape[-2:]
+        canvas[:, :h, current_x:current_x+w] = img * 0.5 + 0.5
+        current_x += w
+    
+    return to_pil_image(canvas)

omnigen2/utils/import_utils.py

@@ -0,0 +1,46 @@
+# Copyright 2024 The HuggingFace 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.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Import utilities: Utilities related to imports and our lazy inits.
+"""
+
+import importlib.util
+import sys
+
+# The package importlib_metadata is in a different place, depending on the python version.
+if sys.version_info < (3, 8):
+    import importlib_metadata
+else:
+    import importlib.metadata as importlib_metadata
+
+def _is_package_available(pkg_name: str):
+    pkg_exists = importlib.util.find_spec(pkg_name) is not None
+    pkg_version = "N/A"
+
+    if pkg_exists:
+        try:
+            pkg_version = importlib_metadata.version(pkg_name)
+        except (ImportError, importlib_metadata.PackageNotFoundError):
+            pkg_exists = False
+
+    return pkg_exists, pkg_version
+
+_triton_available, _triton_version = _is_package_available("triton")
+_flash_attn_available, _flash_attn_version = _is_package_available("flash_attn")
+
+def is_triton_available():
+    return _triton_available
+
+def is_flash_attn_available():
+    return _flash_attn_available

omnigen2/utils/logging_utils.py

@@ -0,0 +1,15 @@
+import logging
+
+class TqdmToLogger(object):
+    """File-like object to redirect tqdm output to a logger."""
+    def __init__(self, logger, level=logging.INFO):
+        self.logger = logger
+        self.level = level
+
+    def write(self, buf):
+        for line in buf.rstrip().splitlines():
+            self.logger.log(self.level, line)
+
+    def flush(self):
+        for handler in self.logger.logger.handlers:
+            handler.flush()

omnigen2/utils/reproducibility.py

@@ -0,0 +1,22 @@
+import random
+import numpy as np
+
+import torch
+
+from diffusers.utils import is_torch_npu_available
+
+
+def worker_init_fn(worker_id, num_processes, num_workers, process_index, seed, same_seed_per_epoch=False):
+    if same_seed_per_epoch:
+        worker_seed = seed + num_processes + num_workers * process_index + worker_id
+    else:
+        worker_seed = torch.initial_seed()
+
+    random.seed(worker_seed)
+    np.random.seed(worker_seed % 2**32)
+    torch.manual_seed(worker_seed)
+
+    if is_torch_npu_available():
+        torch.npu.manual_seed_all(seed)
+    else:
+        torch.cuda.manual_seed_all(seed)

omnigen2/utils/teacache_util.py

@@ -0,0 +1,43 @@
+"""
+Utility for TeaCache
+
+Copyright 2025 BAAI, The OmniGen2 Team and The HuggingFace 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.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+
+@dataclass
+class TeaCacheParams:
+    """
+    TeaCache parameters for `OmniGen2Transformer2DModel`
+    See https://github.com/ali-vilab/TeaCache/ for a more comprehensive understanding
+
+    Args:
+        previous_residual (Optional[torch.Tensor]):
+            The tensor difference between the output and the input of the transformer layers from the previous timestep.
+        previous_modulated_inp (Optional[torch.Tensor]):
+            The modulated input from the previous timestep used to indicate the change of the transformer layer's output.
+        accumulated_rel_l1_distance (float):
+            The accumulated relative L1 distance.
+        is_first_or_last_step (bool):
+            Whether the current timestep is the first or last step.
+    """
+    previous_residual: Optional[torch.Tensor] = None
+    previous_modulated_inp: Optional[torch.Tensor] = None
+    accumulated_rel_l1_distance: float = 0
+    is_first_or_last_step: bool = False