Deploy Gradio app with multiple files

app.py

@@ -0,0 +1,227 @@
+import gradio as gr
+import numpy as np
+import random
+import torch
+import spaces
+
+from PIL import Image
+from diffusers import FlowMatchEulerDiscreteScheduler
+from optimization import optimize_pipeline_
+from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
+from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
+from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
+
+import math
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+
+from PIL import Image
+import os
+import gradio as gr
+from gradio_client import Client, handle_file
+import tempfile
+
+
+# --- Model Loading ---
+dtype = torch.bfloat16
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+pipe = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2509", 
+                                                transformer= QwenImageTransformer2DModel.from_pretrained("linoyts/Qwen-Image-Edit-Rapid-AIO", 
+                                                                                                         subfolder='transformer',
+                                                                                                         torch_dtype=dtype,
+                                                                                                         device_map='cuda'),torch_dtype=dtype).to(device)
+
+pipe.load_lora_weights("autoweeb/Qwen-Image-Edit-2509-Photo-to-Anime", adapter_name="anime")
+pipe.set_adapters(["anime"], adapter_weights=[1.])
+pipe.fuse_lora(adapter_names=["anime"], lora_scale=1.0)
+pipe.unload_lora_weights()
+
+
+
+pipe.transformer.__class__ = QwenImageTransformer2DModel
+pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
+
+optimize_pipeline_(pipe, image=[Image.new("RGB", (1024, 1024)), Image.new("RGB", (1024, 1024))], prompt="prompt")
+
+
+MAX_SEED = np.iinfo(np.int32).max
+
+def _generate_video_segment(input_image_path: str, output_image_path: str, prompt: str, request: gr.Request) -> str:
+    """Generates a single video segment using the external service."""
+    x_ip_token = request.headers['x-ip-token']
+    video_client = Client("multimodalart/wan-2-2-first-last-frame", headers={"x-ip-token": x_ip_token})
+    result = video_client.predict(
+        start_image_pil=handle_file(input_image_path),
+        end_image_pil=handle_file(output_image_path),
+        prompt=prompt, api_name="/generate_video",
+    )
+    return result[0]["video"]
+
+@spaces.GPU
+def convert_to_anime(
+    image,
+    seed,
+    randomize_seed,
+    true_guidance_scale,
+    num_inference_steps,
+    height,
+    width,
+    progress=gr.Progress(track_tqdm=True)
+):
+    prompt = "Convert this photo to anime style"
+    
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    generator = torch.Generator(device=device).manual_seed(seed)
+
+    pil_images = []
+    if image is not None:
+        if isinstance(image, Image.Image):
+            pil_images.append(image.convert("RGB"))
+        elif hasattr(image, "name"):
+            pil_images.append(Image.open(image.name).convert("RGB"))
+
+    if len(pil_images) == 0:
+        raise gr.Error("Please upload an image first.")
+
+    result = pipe(
+        image=pil_images,
+        prompt=prompt,
+        height=height if height != 0 else None,
+        width=width if width != 0 else None,
+        num_inference_steps=num_inference_steps,
+        generator=generator,
+        true_cfg_scale=true_guidance_scale,
+        num_images_per_prompt=1,
+    ).images[0]
+
+    return result, seed
+
+
+# --- UI ---
+css = '''
+#col-container { 
+    max-width: 900px; 
+    margin: 0 auto; 
+    padding: 2rem;
+    font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
+}
+.gradio-container {
+    background: linear-gradient(to bottom, #f5f5f7, #ffffff);
+}
+#title {
+    text-align: center;
+    font-size: 2.5rem;
+    font-weight: 600;
+    color: #1d1d1f;
+    margin-bottom: 0.5rem;
+}
+#description {
+    text-align: center;
+    font-size: 1.1rem;
+    color: #6e6e73;
+    margin-bottom: 2rem;
+}
+.image-container {
+    border-radius: 18px;
+    overflow: hidden;
+    box-shadow: 0 4px 6px rgba(0, 0, 0, 0.07);
+}
+#convert-btn {
+    background: linear-gradient(180deg, #0071e3 0%, #0077ed 100%);
+    border: none;
+    border-radius: 12px;
+    color: white;
+    font-size: 1.1rem;
+    font-weight: 500;
+    padding: 0.75rem 2rem;
+    transition: all 0.3s ease;
+}
+#convert-btn:hover {
+    transform: translateY(-2px);
+    box-shadow: 0 8px 16px rgba(0, 113, 227, 0.3);
+}
+'''
+
+def update_dimensions_on_upload(image):
+    if image is None:
+        return 1024, 1024
+    
+    original_width, original_height = image.size
+    
+    if original_width > original_height:
+        new_width = 1024
+        aspect_ratio = original_height / original_width
+        new_height = int(new_width * aspect_ratio)
+    else:
+        new_height = 1024
+        aspect_ratio = original_width / original_height
+        new_width = int(new_height * aspect_ratio)
+        
+    # Ensure dimensions are multiples of 8
+    new_width = (new_width // 8) * 8
+    new_height = (new_height // 8) * 8
+    
+    return new_width, new_height
+
+
+
+            ["tool_of_the_sea.png", 90, 0, 0, False, 0, True, 1.0, 4, 568, 1024],
+            ["monkey.jpg", -90, 0, 0, False, 0, True, 1.0, 4, 704, 1024],
+            ["metropolis.jpg", 0, 0, -1, False, 0, True, 1.0, 4, 816, 1024],
+            ["disaster_girl.jpg", -45, 0, 1, False, 0, True, 1.0, 4, 768, 1024],
+            ["grumpy.png", 90, 0, 1, False, 0, True, 1.0, 4, 576, 1024]
+        ],
+        inputs=[image,rotate_deg, move_forward,
+        vertical_tilt, wideangle,
+        seed, randomize_seed, true_guidance_scale, num_inference_steps, height, width],
+        outputs=outputs,
+        fn=infer_camera_edit,
+        cache_examples="lazy",
+        elem_id="examples"
+    )
+    
+    # Image upload triggers dimension update and control reset
+    image.upload(
+        fn=update_dimensions_on_upload,
+        inputs=[image],
+        outputs=[width, height]
+    ).then(
+        fn=reset_all,
+        inputs=None,
+        outputs=[rotate_deg, move_forward, vertical_tilt, wideangle, is_reset],
+        queue=False
+    ).then(
+        fn=end_reset, 
+        inputs=None, 
+        outputs=[is_reset], 
+        queue=False
+    )
+
+
+    # Live updates
+    def maybe_infer(is_reset, progress=gr.Progress(track_tqdm=True), *args):
+        if is_reset:
+            return gr.update(), gr.update(), gr.update(), gr.update()
+        else:
+            result_img, result_seed, result_prompt = infer_camera_edit(*args)
+            # Show video button if we have both input and output
+            show_button = args[0] is not None and result_img is not None
+            return result_img, result_seed, result_prompt, gr.update(visible=show_button)
+
+    control_inputs = [
+        image, rotate_deg, move_forward,
+        vertical_tilt, wideangle,
+        seed, randomize_seed, true_guidance_scale, num_inference_steps, height, width, prev_output
+    ]
+    control_inputs_with_flag = [is_reset] + control_inputs
+
+    for control in [rotate_deg, move_forward, vertical_tilt]:
+        control.release(fn=maybe_infer, inputs=control_inputs_with_flag, outputs=outputs + [create_video_button])
+    
+    wideangle.input(fn=maybe_infer, inputs=control_inputs_with_flag, outputs=outputs + [create_video_button])
+    
+    run_event.then(lambda img, *_: img, inputs=[result], outputs=[prev_output])
+
+demo.launch()

optimization.py

@@ -0,0 +1,70 @@
+"""
+"""
+
+from typing import Any
+from typing import Callable
+from typing import ParamSpec
+from torchao.quantization import quantize_
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
+import spaces
+import torch
+from torch.utils._pytree import tree_map
+
+
+P = ParamSpec('P')
+
+
+TRANSFORMER_IMAGE_SEQ_LENGTH_DIM = torch.export.Dim('image_seq_length')
+TRANSFORMER_TEXT_SEQ_LENGTH_DIM = torch.export.Dim('text_seq_length')
+
+TRANSFORMER_DYNAMIC_SHAPES = {
+    'hidden_states': {
+        1: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
+    },
+    'encoder_hidden_states': {
+        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
+    },
+    'encoder_hidden_states_mask': {
+        1: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
+    },
+    'image_rotary_emb': ({
+        0: TRANSFORMER_IMAGE_SEQ_LENGTH_DIM,
+    }, {
+        0: TRANSFORMER_TEXT_SEQ_LENGTH_DIM,
+    }),
+}
+
+
+INDUCTOR_CONFIGS = {
+    'conv_1x1_as_mm': True,
+    'epilogue_fusion': False,
+    'coordinate_descent_tuning': True,
+    'coordinate_descent_check_all_directions': True,
+    'max_autotune': True,
+    'triton.cudagraphs': True,
+}
+
+
+def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kwargs):
+
+    @spaces.GPU(duration=1500)
+    def compile_transformer():
+
+        with spaces.aoti_capture(pipeline.transformer) as call:
+            pipeline(*args, **kwargs)
+
+        dynamic_shapes = tree_map(lambda t: None, call.kwargs)
+        dynamic_shapes |= TRANSFORMER_DYNAMIC_SHAPES
+
+        # quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
+        
+        exported = torch.export.export(
+            mod=pipeline.transformer,
+            args=call.args,
+            kwargs=call.kwargs,
+            dynamic_shapes=dynamic_shapes,
+        )
+
+        return spaces.aoti_compile(exported, INDUCTOR_CONFIGS)
+
+    spaces.aoti_apply(compile_transformer(), pipeline.transformer)

qwenimage/pipeline_qwenimage_edit_plus.py

@@ -0,0 +1,889 @@
+# Copyright 2025 Qwen-Image 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
+import math
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import QwenImageLoraLoaderMixin
+from diffusers.models import AutoencoderKLQwenImage, QwenImageTransformer2DModel
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.qwenimage.pipeline_output import QwenImagePipelineOutput
+
+
+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
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        >>> import torch
+        >>> from PIL import Image
+        >>> from diffusers import QwenImageEditPlusPipeline
+        >>> from diffusers.utils import load_image
+
+        >>> pipe = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2509", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png"
+        ... ).convert("RGB")
+        >>> prompt = (
+        ...     "Make Pikachu hold a sign that says 'Qwen Edit is awesome', yarn art style, detailed, vibrant colors"
+        ... )
+        >>> # Depending on the variant being used, the pipeline call will slightly vary.
+        >>> # Refer to the pipeline documentation for more details.
+        >>> image = pipe(image, prompt, num_inference_steps=50).images[0]
+        >>> image.save("qwenimage_edit_plus.png")
+"""
+
+CONDITION_IMAGE_SIZE = 384 * 384
+VAE_IMAGE_SIZE = 1024 * 1024
+
+
+# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# 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,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    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 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
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = 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")
+
+
+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
+
+
+class QwenImageEditPlusPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
+    r"""
+    The Qwen-Image-Edit pipeline for image editing.
+
+    Args:
+        transformer ([`QwenImageTransformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`Qwen2.5-VL-7B-Instruct`]):
+            [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the
+            [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant.
+        tokenizer (`QwenTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLQwenImage,
+        text_encoder: Qwen2_5_VLForConditionalGeneration,
+        tokenizer: Qwen2Tokenizer,
+        processor: Qwen2VLProcessor,
+        transformer: QwenImageTransformer2DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            processor=processor,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        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.tokenizer_max_length = 1024
+
+        self.prompt_template_encode = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        self.prompt_template_encode_start_idx = 64
+        self.default_sample_size = 128
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden
+    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: Union[str, List[str]] = None,
+        image: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        img_prompt_template = "Picture {}: <|vision_start|><|image_pad|><|vision_end|>"
+        if isinstance(image, list):
+            base_img_prompt = ""
+            for i, img in enumerate(image):
+                base_img_prompt += img_prompt_template.format(i + 1)
+        elif image is not None:
+            base_img_prompt = img_prompt_template.format(1)
+        else:
+            base_img_prompt = ""
+
+        template = self.prompt_template_encode
+
+        drop_idx = self.prompt_template_encode_start_idx
+        txt = [template.format(base_img_prompt + e) for e in prompt]
+
+        model_inputs = self.processor(
+            text=txt,
+            images=image,
+            padding=True,
+            return_tensors="pt",
+        ).to(device)
+
+        outputs = self.text_encoder(
+            input_ids=model_inputs.input_ids,
+            attention_mask=model_inputs.attention_mask,
+            pixel_values=model_inputs.pixel_values,
+            image_grid_thw=model_inputs.image_grid_thw,
+            output_hidden_states=True,
+        )
+
+        hidden_states = outputs.hidden_states[-1]
+        split_hidden_states = self._extract_masked_hidden(hidden_states, model_inputs.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
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        image: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_embeds_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 1024,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            image (`torch.Tensor`, *optional*):
+                image 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._execution_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, image, device)
+
+        _, 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
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline.check_inputs
+    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,
+        callback_on_step_end_tensor_inputs=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} but are {height} and {width}. Dimensions will be resized accordingly"
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        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}")
+
+    @staticmethod
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents
+    def _pack_latents(latents, batch_size, num_channels_latents, height, width):
+        latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
+        latents = latents.permute(0, 2, 4, 1, 3, 5)
+        latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)
+
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents
+    def _unpack_latents(latents, height, width, 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, height // 2, width // 2, channels // 4, 2, 2)
+        latents = latents.permute(0, 3, 1, 4, 2, 5)
+
+        latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width)
+
+        return latents
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline._encode_vae_image
+    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,
+        images,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        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, 1, num_channels_latents, height, width)
+
+        image_latents = None
+        if images is not None:
+            if not isinstance(images, list):
+                images = [images]
+            all_image_latents = []
+            for image in images:
+                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 = self._pack_latents(
+                    image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width
+                )
+                all_image_latents.append(image_latents)
+            image_latents = torch.cat(all_image_latents, dim=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)
+        else:
+            latents = latents.to(device=device, dtype=dtype)
+
+        return latents, image_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
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: Optional[PipelineImageInput] = None,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        true_cfg_scale: float = 4.0,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: Optional[float] = None,
+        num_images_per_prompt: int = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_embeds_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+                `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
+                numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
+                or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
+                list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
+                latents as `image`, but if passing latents directly it is not encoded again.
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts 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 `true_cfg_scale` is
+                not greater than `1`).
+            true_cfg_scale (`float`, *optional*, defaults to 1.0):
+                true_cfg_scale (`float`, *optional*, defaults to 1.0): Guidance scale as defined in [Classifier-Free
+                Diffusion Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of
+                equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is
+                enabled by setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale
+                encourages to generate images that are closely linked to the text `prompt`, usually at the expense of
+                lower image quality.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to None):
+                A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance
+                where the guidance scale is applied during inference through noise prediction rescaling, guidance
+                distilled models take the guidance scale directly as an input parameter during forward pass. Guidance
+                scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images
+                that are closely linked to the text `prompt`, usually at the expense of lower image quality. This
+                parameter in the pipeline is there to support future guidance-distilled models when they come up. It is
+                ignored when not using guidance distilled models. To enable traditional classifier-free guidance,
+                please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should
+                enable classifier-free guidance computations).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will be generated by sampling using the supplied random `generator`.
+            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. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`:
+            [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is a list with the generated images.
+        """
+        image_size = image[-1].size if isinstance(image, list) else image.size
+        calculated_width, calculated_height = calculate_dimensions(1024 * 1024, image_size[0] / image_size[1])
+        height = height or calculated_height
+        width = width or calculated_width
+
+        multiple_of = self.vae_scale_factor * 2
+        width = width // multiple_of * multiple_of
+        height = height // multiple_of * multiple_of
+
+        # 1. Check inputs. Raise error if not correct
+        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,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        # 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. Preprocess image
+        if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels):
+            if not isinstance(image, list):
+                image = [image]
+            condition_image_sizes = []
+            condition_images = []
+            vae_image_sizes = []
+            vae_images = []
+            for img in image:
+                image_width, image_height = img.size
+                condition_width, condition_height = calculate_dimensions(
+                    CONDITION_IMAGE_SIZE, image_width / image_height
+                )
+                vae_width, vae_height = calculate_dimensions(VAE_IMAGE_SIZE, image_width / image_height)
+                condition_image_sizes.append((condition_width, condition_height))
+                vae_image_sizes.append((vae_width, vae_height))
+                condition_images.append(self.image_processor.resize(img, condition_height, condition_width))
+                vae_images.append(self.image_processor.preprocess(img, vae_height, vae_width).unsqueeze(2))
+
+        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 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(
+            image=condition_images,
+            prompt=prompt,
+            prompt_embeds=prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            device=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(
+                image=condition_images,
+                prompt=negative_prompt,
+                prompt_embeds=negative_prompt_embeds,
+                prompt_embeds_mask=negative_prompt_embeds_mask,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents, image_latents = self.prepare_latents(
+            vae_images,
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+        img_shapes = [
+            [
+                (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2),
+                *[
+                    (1, vae_height // self.vae_scale_factor // 2, vae_width // self.vae_scale_factor // 2)
+                    for vae_width, vae_height in vae_image_sizes
+                ],
+            ]
+        ] * batch_size
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
+        image_seq_len = latents.shape[1]
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.15),
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # handle guidance
+        if self.transformer.config.guidance_embeds and guidance_scale is None:
+            raise ValueError("guidance_scale is required for guidance-distilled model.")
+        elif self.transformer.config.guidance_embeds:
+            guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
+            guidance = guidance.expand(latents.shape[0])
+        elif not self.transformer.config.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.config.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
+        
+        image_rotary_emb = self.transformer.pos_embed(img_shapes, txt_seq_lens, device=latents.device)
+        if do_true_cfg:
+            negative_txt_seq_lens = (
+                negative_prompt_embeds_mask.sum(dim=1).tolist()
+                if negative_prompt_embeds_mask is not None
+                else None
+            )
+            uncond_image_rotary_emb = self.transformer.pos_embed(
+                img_shapes, negative_txt_seq_lens, device=latents.device
+            )
+        else:
+            uncond_image_rotary_emb = None
+
+        # 6. Denoising loop
+        self.scheduler.set_begin_index(0)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            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)
+                with self.transformer.cache_context("cond"):
+                    noise_pred = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep / 1000,
+                        guidance=guidance,
+                        encoder_hidden_states_mask=prompt_embeds_mask,
+                        encoder_hidden_states=prompt_embeds,
+                        image_rotary_emb=image_rotary_emb,
+                        attention_kwargs=self.attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_pred[:, : latents.size(1)]
+
+                if do_true_cfg:
+                    with self.transformer.cache_context("uncond"):
+                        neg_noise_pred = self.transformer(
+                            hidden_states=latent_model_input,
+                            timestep=timestep / 1000,
+                            guidance=guidance,
+                            encoder_hidden_states_mask=negative_prompt_embeds_mask,
+                            encoder_hidden_states=negative_prompt_embeds,
+                            image_rotary_emb=uncond_image_rotary_emb,
+                            attention_kwargs=self.attention_kwargs,
+                            return_dict=False,
+                        )[0]
+                    neg_noise_pred = neg_noise_pred[:, : latents.size(1)]
+                    comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
+
+                    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)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = self._unpack_latents(latents, height, width, 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
+            image = self.vae.decode(latents, return_dict=False)[0][:, :, 0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return QwenImagePipelineOutput(images=image)

qwenimage/qwen_fa3_processor.py

@@ -0,0 +1,142 @@
+"""
+Paired with a good language model. Thanks!
+"""
+
+import torch
+from typing import Optional, Tuple
+from diffusers.models.transformers.transformer_qwenimage import apply_rotary_emb_qwen
+
+try:
+    from kernels import get_kernel
+    _k = get_kernel("kernels-community/vllm-flash-attn3")
+    _flash_attn_func = _k.flash_attn_func
+except Exception as e:
+    _flash_attn_func = None
+    _kernels_err = e
+
+
+def _ensure_fa3_available():
+    if _flash_attn_func is None:
+        raise ImportError(
+            "FlashAttention-3 via Hugging Face `kernels` is required. "
+            "Tried `get_kernel('kernels-community/vllm-flash-attn3')` and failed with:\n"
+            f"{_kernels_err}"
+        )
+
+@torch.library.custom_op("flash::flash_attn_func", mutates_args=())
+def flash_attn_func(
+    q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, causal: bool = False
+) -> torch.Tensor:
+    outputs, lse = _flash_attn_func(q, k, v, causal=causal)
+    return outputs
+
+@flash_attn_func.register_fake
+def _(q, k, v, **kwargs):
+    # two outputs:
+    # 1. output: (batch, seq_len, num_heads, head_dim)
+    # 2. softmax_lse: (batch, num_heads, seq_len) with dtype=torch.float32
+    meta_q = torch.empty_like(q).contiguous()
+    return meta_q #, q.new_empty((q.size(0), q.size(2), q.size(1)), dtype=torch.float32)
+
+
+class QwenDoubleStreamAttnProcessorFA3:
+    """
+    FA3-based attention processor for Qwen double-stream architecture.
+    Computes joint attention over concatenated [text, image] streams using vLLM FlashAttention-3
+    accessed via Hugging Face `kernels`.
+
+    Notes / limitations:
+    - General attention masks are not supported here (FA3 path). `is_causal=False` and no arbitrary mask.
+    - Optional windowed attention / sink tokens / softcap can be plumbed through if you use those features.
+    - Expects an available `apply_rotary_emb_qwen` in scope (same as your non-FA3 processor).
+    """
+
+    _attention_backend = "fa3"  # for parity with your other processors, not used internally
+
+    def __init__(self):
+        _ensure_fa3_available()
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        attn,  # Attention module with to_q/to_k/to_v/add_*_proj, norms, to_out, to_add_out, and .heads
+        hidden_states: torch.FloatTensor,                 # (B, S_img, D_model)  image stream
+        encoder_hidden_states: torch.FloatTensor = None,  # (B, S_txt, D_model)  text stream
+        encoder_hidden_states_mask: torch.FloatTensor = None,  # unused in FA3 path
+        attention_mask: Optional[torch.FloatTensor] = None,    # unused in FA3 path
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # (img_freqs, txt_freqs)
+    ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+        if encoder_hidden_states is None:
+            raise ValueError("QwenDoubleStreamAttnProcessorFA3 requires encoder_hidden_states (text stream).")
+        if attention_mask is not None:
+            # FA3 kernel path here does not consume arbitrary masks; fail fast to avoid silent correctness issues.
+            raise NotImplementedError("attention_mask is not supported in this FA3 implementation.")
+
+        _ensure_fa3_available()
+
+        B, S_img, _ = hidden_states.shape
+        S_txt = encoder_hidden_states.shape[1]
+
+        # ---- QKV projections (image/sample stream) ----
+        img_q = attn.to_q(hidden_states)   # (B, S_img, D)
+        img_k = attn.to_k(hidden_states)
+        img_v = attn.to_v(hidden_states)
+
+        # ---- QKV projections (text/context stream) ----
+        txt_q = attn.add_q_proj(encoder_hidden_states)  # (B, S_txt, D)
+        txt_k = attn.add_k_proj(encoder_hidden_states)
+        txt_v = attn.add_v_proj(encoder_hidden_states)
+
+        # ---- Reshape to (B, S, H, D_h) ----
+        H = attn.heads
+        img_q = img_q.unflatten(-1, (H, -1))
+        img_k = img_k.unflatten(-1, (H, -1))
+        img_v = img_v.unflatten(-1, (H, -1))
+
+        txt_q = txt_q.unflatten(-1, (H, -1))
+        txt_k = txt_k.unflatten(-1, (H, -1))
+        txt_v = txt_v.unflatten(-1, (H, -1))
+
+        # ---- Q/K normalization (per your module contract) ----
+        if getattr(attn, "norm_q", None) is not None:
+            img_q = attn.norm_q(img_q)
+        if getattr(attn, "norm_k", None) is not None:
+            img_k = attn.norm_k(img_k)
+        if getattr(attn, "norm_added_q", None) is not None:
+            txt_q = attn.norm_added_q(txt_q)
+        if getattr(attn, "norm_added_k", None) is not None:
+            txt_k = attn.norm_added_k(txt_k)
+
+        # ---- RoPE (Qwen variant) ----
+        if image_rotary_emb is not None:
+            img_freqs, txt_freqs = image_rotary_emb
+            # expects tensors shaped (B, S, H, D_h)
+            img_q = apply_rotary_emb_qwen(img_q, img_freqs, use_real=False)
+            img_k = apply_rotary_emb_qwen(img_k, img_freqs, use_real=False)
+            txt_q = apply_rotary_emb_qwen(txt_q, txt_freqs, use_real=False)
+            txt_k = apply_rotary_emb_qwen(txt_k, txt_freqs, use_real=False)
+
+        # ---- Joint attention over [text, image] along sequence axis ----
+        # Shapes: (B, S_total, H, D_h)
+        q = torch.cat([txt_q, img_q], dim=1)
+        k = torch.cat([txt_k, img_k], dim=1)
+        v = torch.cat([txt_v, img_v], dim=1)
+
+        # FlashAttention-3 path expects (B, S, H, D_h) and returns (out, softmax_lse)
+        out = flash_attn_func(q, k, v, causal=False)  # out: (B, S_total, H, D_h)
+
+        # ---- Back to (B, S, D_model) ----
+        out = out.flatten(2, 3).to(q.dtype)
+
+        # Split back to text / image segments
+        txt_attn_out = out[:, :S_txt, :]
+        img_attn_out = out[:, S_txt:, :]
+
+        # ---- Output projections ----
+        img_attn_out = attn.to_out[0](img_attn_out)
+        if len(attn.to_out) > 1:
+            img_attn_out = attn.to_out[1](img_attn_out)  # dropout if present
+
+        txt_attn_out = attn.to_add_out(txt_attn_out)
+
+        return img_attn_out, txt_attn_out

qwenimage/transformer_qwenimage.py

@@ -0,0 +1,642 @@
+# Copyright 2025 Qwen-Image Team, 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 functools
+import math
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
+from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.attention import FeedForward, AttentionMixin
+from diffusers.models.attention_dispatch import dispatch_attention_fn
+from diffusers.models.attention_processor import Attention
+from diffusers.models.cache_utils import CacheMixin
+from diffusers.models.embeddings import TimestepEmbedding, Timesteps
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormContinuous, RMSNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_timestep_embedding(
+    timesteps: torch.Tensor,
+    embedding_dim: int,
+    flip_sin_to_cos: bool = False,
+    downscale_freq_shift: float = 1,
+    scale: float = 1,
+    max_period: int = 10000,
+) -> torch.Tensor:
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
+
+    Args
+        timesteps (torch.Tensor):
+            a 1-D Tensor of N indices, one per batch element. These may be fractional.
+        embedding_dim (int):
+            the dimension of the output.
+        flip_sin_to_cos (bool):
+            Whether the embedding order should be `cos, sin` (if True) or `sin, cos` (if False)
+        downscale_freq_shift (float):
+            Controls the delta between frequencies between dimensions
+        scale (float):
+            Scaling factor applied to the embeddings.
+        max_period (int):
+            Controls the maximum frequency of the embeddings
+    Returns
+        torch.Tensor: an [N x dim] Tensor of positional embeddings.
+    """
+    assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
+
+    half_dim = embedding_dim // 2
+    exponent = -math.log(max_period) * torch.arange(
+        start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
+    )
+    exponent = exponent / (half_dim - downscale_freq_shift)
+
+    emb = torch.exp(exponent).to(timesteps.dtype)
+    emb = timesteps[:, None].float() * emb[None, :]
+
+    # scale embeddings
+    emb = scale * emb
+
+    # concat sine and cosine embeddings
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
+
+    # flip sine and cosine embeddings
+    if flip_sin_to_cos:
+        emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
+
+    # zero pad
+    if embedding_dim % 2 == 1:
+        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
+    return emb
+
+
+def apply_rotary_emb_qwen(
+    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, S, H, 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, CogView4 and Cosmos
+            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:
+        x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
+        freqs_cis = freqs_cis.unsqueeze(1)
+        x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3)
+
+        return x_out.type_as(x)
+
+
+class QwenTimestepProjEmbeddings(nn.Module):
+    def __init__(self, embedding_dim):
+        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)
+
+    def forward(self, timestep, hidden_states):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))  # (N, D)
+
+        conditioning = timesteps_emb
+
+        return conditioning
+
+
+class QwenEmbedRope(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.rope_cache = {}
+
+        # DO NOT USING REGISTER BUFFER HERE, IT WILL CAUSE COMPLEX NUMBERS LOSE ITS IMAGINARY PART
+        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
+        for idx, fhw in enumerate(video_fhw):
+            frame, height, width = fhw
+            rope_key = f"{idx}_{height}_{width}"
+
+            if not torch.compiler.is_compiling():
+                if rope_key not in self.rope_cache:
+                    self.rope_cache[rope_key] = self._compute_video_freqs(frame, height, width, idx)
+                video_freq = self.rope_cache[rope_key]
+            else:
+                video_freq = self._compute_video_freqs(frame, height, width, idx)
+            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_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.lru_cache(maxsize=None)
+    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()
+
+
+class QwenDoubleStreamAttnProcessor2_0:
+    """
+    Attention processor for Qwen double-stream architecture, matching DoubleStreamLayerMegatron logic. This processor
+    implements joint attention computation where text and image streams are processed together.
+    """
+
+    _attention_backend = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "QwenDoubleStreamAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,  # Image stream
+        encoder_hidden_states: torch.FloatTensor = None,  # Text stream
+        encoder_hidden_states_mask: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        if encoder_hidden_states is None:
+            raise ValueError("QwenDoubleStreamAttnProcessor2_0 requires encoder_hidden_states (text stream)")
+
+        seq_txt = encoder_hidden_states.shape[1]
+
+        # Compute QKV for image stream (sample projections)
+        img_query = attn.to_q(hidden_states)
+        img_key = attn.to_k(hidden_states)
+        img_value = attn.to_v(hidden_states)
+
+        # Compute QKV for text stream (context projections)
+        txt_query = attn.add_q_proj(encoder_hidden_states)
+        txt_key = attn.add_k_proj(encoder_hidden_states)
+        txt_value = attn.add_v_proj(encoder_hidden_states)
+
+        # Reshape for multi-head attention
+        img_query = img_query.unflatten(-1, (attn.heads, -1))
+        img_key = img_key.unflatten(-1, (attn.heads, -1))
+        img_value = img_value.unflatten(-1, (attn.heads, -1))
+
+        txt_query = txt_query.unflatten(-1, (attn.heads, -1))
+        txt_key = txt_key.unflatten(-1, (attn.heads, -1))
+        txt_value = txt_value.unflatten(-1, (attn.heads, -1))
+
+        # Apply QK normalization
+        if attn.norm_q is not None:
+            img_query = attn.norm_q(img_query)
+        if attn.norm_k is not None:
+            img_key = attn.norm_k(img_key)
+        if attn.norm_added_q is not None:
+            txt_query = attn.norm_added_q(txt_query)
+        if attn.norm_added_k is not None:
+            txt_key = attn.norm_added_k(txt_key)
+
+        # Apply RoPE
+        if image_rotary_emb is not None:
+            img_freqs, txt_freqs = image_rotary_emb
+            img_query = apply_rotary_emb_qwen(img_query, img_freqs, use_real=False)
+            img_key = apply_rotary_emb_qwen(img_key, img_freqs, use_real=False)
+            txt_query = apply_rotary_emb_qwen(txt_query, txt_freqs, use_real=False)
+            txt_key = apply_rotary_emb_qwen(txt_key, txt_freqs, use_real=False)
+
+        # Concatenate for joint attention
+        # Order: [text, image]
+        joint_query = torch.cat([txt_query, img_query], dim=1)
+        joint_key = torch.cat([txt_key, img_key], dim=1)
+        joint_value = torch.cat([txt_value, img_value], dim=1)
+
+        # Compute joint attention
+        joint_hidden_states = dispatch_attention_fn(
+            joint_query,
+            joint_key,
+            joint_value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+        )
+
+        # Reshape back
+        joint_hidden_states = joint_hidden_states.flatten(2, 3)
+        joint_hidden_states = joint_hidden_states.to(joint_query.dtype)
+
+        # Split attention outputs back
+        txt_attn_output = joint_hidden_states[:, :seq_txt, :]  # Text part
+        img_attn_output = joint_hidden_states[:, seq_txt:, :]  # Image part
+
+        # Apply output projections
+        img_attn_output = attn.to_out[0](img_attn_output)
+        if len(attn.to_out) > 1:
+            img_attn_output = attn.to_out[1](img_attn_output)  # dropout
+
+        txt_attn_output = attn.to_add_out(txt_attn_output)
+
+        return img_attn_output, txt_attn_output
+
+
+@maybe_allow_in_graph
+class QwenImageTransformerBlock(nn.Module):
+    def __init__(
+        self, dim: int, num_attention_heads: int, attention_head_dim: int, qk_norm: str = "rms_norm", eps: float = 1e-6
+    ):
+        super().__init__()
+
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+
+        # Image processing modules
+        self.img_mod = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(dim, 6 * dim, bias=True),  # For scale, shift, gate for norm1 and norm2
+        )
+        self.img_norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,  # Enable cross attention for joint computation
+            added_kv_proj_dim=dim,  # Enable added KV projections for text stream
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            context_pre_only=False,
+            bias=True,
+            processor=QwenDoubleStreamAttnProcessor2_0(),
+            qk_norm=qk_norm,
+            eps=eps,
+        )
+        self.img_norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.img_mlp = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+        # Text processing modules
+        self.txt_mod = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(dim, 6 * dim, bias=True),  # For scale, shift, gate for norm1 and norm2
+        )
+        self.txt_norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        # Text doesn't need separate attention - it's handled by img_attn joint computation
+        self.txt_norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=eps)
+        self.txt_mlp = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+    def _modulate(self, x, mod_params):
+        """Apply modulation to input tensor"""
+        shift, scale, gate = mod_params.chunk(3, dim=-1)
+        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1), gate.unsqueeze(1)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_mask: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Get modulation parameters for both streams
+        img_mod_params = self.img_mod(temb)  # [B, 6*dim]
+        txt_mod_params = self.txt_mod(temb)  # [B, 6*dim]
+
+        # Split modulation parameters for norm1 and norm2
+        img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)  # Each [B, 3*dim]
+        txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)  # Each [B, 3*dim]
+
+        # Process image stream - norm1 + modulation
+        img_normed = self.img_norm1(hidden_states)
+        img_modulated, img_gate1 = self._modulate(img_normed, img_mod1)
+
+        # Process text stream - norm1 + modulation
+        txt_normed = self.txt_norm1(encoder_hidden_states)
+        txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1)
+
+        # Use QwenAttnProcessor2_0 for joint attention computation
+        # This directly implements the DoubleStreamLayerMegatron logic:
+        # 1. Computes QKV for both streams
+        # 2. Applies QK normalization and RoPE
+        # 3. Concatenates and runs joint attention
+        # 4. Splits results back to separate streams
+        joint_attention_kwargs = joint_attention_kwargs or {}
+        attn_output = self.attn(
+            hidden_states=img_modulated,  # Image stream (will be processed as "sample")
+            encoder_hidden_states=txt_modulated,  # Text stream (will be processed as "context")
+            encoder_hidden_states_mask=encoder_hidden_states_mask,
+            image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
+        )
+
+        # QwenAttnProcessor2_0 returns (img_output, txt_output) when encoder_hidden_states is provided
+        img_attn_output, txt_attn_output = attn_output
+
+        # Apply attention gates and add residual (like in Megatron)
+        hidden_states = hidden_states + img_gate1 * img_attn_output
+        encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
+
+        # Process image stream - norm2 + MLP
+        img_normed2 = self.img_norm2(hidden_states)
+        img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
+        img_mlp_output = self.img_mlp(img_modulated2)
+        hidden_states = hidden_states + img_gate2 * img_mlp_output
+
+        # Process text stream - norm2 + MLP
+        txt_normed2 = self.txt_norm2(encoder_hidden_states)
+        txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
+        txt_mlp_output = self.txt_mlp(txt_modulated2)
+        encoder_hidden_states = encoder_hidden_states + txt_gate2 * txt_mlp_output
+
+        # Clip to prevent overflow for fp16
+        if encoder_hidden_states.dtype == torch.float16:
+            encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+        if hidden_states.dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+
+        return encoder_hidden_states, hidden_states
+
+
+class QwenImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin, AttentionMixin):
+    """
+    The Transformer model introduced in Qwen.
+
+    Args:
+        patch_size (`int`, defaults to `2`):
+            Patch size to turn the input data into small patches.
+        in_channels (`int`, defaults to `64`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `None`):
+            The number of channels in the output. If not specified, it defaults to `in_channels`.
+        num_layers (`int`, defaults to `60`):
+            The number of layers of dual stream DiT blocks to use.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of dimensions to use for each attention head.
+        num_attention_heads (`int`, defaults to `24`):
+            The number of attention heads to use.
+        joint_attention_dim (`int`, defaults to `3584`):
+            The number of dimensions to use for the joint attention (embedding/channel dimension of
+            `encoder_hidden_states`).
+        guidance_embeds (`bool`, defaults to `False`):
+            Whether to use guidance embeddings for guidance-distilled variant of the model.
+        axes_dims_rope (`Tuple[int]`, defaults to `(16, 56, 56)`):
+            The dimensions to use for the rotary positional embeddings.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["QwenImageTransformerBlock"]
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
+    _repeated_blocks = ["QwenImageTransformerBlock"]
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 64,
+        out_channels: Optional[int] = 16,
+        num_layers: int = 60,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        joint_attention_dim: int = 3584,
+        guidance_embeds: bool = False,  # TODO: this should probably be removed
+        axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
+    ):
+        super().__init__()
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pos_embed = QwenEmbedRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
+
+        self.time_text_embed = QwenTimestepProjEmbeddings(embedding_dim=self.inner_dim)
+
+        self.txt_norm = RMSNorm(joint_attention_dim, eps=1e-6)
+
+        self.img_in = nn.Linear(in_channels, self.inner_dim)
+        self.txt_in = nn.Linear(joint_attention_dim, self.inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                QwenImageTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor = None,
+        encoder_hidden_states_mask: torch.Tensor = None,
+        timestep: torch.LongTensor = None,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        guidance: torch.Tensor = None,  # TODO: this should probably be removed
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[torch.Tensor, Transformer2DModelOutput]:
+        """
+        The [`QwenTransformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`):
+                Input `hidden_states`.
+            encoder_hidden_states (`torch.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            encoder_hidden_states_mask (`torch.Tensor` of shape `(batch_size, text_sequence_length)`):
+                Mask of the input conditions.
+            timestep ( `torch.LongTensor`):
+                Used to indicate denoising step.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        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 `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        hidden_states = self.img_in(hidden_states)
+
+        timestep = timestep.to(hidden_states.dtype)
+        encoder_hidden_states = self.txt_norm(encoder_hidden_states)
+        encoder_hidden_states = self.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance.to(hidden_states.dtype) * 1000
+
+        temb = (
+            self.time_text_embed(timestep, hidden_states)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, hidden_states)
+        )
+
+        for index_block, block in enumerate(self.transformer_blocks):
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    encoder_hidden_states_mask,
+                    temb,
+                    image_rotary_emb,
+                )
+
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=attention_kwargs,
+                )
+
+        # Use only the image part (hidden_states) from the dual-stream blocks
+        hidden_states = self.norm_out(hidden_states, temb)
+        output = self.proj_out(hidden_states)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

requirements.txt

@@ -0,0 +1,10 @@
+git+https://github.com/huggingface/diffusers.git
+transformers
+accelerate
+safetensors
+sentencepiece
+dashscope
+kernels
+torchvision
+peft
+torchao==0.11.0