[New model] Support model qwen image layered (#381)

Signed-off-by: dengyunyang <584797741@qq.com>

docs/models/supported_models.md

@@ -22,6 +22,7 @@ th {
 | `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` |
 | `QwenImageEditPipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image-Edit` |
 | `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2509 | `Qwen/Qwen-Image-Edit-2509` |
+| `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` |
 |`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` |
 | `WanPipeline` | Wan2.2 | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` |
 | `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` |

examples/offline_inference/image_to_image/image_edit.py

@@ -22,6 +22,17 @@ Usage (multiple images):
         --cfg_scale 4.0 \
         --guidance_scale 1.0
 
+Usage (layered):
+    python image_layered.py \
+        --model "Qwen/Qwen-Image-Layered" \
+        --image input.png \
+        --prompt "" \
+        --output "layered" \
+        --num_inference_steps 50 \
+        --cfg_scale 4.0 \
+        --layers 4 \
+        --color-format "RGBA"
+
 For more options, run:
     python image_edit.py --help
 """
@@ -100,7 +111,7 @@ def parse_args() -> argparse.Namespace:
         "--output",
         type=str,
         default="output_image_edit.png",
-        help="Path to save the edited image (PNG).",
+        help=("Path to save the edited image (PNG). Or prefix for Qwen-Image-Layered model save images(PNG)."),
     )
     parser.add_argument(
         "--num_outputs_per_prompt",
@@ -132,6 +143,21 @@ def parse_args() -> argparse.Namespace:
         help="Number of GPUs used for ulysses sequence parallelism.",
     )
 
+    parser.add_argument("--layers", type=int, default=4, help="Number of layers to decompose the input image into.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=640,
+        help="Bucket in (640, 1024) to determine the condition and output resolution",
+    )
+
+    parser.add_argument(
+        "--color-format",
+        type=str,
+        default="RGB",
+        help="For Qwen-Image-Layered, set to RGBA.",
+    )
+
     return parser.parse_args()
 
 
@@ -143,7 +169,8 @@ def main():
     for image_path in args.image:
         if not os.path.exists(image_path):
             raise FileNotFoundError(f"Input image not found: {image_path}")
-        img = Image.open(image_path).convert("RGB")
+
+        img = Image.open(image_path).convert(args.color_format)
         input_images.append(img)
 
     # Use single image or list based on number of inputs
@@ -226,6 +253,7 @@ def main():
         guidance_scale=args.guidance_scale,
         num_inference_steps=args.num_inference_steps,
         num_outputs_per_prompt=args.num_outputs_per_prompt,
+        layers=args.layers,
     )
     generation_end = time.perf_counter()
     generation_time = generation_end - generation_start
@@ -240,13 +268,19 @@ def main():
     stem = output_path.stem or "output_image_edit"
 
     if args.num_outputs_per_prompt <= 1:
-        images[0].save(output_path)
-        print(f"Saved edited image to {os.path.abspath(output_path)}")
+        img = images[0]
+        img = img if isinstance(img, list) else [img]
+        for sub_idx, sub_img in enumerate(img):
+            save_path = output_path.parent / f"{stem}_{sub_idx}{suffix}"
+            sub_img.save(save_path)
+            print(f"Saved edited image to {os.path.abspath(save_path)}")
     else:
         for idx, img in enumerate(images):
-            save_path = output_path.parent / f"{stem}_{idx}{suffix}"
-            img.save(save_path)
-            print(f"Saved edited image to {os.path.abspath(save_path)}")
+            img = img if isinstance(img, list) else [img]
+            for sub_idx, sub_img in enumerate(img):
+                save_path = output_path.parent / f"{stem}_{idx}_{sub_idx}{suffix}"
+                sub_img.save(save_path)
+                print(f"Saved edited image to {os.path.abspath(save_path)}")
 
 
 if __name__ == "__main__":

vllm_omni/diffusion/diffusion_engine.py

@@ -48,7 +48,7 @@ class DiffusionEngine:
             logger.info("Generation completed successfully.")
 
             postprocess_start_time = time.time()
-            result = self.post_process_func(output.output)
+            result = self.post_process_func(output.output) if self.post_process_func is not None else output.output
             postprocess_time = time.time() - postprocess_start_time
             logger.info(f"Post-processing completed in {postprocess_time:.4f} seconds")
 

vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image_layered.py

@@ -0,0 +1,891 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import inspect
+import json
+import logging
+import math
+import os
+from collections.abc import Iterable
+from typing import Any
+
+import numpy as np
+import PIL.Image
+import torch
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers.scheduling_flow_match_euler_discrete import (
+    FlowMatchEulerDiscreteScheduler,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from torch import nn
+from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor
+from vllm.model_executor.models.utils import AutoWeightsLoader
+
+from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
+from vllm_omni.diffusion.distributed.utils import get_local_device
+from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
+from vllm_omni.diffusion.models.qwen_image.autoencoder_kl_qwenimage import (
+    AutoencoderKLQwenImage,
+)
+from vllm_omni.diffusion.models.qwen_image.qwen_image_transformer import (
+    QwenImageTransformer2DModel,
+)
+from vllm_omni.diffusion.request import OmniDiffusionRequest
+from vllm_omni.model_executor.model_loader.weight_utils import (
+    download_weights_from_hf_specific,
+)
+
+logger = logging.getLogger(__name__)
+
+
+# Interface called in diffusion engine
+def get_qwen_image_layered_pre_process_func(
+    od_config: OmniDiffusionConfig,
+):
+    """Pre-processing function for QwenImageLayeredPipeline."""
+    model_name = od_config.model
+    if os.path.exists(model_name):
+        model_path = model_name
+    else:
+        model_path = download_weights_from_hf_specific(model_name, None, ["*"])
+
+    vae_config_path = os.path.join(model_path, "vae/config.json")
+    with open(vae_config_path) as f:
+        vae_config = json.load(f)
+        vae_scale_factor = 2 ** len(vae_config["temporal_downsample"]) if "temporal_downsample" in vae_config else 8
+        latent_channels = vae_config.get("z_dim", 16)
+
+    # QwenImage latents are turned into 2x2 patches and packed.
+    # This means the latent width and height has to be divisible
+    # by the patch size. So the vae scale factor is multiplied
+    # by the patch size to account for this
+    image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor * 2)
+
+    def pre_process_func(
+        requests: list[OmniDiffusionRequest],
+    ):
+        """Pre-process requests for QwenImageLayeredPipeline."""
+        for req in requests:
+            image = req.pil_image
+
+            # 1. calculate dimensions
+            image_size = image[0].size if isinstance(image, list) else image.size
+            assert req.resolution in [640, 1024], f"resolution must be either 640 or 1024, but got {req.resolution}"
+            calculated_width, calculated_height = calculate_dimensions(
+                req.resolution * req.resolution, image_size[0] / image_size[1]
+            )
+            height = calculated_height
+            width = calculated_width
+
+            multiple_of = vae_scale_factor * 2
+            width = width // multiple_of * multiple_of
+            height = height // multiple_of * multiple_of
+
+            # Store calculated dimensions in request
+            req.calculated_height = calculated_height
+            req.calculated_width = calculated_width
+            req.height = height
+            req.width = width
+
+            # 2. Preprocess image
+            if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == latent_channels):
+                image = image_processor.resize(image, calculated_height, calculated_width)
+                prompt_image = image
+                image = image_processor.preprocess(image, calculated_height, calculated_width)
+                image = image.unsqueeze(2)
+                # image = image.to(dtype=self.text_encoder.dtype)  # do it later
+
+                # Store preprocessed image and prompt image in request
+                req.preprocessed_image = image
+                req.prompt_image = prompt_image
+
+        return requests
+
+    return pre_process_func
+
+
+# Copied from diffusers to avoid version coupling.
+# Upstream code merged on 2025-12-18.
+def calculate_dimensions(target_area, ratio):
+    width = math.sqrt(target_area * ratio)
+    height = width / ratio
+
+    width = round(width / 32) * 32
+    height = round(height / 32) * 32
+
+    return width, height
+
+
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: int | None = None,
+    device: str | torch.device | None = None,
+    timesteps: list[int] | None = None,
+    sigmas: list[float] | None = None,
+    **kwargs,
+):
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    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)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, 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
+
+
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class QwenImageLayeredPipeline(
+    nn.Module,
+):
+    def __init__(
+        self,
+        *,
+        od_config: OmniDiffusionConfig,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.od_config = od_config
+        self.device = get_local_device()
+        model = od_config.model
+        # Check if model is a local path
+        local_files_only = os.path.exists(model)
+
+        # modules keep same as transformers & diffusers
+        self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
+            model, subfolder="scheduler", local_files_only=local_files_only
+        )
+        self.text_encoder = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+            model, subfolder="text_encoder", local_files_only=local_files_only
+        )
+        self.vae = AutoencoderKLQwenImage.from_pretrained(model, subfolder="vae", local_files_only=local_files_only).to(
+            self.device
+        )
+        self.tokenizer = Qwen2Tokenizer.from_pretrained(model, subfolder="tokenizer", local_files_only=local_files_only)
+        self.processor = Qwen2VLProcessor.from_pretrained(
+            model, subfolder="processor", local_files_only=local_files_only
+        )
+
+        # modules re-implemented for vLLM-Omni
+        self.weights_sources = [
+            DiffusersPipelineLoader.ComponentSource(
+                model_or_path=od_config.model,
+                subfolder="transformer",
+                revision=None,
+                prefix="transformer.",
+                fall_back_to_pt=True,
+            )
+        ]
+
+        use_additional_t_cond = od_config.tf_model_config.use_additional_t_cond
+        zero_cond_t = od_config.tf_model_config.zero_cond_t
+        use_layer3d_rope = od_config.tf_model_config.use_layer3d_rope
+        guidance_embeds = od_config.tf_model_config.guidance_embeds
+
+        self.transformer = QwenImageTransformer2DModel(
+            od_config=od_config,
+            use_additional_t_cond=use_additional_t_cond,
+            zero_cond_t=zero_cond_t,
+            use_layer3d_rope=use_layer3d_rope,
+            guidance_embeds=guidance_embeds,
+        )
+
+        # Pipeline configuration & processing parameters
+        self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16
+        # QwenImage latents are turned into 2x2 patches and packed.
+        # This means the latent width and height has to be divisible
+        # by the patch size. So the vae scale factor is multiplied
+        # by the patch size to account for this
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
+        self.vl_processor = self.processor
+        self.tokenizer_max_length = 1024
+
+        self.prompt_template_encode = (
+            "<|im_start|>system\nDescribe the image by detailing the color, "
+            "shape, size, texture, quantity, text, spatial relationships of the objects and "
+            "background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        )
+        self.prompt_template_encode_start_idx = 34
+        self.image_caption_prompt_cn = (
+            """<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n# """
+            """图像标注器\n你是一个专业的图像标注器。请基于输入图像，撰写图注:\n1.
+使用自然、描述性的语言撰写图注，不要使用结构化形式或富文本形式。\n2. 通过加入以下内容，丰富图注细节：\n """
+            """- 对象的属性：如数量、颜色、形状、大小、位置、材质、状态、动作等\n -
+对象间的视觉关系：如空间关系、功能关系、动作关系、从属关系、比较关系、因果关系等\n - 环境细节：例如天气、"""
+            """光照、颜色、纹理、气氛等\n - 文字内容：识别图像中清晰可见的文字，不做翻译和解释，用引号在"""
+            """图注中强调\n3.
+保持真实性与准确性：\n - 不要使用笼统的描述\n -
+描述图像中所有可见的信息，但不要加入没有在图像中出现"""
+            """的内容\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>assistant\n"""
+        )
+        self.image_caption_prompt_en = (
+            """<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"""
+            """<|im_start|>user\n# Image Annotator\nYou are a professional
+image annotator. Please write an image caption based on the input image:\n1. Write the caption using natural,
+descriptive language without structured formats or rich text.\n2. Enrich caption details by including: \n - Object
+attributes, such as quantity, color, shape, size, material, state, position, actions, and so on\n - Vision Relations
+between objects, such as spatial relations, functional relations, possessive relations, attachment relations, action
+relations, comparative relations, causal relations, and so on\n - Environmental details, such as weather, lighting,
+colors, textures, atmosphere, and so on\n - Identify the text clearly visible in the image, without translation or
+explanation, and highlight it in the caption with quotation marks\n3. Maintain authenticity and accuracy:\n - Avoid
+generalizations\n - Describe all visible information in the image, while do not add information not explicitly shown in
+the image\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>assistant\n"""
+        )
+        self.default_sample_size = 128
+
+        # VLLM-omni related, Initialize cache backend to None (will be set by worker if needed)
+        self._cache_backend = None
+        self.stage = None
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_embeds_mask=None,
+        negative_prompt_embeds_mask=None,
+        max_sequence_length=None,
+    ):
+        if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
+            logger.warning(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} "
+                f"but are {height} and {width}. Dimensions will be resized accordingly"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and prompt_embeds_mask is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate"
+                " `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+        if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make"
+                " sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to "
+                "generate `negative_prompt_embeds`."
+            )
+
+        if max_sequence_length is not None and max_sequence_length > 1024:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}")
+
+    def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor):
+        bool_mask = mask.bool()
+        valid_lengths = bool_mask.sum(dim=1)
+        selected = hidden_states[bool_mask]
+        split_result = torch.split(selected, valid_lengths.tolist(), dim=0)
+
+        return split_result
+
+    def _get_qwen_prompt_embeds(
+        self,
+        prompt: str | list[str] | None = None,
+        device: torch.device | None = None,
+        dtype: torch.dtype | None = None,
+    ):
+        device = device or self.device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        template = self.prompt_template_encode
+        drop_idx = self.prompt_template_encode_start_idx
+        txt = [template.format(e) for e in prompt]
+        txt_tokens = self.tokenizer(
+            txt,
+            padding=True,
+            return_tensors="pt",
+        ).to(device)
+        encoder_hidden_states = self.text_encoder(
+            input_ids=txt_tokens.input_ids,
+            attention_mask=txt_tokens.attention_mask,
+            output_hidden_states=True,
+        )
+        hidden_states = encoder_hidden_states.hidden_states[-1]
+        split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask)
+        split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
+        attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states]
+        max_seq_len = max([e.size(0) for e in split_hidden_states])
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states]
+        )
+        encoder_attention_mask = torch.stack(
+            [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list]
+        )
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        return prompt_embeds, encoder_attention_mask
+
+    def encode_prompt(
+        self,
+        prompt: str | list[str],
+        device: torch.device | None = None,
+        num_images_per_prompt: int = 1,
+        prompt_embeds: torch.Tensor | None = None,
+        prompt_embeds_mask: torch.Tensor | None = None,
+        max_sequence_length: int = 1024,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            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.
+        """
+        device = device or self.device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt)
+
+        prompt_embeds = prompt_embeds[:, :max_sequence_length]
+        prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length]
+
+        _, seq_len, _ = prompt_embeds.shape
+        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_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len)
+
+        return prompt_embeds, prompt_embeds_mask
+
+    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+        if isinstance(generator, list):
+            image_latents = [
+                retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax")
+                for i in range(image.shape[0])
+            ]
+            image_latents = torch.cat(image_latents, dim=0)
+        else:
+            image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax")
+        latents_mean = (
+            torch.tensor(self.vae.config.latents_mean)
+            .view(1, self.latent_channels, 1, 1, 1)
+            .to(image_latents.device, image_latents.dtype)
+        )
+        latents_std = (
+            torch.tensor(self.vae.config.latents_std)
+            .view(1, self.latent_channels, 1, 1, 1)
+            .to(image_latents.device, image_latents.dtype)
+        )
+        image_latents = (image_latents - latents_mean) / latents_std
+
+        return image_latents
+
+    def prepare_latents(
+        self,
+        image,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        layers,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
+
+        shape = (
+            batch_size,
+            layers + 1,
+            num_channels_latents,
+            height,
+            width,
+        )  ### the generated first image is combined image
+
+        image_latents = None
+        if image is not None:
+            image = image.to(device=device, dtype=dtype)
+            if image.shape[1] != self.latent_channels:
+                image_latents = self._encode_vae_image(image=image, generator=generator)
+            else:
+                image_latents = image
+            if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+                # expand init_latents for batch_size
+                additional_image_per_prompt = batch_size // image_latents.shape[0]
+                image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+            elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+                raise ValueError(
+                    f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+                )
+            else:
+                image_latents = torch.cat([image_latents], dim=0)
+
+            image_latent_height, image_latent_width = image_latents.shape[3:]
+            image_latents = image_latents.permute(0, 2, 1, 3, 4)  # (b, c, f, h, w) -> (b, f, c, h, w)
+            image_latents = self._pack_latents(
+                image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width, 1
+            )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width, layers + 1)
+        else:
+            latents = latents.to(device=device, dtype=dtype)
+
+        return latents, image_latents
+
+    def get_image_caption(self, prompt_image, use_en_prompt=True, device=None):
+        if use_en_prompt:
+            prompt = self.image_caption_prompt_en
+        else:
+            prompt = self.image_caption_prompt_cn
+        model_inputs = self.vl_processor(
+            text=prompt,
+            images=prompt_image,
+            padding=True,
+            return_tensors="pt",
+        ).to(device)
+        generated_ids = self.text_encoder.generate(**model_inputs, max_new_tokens=512)
+        generated_ids_trimmed = [
+            out_ids[len(in_ids) :] for in_ids, out_ids in zip(model_inputs.input_ids, generated_ids)
+        ]
+        output_text = self.vl_processor.batch_decode(
+            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+        )[0]
+        return output_text.strip()
+
+    @staticmethod
+    def _pack_latents(latents, batch_size, num_channels_latents, height, width, layers):
+        latents = latents.view(batch_size, layers, num_channels_latents, height // 2, 2, width // 2, 2)
+        latents = latents.permute(0, 1, 3, 5, 2, 4, 6)
+        latents = latents.reshape(batch_size, layers * (height // 2) * (width // 2), num_channels_latents * 4)
+
+        return latents
+
+    @staticmethod
+    def _unpack_latents(latents, height, width, layers, vae_scale_factor):
+        batch_size, num_patches, channels = latents.shape
+
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (vae_scale_factor * 2))
+        width = 2 * (int(width) // (vae_scale_factor * 2))
+
+        latents = latents.view(batch_size, layers + 1, height // 2, width // 2, channels // 4, 2, 2)
+        latents = latents.permute(0, 1, 4, 2, 5, 3, 6)
+
+        latents = latents.reshape(batch_size, layers + 1, channels // (2 * 2), height, width)
+        latents = latents.permute(0, 2, 1, 3, 4)  # (b, c, f, h, w)
+
+        return latents
+
+    def diffuse(
+        self,
+        prompt_embeds,
+        prompt_embeds_mask,
+        negative_prompt_embeds,
+        negative_prompt_embeds_mask,
+        latents,
+        image_latents,
+        img_shapes,
+        txt_seq_lens,
+        negative_txt_seq_lens,
+        timesteps,
+        do_true_cfg,
+        guidance,
+        true_cfg_scale,
+        cfg_normalize,
+        is_rgb,
+    ):
+        self.scheduler.set_begin_index(0)
+        for i, t in enumerate(timesteps):
+            if self.interrupt:
+                continue
+
+            self._current_timestep = t
+
+            latent_model_input = latents
+            if image_latents is not None:
+                latent_model_input = torch.cat([latents, image_latents], dim=1)
+
+            # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+            timestep = t.expand(latents.shape[0]).to(latents.dtype)
+            # Forward pass for positive prompt (or unconditional if no CFG)
+            # cache_branch is passed to hook for CFG-aware state management
+            transformer_kwargs = {
+                "hidden_states": latent_model_input,
+                "timestep": timestep / 1000,
+                "guidance": guidance,
+                "encoder_hidden_states_mask": prompt_embeds_mask,
+                "encoder_hidden_states": prompt_embeds,
+                "img_shapes": img_shapes,
+                "txt_seq_lens": txt_seq_lens,
+                "attention_kwargs": self.attention_kwargs,
+                "additional_t_cond": is_rgb,
+                "return_dict": False,
+            }
+            if self._cache_backend is not None:
+                transformer_kwargs["cache_branch"] = "positive"
+
+            noise_pred = self.transformer(**transformer_kwargs)[0]
+            noise_pred = noise_pred[:, : latents.size(1)]
+
+            if do_true_cfg:
+                neg_transformer_kwargs = {
+                    "hidden_states": latent_model_input,
+                    "timestep": timestep / 1000,
+                    "guidance": guidance,
+                    "encoder_hidden_states_mask": negative_prompt_embeds_mask,
+                    "encoder_hidden_states": negative_prompt_embeds,
+                    "img_shapes": img_shapes,
+                    "txt_seq_lens": negative_txt_seq_lens,
+                    "attention_kwargs": self.attention_kwargs,
+                    "additional_t_cond": is_rgb,
+                    "return_dict": False,
+                }
+                if self._cache_backend is not None:
+                    neg_transformer_kwargs["cache_branch"] = "negative"
+                neg_noise_pred = self.transformer(**neg_transformer_kwargs)[0]
+
+                neg_noise_pred = neg_noise_pred[:, : latents.size(1)]
+                comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
+
+                if cfg_normalize:
+                    cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
+                    noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
+                    noise_pred = comb_pred * (cond_norm / noise_norm)
+                else:
+                    noise_pred = comb_pred
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+            # no need callback now
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    def forward(
+        self,
+        req: OmniDiffusionRequest,
+        image: PIL.Image.Image | torch.Tensor | None = None,
+        prompt: str | list[str] | None = None,
+        negative_prompt: str | list[str] = None,
+        true_cfg_scale: float = 4.0,
+        layers: int | None = 4,
+        num_inference_steps: int = 50,
+        sigmas: list[float] | None = None,
+        guidance_scale: float | None = None,
+        num_images_per_prompt: int = 1,
+        generator: torch.Generator | list[torch.Generator] | None = None,
+        latents: torch.Tensor | None = None,
+        prompt_embeds: torch.Tensor | None = None,
+        prompt_embeds_mask: torch.Tensor | None = None,
+        negative_prompt_embeds: torch.Tensor | None = None,
+        negative_prompt_embeds_mask: torch.Tensor | None = None,
+        output_type: str | None = "pil",
+        attention_kwargs: dict[str, Any] | None = None,
+        max_sequence_length: int = 512,
+        resolution: int = 640,
+        cfg_normalize: bool = False,
+        use_en_prompt: bool = False,
+    ) -> DiffusionOutput:
+        """Forward pass for image layered."""
+
+        # 1. Get preprocessed image from request (pre-processing is done in DiffusionEngine)
+        # Override parameters from request if provided
+        prompt = req.prompt if req.prompt is not None else prompt
+        layers = req.layers if req.layers is not None else layers
+        resolution = req.resolution if req.resolution is not None else resolution
+        cfg_normalize = req.cfg_normalize if req.cfg_normalize is not None else cfg_normalize
+        use_en_prompt = req.use_en_prompt if req.use_en_prompt is not None else use_en_prompt
+        negative_prompt = req.negative_prompt if req.negative_prompt is not None else negative_prompt
+        num_inference_steps = req.num_inference_steps or num_inference_steps
+        generator = req.generator or generator
+        true_cfg_scale = req.true_cfg_scale or true_cfg_scale
+        req_num_outputs = getattr(req, "num_outputs_per_prompt", None)
+        if req_num_outputs and req_num_outputs > 0:
+            num_images_per_prompt = req_num_outputs
+
+        if hasattr(req, "preprocessed_image"):
+            prompt_image = req.prompt_image
+            image = req.preprocessed_image
+            image = image.to(dtype=self.text_encoder.dtype)  # Now we get the type
+            calculated_height = req.calculated_height
+            calculated_width = req.calculated_width
+            height = req.height
+            width = req.width
+        else:
+            # fallback to run pre-processing in pipeline (debug only)
+            image_size = image[0].size if isinstance(image, list) else image.size
+            assert resolution in [640, 1024], f"resolution must be either 640 or 1024, but got {resolution}"
+            calculated_width, calculated_height = calculate_dimensions(
+                resolution * resolution, image_size[0] / image_size[1]
+            )
+            height = calculated_height
+            width = calculated_width
+
+            multiple_of = self.vae_scale_factor * 2
+            width = width // multiple_of * multiple_of
+            height = height // multiple_of * multiple_of
+
+            if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels):
+                image = self.image_processor.resize(image, calculated_height, calculated_width)
+                prompt_image = image
+                image = self.image_processor.preprocess(image, calculated_height, calculated_width)
+                image = image.unsqueeze(2)
+                image = image.to(dtype=self.text_encoder.dtype)
+
+        # 2. check inputs
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            negative_prompt_embeds_mask=negative_prompt_embeds_mask,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        # 3. encode prompot & negative prompt
+        if prompt is None or prompt == "" or prompt == " ":
+            prompt = self.get_image_caption(prompt_image, use_en_prompt=use_en_prompt, device=self.device)
+        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]
+
+        has_neg_prompt = negative_prompt is not None or (
+            negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None
+        )
+
+        if true_cfg_scale > 1 and not has_neg_prompt:
+            logger.warning(
+                f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free "
+                f"guidance is not enabled since no negative_prompt is provided."
+            )
+        elif true_cfg_scale <= 1 and has_neg_prompt:
+            logger.warning(
+                " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1"
+            )
+
+        do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
+        prompt_embeds, prompt_embeds_mask = self.encode_prompt(
+            prompt=prompt,
+            prompt_embeds=prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            device=self.device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+        )
+        if do_true_cfg:
+            negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(
+                prompt=negative_prompt,
+                prompt_embeds=negative_prompt_embeds,
+                prompt_embeds_mask=negative_prompt_embeds_mask,
+                device=self.device,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.in_channels // 4
+        latents, image_latents = self.prepare_latents(
+            image,
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            layers,
+            prompt_embeds.dtype,
+            self.device,
+            generator,
+            latents,
+        )
+        img_shapes = [
+            [
+                *[
+                    (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)
+                    for _ in range(layers + 1)
+                ],
+                (1, calculated_height // self.vae_scale_factor // 2, calculated_width // self.vae_scale_factor // 2),
+            ]
+        ] * batch_size
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas
+        base_seqlen = 256 * 256 / 16 / 16
+        mu = (image_latents.shape[1] / base_seqlen) ** 0.5
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            self.device,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        self._num_timesteps = len(timesteps)
+
+        # handle guidance
+        if self.transformer.guidance_embeds and guidance_scale is None:
+            raise ValueError("guidance_scale is required for guidance-distilled model.")
+        elif self.transformer.guidance_embeds:
+            guidance = torch.full([1], guidance_scale, device=self.device, dtype=torch.float32)
+            guidance = guidance.expand(latents.shape[0])
+        elif not self.transformer.guidance_embeds and guidance_scale is not None:
+            logger.warning(
+                f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled."
+            )
+            guidance = None
+        elif not self.transformer.guidance_embeds and guidance_scale is None:
+            guidance = None
+
+        if self.attention_kwargs is None:
+            self._attention_kwargs = {}
+
+        txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist() if prompt_embeds_mask is not None else None
+        negative_txt_seq_lens = (
+            negative_prompt_embeds_mask.sum(dim=1).tolist() if negative_prompt_embeds_mask is not None else None
+        )
+        is_rgb = torch.tensor([0] * batch_size).to(device=self.device, dtype=torch.long)
+
+        latents = self.diffuse(
+            prompt_embeds,
+            prompt_embeds_mask,
+            negative_prompt_embeds,
+            negative_prompt_embeds_mask,
+            latents,
+            image_latents,
+            img_shapes,
+            txt_seq_lens,
+            negative_txt_seq_lens,
+            timesteps,
+            do_true_cfg,
+            guidance,
+            true_cfg_scale,
+            cfg_normalize,
+            is_rgb,
+        )
+
+        self._current_timestep = None
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = self._unpack_latents(latents, height, width, layers, self.vae_scale_factor)
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+
+            b, c, f, h, w = latents.shape
+
+            latents = latents[:, :, 1:]  # remove the first frame as it is the origin input
+
+            latents = latents.permute(0, 2, 1, 3, 4).view(-1, c, 1, h, w)
+
+            image = self.vae.decode(latents, return_dict=False)[0]  # (b f) c 1 h w
+
+            image = image.squeeze(2)
+            # Maybe extract post process in the future
+            image = self.image_processor.postprocess(image, output_type=output_type)
+            images = []
+            for bidx in range(b):
+                images.append(image[bidx * f : (bidx + 1) * f])
+
+        return DiffusionOutput(output=images)
+
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        loader = AutoWeightsLoader(self)
+        return loader.load_weights(weights)

vllm_omni/diffusion/models/qwen_image/qwen_image_transformer.py

@@ -32,21 +32,144 @@ logger = init_logger(__name__)
 
 
 class QwenTimestepProjEmbeddings(nn.Module):
-    def __init__(self, embedding_dim):
+    def __init__(self, embedding_dim, use_additional_t_cond=False):
         super().__init__()
 
         self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
         self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+        self.use_additional_t_cond = use_additional_t_cond
+        if use_additional_t_cond:
+            self.addition_t_embedding = nn.Embedding(2, embedding_dim)
 
-    def forward(self, timestep, hidden_states):
+    def forward(self, timestep, hidden_states, addition_t_cond=None):
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))  # (N, D)
 
         conditioning = timesteps_emb
+        if self.use_additional_t_cond:
+            if addition_t_cond is None:
+                raise ValueError("When additional_t_cond is True, addition_t_cond must be provided.")
+            addition_t_emb = self.addition_t_embedding(addition_t_cond)
+            addition_t_emb = addition_t_emb.to(dtype=hidden_states.dtype)
+            conditioning = conditioning + addition_t_emb
 
         return conditioning
 
 
+class QwenEmbedLayer3DRope(nn.Module):
+    def __init__(self, theta: int, axes_dim: list[int], scale_rope=False):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+        pos_index = torch.arange(4096)
+        neg_index = torch.arange(4096).flip(0) * -1 - 1
+        self.pos_freqs = torch.cat(
+            [
+                self.rope_params(pos_index, self.axes_dim[0], self.theta),
+                self.rope_params(pos_index, self.axes_dim[1], self.theta),
+                self.rope_params(pos_index, self.axes_dim[2], self.theta),
+            ],
+            dim=1,
+        )
+        self.neg_freqs = torch.cat(
+            [
+                self.rope_params(neg_index, self.axes_dim[0], self.theta),
+                self.rope_params(neg_index, self.axes_dim[1], self.theta),
+                self.rope_params(neg_index, self.axes_dim[2], self.theta),
+            ],
+            dim=1,
+        )
+
+        self.scale_rope = scale_rope
+
+    def rope_params(self, index, dim, theta=10000):
+        """
+        Args:
+            index: [0, 1, 2, 3] 1D Tensor representing the position index of the token
+        """
+        assert dim % 2 == 0
+        freqs = torch.outer(index, 1.0 / torch.pow(theta, torch.arange(0, dim, 2).to(torch.float32).div(dim)))
+        freqs = torch.polar(torch.ones_like(freqs), freqs)
+        return freqs
+
+    def forward(self, video_fhw, txt_seq_lens, device):
+        """
+        Args: video_fhw: [frame, height, width] a list of 3 integers representing the shape of the video Args:
+        txt_length: [bs] a list of 1 integers representing the length of the text
+        """
+        if self.pos_freqs.device != device:
+            self.pos_freqs = self.pos_freqs.to(device)
+            self.neg_freqs = self.neg_freqs.to(device)
+
+        if isinstance(video_fhw, list):
+            video_fhw = video_fhw[0]
+        if not isinstance(video_fhw, list):
+            video_fhw = [video_fhw]
+
+        vid_freqs = []
+        max_vid_index = 0
+        layer_num = len(video_fhw) - 1
+        for idx, fhw in enumerate(video_fhw):
+            frame, height, width = fhw
+            if idx != layer_num:
+                video_freq = self._compute_video_freqs(frame, height, width, idx)
+            else:
+                ### For the condition image, we set the layer index to -1
+                video_freq = self._compute_condition_freqs(frame, height, width)
+            video_freq = video_freq.to(device)
+            vid_freqs.append(video_freq)
+
+            if self.scale_rope:
+                max_vid_index = max(height // 2, width // 2, max_vid_index)
+            else:
+                max_vid_index = max(height, width, max_vid_index)
+
+        max_vid_index = max(max_vid_index, layer_num)
+        max_len = max(txt_seq_lens)
+        txt_freqs = self.pos_freqs[max_vid_index : max_vid_index + max_len, ...]
+        vid_freqs = torch.cat(vid_freqs, dim=0)
+
+        return vid_freqs, txt_freqs
+
+    @functools.cache
+    def _compute_video_freqs(self, frame, height, width, idx=0):
+        seq_lens = frame * height * width
+        freqs_pos = self.pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+        freqs_neg = self.neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+
+        freqs_frame = freqs_pos[0][idx : idx + frame].view(frame, 1, 1, -1).expand(frame, height, width, -1)
+        if self.scale_rope:
+            freqs_height = torch.cat([freqs_neg[1][-(height - height // 2) :], freqs_pos[1][: height // 2]], dim=0)
+            freqs_height = freqs_height.view(1, height, 1, -1).expand(frame, height, width, -1)
+            freqs_width = torch.cat([freqs_neg[2][-(width - width // 2) :], freqs_pos[2][: width // 2]], dim=0)
+            freqs_width = freqs_width.view(1, 1, width, -1).expand(frame, height, width, -1)
+        else:
+            freqs_height = freqs_pos[1][:height].view(1, height, 1, -1).expand(frame, height, width, -1)
+            freqs_width = freqs_pos[2][:width].view(1, 1, width, -1).expand(frame, height, width, -1)
+
+        freqs = torch.cat([freqs_frame, freqs_height, freqs_width], dim=-1).reshape(seq_lens, -1)
+        return freqs.clone().contiguous()
+
+    @functools.cache
+    def _compute_condition_freqs(self, frame, height, width):
+        seq_lens = frame * height * width
+        freqs_pos = self.pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+        freqs_neg = self.neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+
+        freqs_frame = freqs_neg[0][-1:].view(frame, 1, 1, -1).expand(frame, height, width, -1)
+        if self.scale_rope:
+            freqs_height = torch.cat([freqs_neg[1][-(height - height // 2) :], freqs_pos[1][: height // 2]], dim=0)
+            freqs_height = freqs_height.view(1, height, 1, -1).expand(frame, height, width, -1)
+            freqs_width = torch.cat([freqs_neg[2][-(width - width // 2) :], freqs_pos[2][: width // 2]], dim=0)
+            freqs_width = freqs_width.view(1, 1, width, -1).expand(frame, height, width, -1)
+        else:
+            freqs_height = freqs_pos[1][:height].view(1, height, 1, -1).expand(frame, height, width, -1)
+            freqs_width = freqs_pos[2][:width].view(1, 1, width, -1).expand(frame, height, width, -1)
+
+        freqs = torch.cat([freqs_frame, freqs_height, freqs_width], dim=-1).reshape(seq_lens, -1)
+        return freqs.clone().contiguous()
+
+
 class QwenEmbedRope(nn.Module):
     def __init__(self, theta: int, axes_dim: list[int], scale_rope=False):
         super().__init__()
@@ -493,6 +616,8 @@ class QwenImageTransformer2DModel(nn.Module):
         guidance_embeds: bool = False,  # TODO: this should probably be removed
         axes_dims_rope: tuple[int, int, int] = (16, 56, 56),
         zero_cond_t: bool = False,
+        use_additional_t_cond: bool = False,
+        use_layer3d_rope: bool = False,
     ):
         super().__init__()
         model_config = od_config.tf_model_config
@@ -503,9 +628,14 @@ class QwenImageTransformer2DModel(nn.Module):
         self.inner_dim = num_attention_heads * attention_head_dim
         self.guidance_embeds = guidance_embeds
 
-        self.pos_embed = QwenEmbedRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
+        if not use_layer3d_rope:
+            self.pos_embed = QwenEmbedRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
+        else:
+            self.pos_embed = QwenEmbedLayer3DRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
 
-        self.time_text_embed = QwenTimestepProjEmbeddings(embedding_dim=self.inner_dim)
+        self.time_text_embed = QwenTimestepProjEmbeddings(
+            embedding_dim=self.inner_dim, use_additional_t_cond=use_additional_t_cond
+        )
 
         self.txt_norm = RMSNorm(joint_attention_dim, eps=1e-6)
 
@@ -540,6 +670,7 @@ class QwenImageTransformer2DModel(nn.Module):
         txt_seq_lens: list[int] | None = None,
         guidance: torch.Tensor = None,  # TODO: this should probably be removed
         attention_kwargs: dict[str, Any] | None = None,
+        additional_t_cond=None,
         return_dict: bool = True,
     ) -> torch.Tensor | Transformer2DModelOutput:
         """
@@ -599,9 +730,9 @@ class QwenImageTransformer2DModel(nn.Module):
             guidance = guidance.to(hidden_states.dtype) * 1000
 
         temb = (
-            self.time_text_embed(timestep, hidden_states)
+            self.time_text_embed(timestep, hidden_states, additional_t_cond)
             if guidance is None
-            else self.time_text_embed(timestep, guidance, hidden_states)
+            else self.time_text_embed(timestep, guidance, hidden_states, additional_t_cond)
         )
 
         image_rotary_emb = self.pos_embed(img_shapes, txt_seq_lens, device=hidden_states.device)

vllm_omni/diffusion/registry.py

@@ -24,6 +24,11 @@ _DIFFUSION_MODELS = {
         "pipeline_qwen_image_edit_plus",
         "QwenImageEditPlusPipeline",
     ),
+    "QwenImageLayeredPipeline": (
+        "qwen_image",
+        "pipeline_qwen_image_layered",
+        "QwenImageLayeredPipeline",
+    ),
     "ZImagePipeline": (
         "z_image",
         "pipeline_z_image",
@@ -93,6 +98,7 @@ _DIFFUSION_PRE_PROCESS_FUNCS = {
     # where mod_folder and mod_relname are  defined and mapped using `_DIFFUSION_MODELS` via the `arch` key
     "QwenImageEditPipeline": "get_qwen_image_edit_pre_process_func",
     "QwenImageEditPlusPipeline": "get_qwen_image_edit_plus_pre_process_func",
+    "QwenImageLayeredPipeline": "get_qwen_image_layered_pre_process_func",
 }
 
 
@@ -107,9 +113,7 @@ def _load_process_func(od_config: OmniDiffusionConfig, func_name: str):
 
 def get_diffusion_post_process_func(od_config: OmniDiffusionConfig):
     if od_config.model_class_name not in _DIFFUSION_POST_PROCESS_FUNCS:
-        raise ValueError(
-            f"Post process function for model class {od_config.model_class_name} not found in diffusion model registry."
-        )
+        return None
     func_name = _DIFFUSION_POST_PROCESS_FUNCS[od_config.model_class_name]
     return _load_process_func(od_config, func_name)
 

vllm_omni/diffusion/request.py

@@ -66,6 +66,18 @@ class OmniDiffusionRequest:
     seed: int | None = None
     seeds: list[int] | None = None
 
+    # layered info
+    layers: int = 4
+
+    # cfg info
+    cfg_normalize: bool = False
+
+    # caption language
+    use_en_prompt: bool = False
+
+    # different bucket in (640, 1024) to determine the condition and output resolution
+    resolution: int = 640
+
     # Tracking if embeddings are already processed
     is_prompt_processed: bool = False
 

vllm_omni/diffusion/utils/hf_utils.py

@@ -1,6 +1,8 @@
+import os
 from functools import lru_cache
 
 from vllm.logger import init_logger
+from vllm.transformers_utils.config import get_hf_file_to_dict
 
 logger = init_logger(__name__)
 
@@ -14,8 +16,49 @@ def load_diffusers_config(model_name) -> dict:
 
 @lru_cache
 def is_diffusion_model(model_name: str) -> bool:
+    """Check if a model is a diffusion model.
+
+    Uses multiple fallback strategies to detect diffusion models:
+    1. Check local file system for model_index.json (fastest, no imports)
+    2. Check using vllm's get_hf_file_to_dict utility
+    3. Try the standard diffusers approach (may fail due to import issues)
+    """
+    # Strategy 1: Check local file system first (fastest, avoids import issues)
+    if os.path.isdir(model_name):
+        model_index_path = os.path.join(model_name, "model_index.json")
+        if os.path.exists(model_index_path):
+            try:
+                import json
+
+                with open(model_index_path) as f:
+                    config_dict = json.load(f)
+                if config_dict.get("_class_name") and config_dict.get("_diffusers_version"):
+                    logger.debug("Detected diffusion model via local model_index.json")
+                    return True
+            except Exception as e:
+                logger.debug("Failed to read local model_index.json: %s", e)
+
+    # Strategy 2: Check using vllm's utility (works for both local and remote models)
+    try:
+        # Try to get model_index.json using vllm's utility
+        config_dict = get_hf_file_to_dict("model_index.json", model_name)
+        # Verify it has the required fields for a diffusers model
+        if config_dict.get("_class_name") and config_dict.get("_diffusers_version"):
+            logger.debug("Detected diffusion model via model_index.json")
+            return True
+    except Exception as e:
+        logger.debug("Failed to check model_index.json via get_hf_file_to_dict: %s", e)
+
+    # Strategy 3: Try the standard diffusers approach (may fail due to import issues)
+    # This is last because it requires importing diffusers/xformers/flash_attn
+    # which may have compatibility issues
     try:
         load_diffusers_config(model_name)
         return True
-    except Exception:
-        return False
+    except (ImportError, ModuleNotFoundError) as e:
+        logger.debug("Failed to import diffusers dependencies: %s", e)
+        logger.debug("This may be due to flash_attn/PyTorch version mismatch")
+    except Exception as e:
+        logger.debug("Failed to load diffusers config via DiffusionPipeline: %s", e)
+
+    return False