[diffusion] z-image support (#149)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: Samit <285365963@qq.com>

docs/models/supported_models.md

@@ -20,6 +20,7 @@ th {
 | `Qwen3OmniMoeForConditionalGeneration` | Qwen3-Omni | `Qwen/Qwen3-Omni-30B-A3B-Instruct` |
 | `Qwen2_5OmniForConditionalGeneration` | Qwen2.5-Omni | `Qwen/Qwen2.5-Omni-7B`, `Qwen/Qwen2.5-Omni-3B` |
 | `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` |
+|`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` |
 
 
 ## List of Supported Models for Ascend NPU

vllm_omni/diffusion/data.py

@@ -21,6 +21,8 @@ class OmniDiffusionConfig:
 
     model_class_name: str | None = None
 
+    dtype: torch.dtype = torch.bfloat16
+
     # Attention
     # attention_backend: str = None
 

vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image.py

@@ -20,7 +20,6 @@ from diffusers.schedulers.scheduling_flow_match_euler_discrete import (
 from diffusers.utils.torch_utils import randn_tensor
 from torch import nn
 from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer
-from vllm.model_executor.model_loader.utils import set_default_torch_dtype
 
 from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
 from vllm_omni.diffusion.distributed.utils import get_local_device
@@ -236,7 +235,7 @@ class QwenImagePipeline(
     def __init__(
         self,
         *,
-        od_config: OmniDiffusionConfig = None,
+        od_config: OmniDiffusionConfig,
         prefix: str = "",
     ):
         super().__init__()
@@ -250,23 +249,14 @@ class QwenImagePipeline(
         self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
             model, subfolder="scheduler", local_files_only=local_files_only
         )
-        logger.info("Loaded Qwen-Image scheduler successfully")
-
-        with set_default_torch_dtype(torch.bfloat16):
-            self.text_encoder = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-                model, subfolder="text_encoder", local_files_only=local_files_only
-            )
-            logger.info("Loaded Qwen-Image text encoder successfully")
-            self.vae = AutoencoderKLQwenImage.from_pretrained(
-                model, subfolder="vae", local_files_only=local_files_only
-            ).to(self.device)
-            logger.info("Loaded Qwen-Image VAE successfully")
-            self.transformer = QwenImageTransformer2DModel()
-            logger.info("Initialized Qwen-Image transformer successfully.")
-            self.tokenizer = Qwen2Tokenizer.from_pretrained(
-                model, subfolder="tokenizer", local_files_only=local_files_only
-            )
-            logger.info("Loaded Qwen-Image tokenizer successfully.")
+        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.transformer = QwenImageTransformer2DModel()
+        self.tokenizer = Qwen2Tokenizer.from_pretrained(model, subfolder="tokenizer", local_files_only=local_files_only)
 
         self.stage = None
 

vllm_omni/diffusion/models/z_image/pipeline_z_image.py

@@ -0,0 +1,648 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright 2025 Alibaba Z-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 json
+import os
+from typing import Any, Callable
+
+import torch
+import torch.nn as nn
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.models.autoencoders import AutoencoderKL
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import logging
+from diffusers.utils.torch_utils import randn_tensor
+from transformers import AutoModel, AutoTokenizer
+
+from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
+from vllm_omni.diffusion.distributed.utils import get_local_device
+from vllm_omni.diffusion.models.z_image.z_image_transformer import (
+    ZImageTransformer2DModel,
+)
+from vllm_omni.diffusion.request import OmniDiffusionRequest
+from vllm_omni.model_executor.model_loader.weight_utils import (
+    download_weights_from_hf_specific,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_post_process_func(
+    od_config: OmniDiffusionConfig,
+):
+    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["block_out_channels"]) - 1) if "block_out_channels" in vae_config else 8
+
+    image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor * 2)
+
+    def post_process_func(
+        images: torch.Tensor,
+    ):
+        return image_processor.postprocess(images)
+
+    return post_process_func
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.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: int | None = None,
+    device: str | torch.device | None = None,
+    timesteps: list[int] | None = None,
+    sigmas: list[float] | None = None,
+    **kwargs,
+) -> tuple[torch.Tensor, int]:
+    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
+
+
+class ZImagePipeline(nn.Module):
+    def __init__(
+        self,
+        *,
+        od_config: OmniDiffusionConfig = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.od_config = od_config
+        self._execution_device = get_local_device()
+        model = od_config.model
+        local_files_only = os.path.exists(model)
+        self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
+            model, subfolder="scheduler", local_files_only=local_files_only
+        )
+
+        self.text_encoder = AutoModel.from_pretrained(
+            model, subfolder="text_encoder", local_files_only=local_files_only
+        )
+        self.vae = AutoencoderKL.from_pretrained(model, subfolder="vae", local_files_only=local_files_only).to(
+            self._execution_device
+        )
+        self.transformer = ZImageTransformer2DModel()
+        self.tokenizer = AutoTokenizer.from_pretrained(model, subfolder="tokenizer", local_files_only=local_files_only)
+
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
+        )
+
+    def encode_prompt(
+        self,
+        prompt: str | list[str],
+        device: torch.device | None = None,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: str | list[str] | None = None,
+        prompt_embeds: list[torch.FloatTensor] | None = None,
+        negative_prompt_embeds: torch.FloatTensor | None = None,
+        max_sequence_length: int = 512,
+    ):
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt_embeds = self._encode_prompt(
+            prompt=prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            max_sequence_length=max_sequence_length,
+        )
+
+        if do_classifier_free_guidance:
+            if negative_prompt is None:
+                negative_prompt = ["" for _ in prompt]
+            else:
+                negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            assert len(prompt) == len(negative_prompt)
+            negative_prompt_embeds = self._encode_prompt(
+                prompt=negative_prompt,
+                device=device,
+                prompt_embeds=negative_prompt_embeds,
+                max_sequence_length=max_sequence_length,
+            )
+        else:
+            negative_prompt_embeds = []
+        return prompt_embeds, negative_prompt_embeds
+
+    def _encode_prompt(
+        self,
+        prompt: str | list[str],
+        device: torch.device | None = None,
+        prompt_embeds: list[torch.FloatTensor] | None = None,
+        max_sequence_length: int = 512,
+    ) -> list[torch.FloatTensor]:
+        device = device or self._execution_device
+
+        if prompt_embeds is not None:
+            return prompt_embeds
+
+        if isinstance(prompt, str):
+            prompt = [prompt]
+
+        for i, prompt_item in enumerate(prompt):
+            messages = [
+                {"role": "user", "content": prompt_item},
+            ]
+            prompt_item = self.tokenizer.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True,
+                enable_thinking=True,
+            )
+            prompt[i] = prompt_item
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids.to(device)
+        prompt_masks = text_inputs.attention_mask.to(device).bool()
+
+        prompt_embeds = self.text_encoder(
+            input_ids=text_input_ids,
+            attention_mask=prompt_masks,
+            output_hidden_states=True,
+        ).hidden_states[-2]
+
+        embeddings_list = []
+
+        for i in range(len(prompt_embeds)):
+            embeddings_list.append(prompt_embeds[i][prompt_masks[i]])
+
+        return embeddings_list
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
+
+        shape = (batch_size, num_channels_latents, height, width)
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            if latents.shape != shape:
+                raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+            latents = latents.to(device)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1
+
+    @property
+    def joint_attention_kwargs(self):
+        return self._joint_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    def forward(
+        self,
+        req: OmniDiffusionRequest,
+        prompt: str | list[str] | None = None,
+        height: int | None = None,
+        width: int | None = None,
+        num_inference_steps: int = 50,
+        sigmas: list[float] | None = None,
+        guidance_scale: float = 5.0,
+        cfg_normalization: bool = False,
+        cfg_truncation: float = 1.0,
+        negative_prompt: str | list[str] | None = None,
+        num_images_per_prompt: int | None = 1,
+        generator: torch.Generator | list[torch.Generator] | None = None,
+        latents: torch.FloatTensor | None = None,
+        prompt_embeds: list[torch.FloatTensor] | None = None,
+        negative_prompt_embeds: list[torch.FloatTensor] | None = None,
+        output_type: str | None = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: dict[str, Any] | None = None,
+        callback_on_step_end: Callable[[int, int, dict], None] | None = None,
+        callback_on_step_end_tensor_inputs: list[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ) -> DiffusionOutput:
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            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.
+            height (`int`, *optional*, defaults to 1024):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to 1024):
+                The width in pixels of the generated image.
+            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 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). 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.
+            cfg_normalization (`bool`, *optional*, defaults to False):
+                Whether to apply configuration normalization.
+            cfg_truncation (`float`, *optional*, defaults to 1.0):
+                The truncation value for configuration.
+            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 `guidance_scale` is
+                less than `1`).
+            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.FloatTensor`, *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 (`list[torch.FloatTensor]`, *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 (`list[torch.FloatTensor]`, *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.stable_diffusion.ZImagePipelineOutput`] instead of a plain
+                tuple.
+            joint_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`, *optional*, defaults to 512):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if
+            `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
+            generated images.
+        """
+        prompt = req.prompt
+        negative_prompt = req.negative_prompt
+        height: int = req.height or 1024
+        width: int = req.width or 1024
+        num_inference_steps = req.num_inference_steps or 50
+        generator = req.generator
+        guidance_scale = req.guidance_scale if req.guidance_rescale is not None else guidance_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
+
+        vae_scale = self.vae_scale_factor * 2
+        if height % vae_scale != 0:
+            raise ValueError(
+                f"Height must be divisible by {vae_scale} (got {height}). "
+                f"Please adjust the height to a multiple of {vae_scale}."
+            )
+        if width % vae_scale != 0:
+            raise ValueError(
+                f"Width must be divisible by {vae_scale} (got {width}). "
+                f"Please adjust the width to a multiple of {vae_scale}."
+            )
+
+        device = self._execution_device
+
+        self._guidance_scale = guidance_scale
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+        self._cfg_normalization = cfg_normalization
+        self._cfg_truncation = cfg_truncation
+        # 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 = len(prompt_embeds)
+
+        # If prompt_embeds is provided and prompt is None, skip encoding
+        if prompt_embeds is not None and prompt is None:
+            if self.do_classifier_free_guidance and negative_prompt_embeds is None:
+                raise ValueError(
+                    "When `prompt_embeds` is provided without `prompt`, "
+                    "`negative_prompt_embeds` must also be provided for classifier-free guidance."
+                )
+        else:
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+            ) = self.encode_prompt(
+                prompt=prompt,
+                negative_prompt=negative_prompt,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+                prompt_embeds=prompt_embeds,
+                negative_prompt_embeds=negative_prompt_embeds,
+                device=device,
+                max_sequence_length=max_sequence_length,
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.in_channels
+
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # Repeat prompt_embeds for num_images_per_prompt
+        if num_images_per_prompt > 1:
+            prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)]
+            if self.do_classifier_free_guidance and negative_prompt_embeds:
+                negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)]
+
+        actual_batch_size = batch_size * num_images_per_prompt
+        image_seq_len = (latents.shape[2] // 2) * (latents.shape[3] // 2)
+
+        # 5. Prepare timesteps
+        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),
+        )
+        self.scheduler.sigma_min = 0.0
+        scheduler_kwargs = {"mu": mu}
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            **scheduler_kwargs,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Denoising loop
+        for i, t in enumerate(timesteps):
+            if self.interrupt:
+                continue
+
+            # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+            timestep = t.expand(latents.shape[0])
+            timestep = (1000 - timestep) / 1000
+            # Normalized time for time-aware config (0 at start, 1 at end)
+            t_norm = timestep[0].item()
+
+            # Handle cfg truncation
+            current_guidance_scale = self.guidance_scale
+            if (
+                self.do_classifier_free_guidance
+                and self._cfg_truncation is not None
+                and float(self._cfg_truncation) <= 1
+            ):
+                if t_norm > self._cfg_truncation:
+                    current_guidance_scale = 0.0
+
+            # Run CFG only if configured AND scale is non-zero
+            apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0
+
+            if apply_cfg:
+                latents_typed = latents.to(self.transformer.dtype)
+                latent_model_input = latents_typed.repeat(2, 1, 1, 1)
+                prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds
+                timestep_model_input = timestep.repeat(2)
+            else:
+                latent_model_input = latents.to(self.transformer.dtype)
+                prompt_embeds_model_input = prompt_embeds
+                timestep_model_input = timestep
+
+            latent_model_input = latent_model_input.unsqueeze(2)
+            latent_model_input_list = list(latent_model_input.unbind(dim=0))
+
+            model_out_list = self.transformer(
+                latent_model_input_list,
+                timestep_model_input,
+                prompt_embeds_model_input,
+            )[0]
+
+            if apply_cfg:
+                # Perform CFG
+                pos_out = model_out_list[:actual_batch_size]
+                neg_out = model_out_list[actual_batch_size:]
+
+                noise_pred = []
+                for j in range(actual_batch_size):
+                    pos = pos_out[j].float()
+                    neg = neg_out[j].float()
+
+                    pred = pos + current_guidance_scale * (pos - neg)
+
+                    # Renormalization
+                    if self._cfg_normalization and float(self._cfg_normalization) > 0.0:
+                        ori_pos_norm = torch.linalg.vector_norm(pos)
+                        new_pos_norm = torch.linalg.vector_norm(pred)
+                        max_new_norm = ori_pos_norm * float(self._cfg_normalization)
+                        if new_pos_norm > max_new_norm:
+                            pred = pred * (max_new_norm / new_pos_norm)
+
+                    noise_pred.append(pred)
+
+                noise_pred = torch.stack(noise_pred, dim=0)
+            else:
+                noise_pred = torch.stack([t.float() for t in model_out_list], dim=0)
+
+            noise_pred = noise_pred.squeeze(2)
+            noise_pred = -noise_pred
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0]
+            assert latents.dtype == torch.float32
+
+            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)
+                negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = latents.to(self.vae.dtype)
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+            # image = self.image_processor.postprocess(image, output_type=output_type)
+
+        return DiffusionOutput(output=image)
+
+    def load_weights(self):
+        self.load_transformer()
+
+    def load_transformer(self):
+        import glob
+        import os
+
+        # Define the weight iterator
+        def weight_iterator(transformer_path):
+            if not os.path.exists(transformer_path):
+                logger.warning(f"Path {transformer_path} does not exist.")
+                return
+
+            # Look for safetensors first
+            safetensors_files = glob.glob(os.path.join(transformer_path, "*.safetensors"))
+            if safetensors_files:
+                try:
+                    from safetensors.torch import load_file
+                except ImportError:
+                    logger.warning("safetensors not installed, cannot load .safetensors files.")
+                    return
+
+                for file_path in safetensors_files:
+                    state_dict = load_file(file_path)
+                    for name, tensor in state_dict.items():
+                        yield name, tensor
+            else:
+                # Fallback to bin
+                bin_files = glob.glob(os.path.join(transformer_path, "*.bin"))
+                for file_path in bin_files:
+                    state_dict = torch.load(file_path)
+                    for name, tensor in state_dict.items():
+                        yield name, tensor
+
+        try:
+            # Get model path from config or download from HF
+            model_name = self.od_config.model
+            if os.path.exists(model_name):
+                model_path = model_name
+            else:
+                model_path = download_weights_from_hf_specific(model_name, None, ["*"])
+
+            transformer_path = os.path.join(model_path, "transformer")
+            self.transformer.load_weights(weight_iterator(transformer_path))
+
+        except Exception as e:
+            print(f"An error occurred: {e}")
+            raise e

vllm_omni/diffusion/models/z_image/z_image_transformer.py

@@ -0,0 +1,668 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright 2025 Alibaba Z-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 math
+from collections.abc import Iterable
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.models.attention_dispatch import dispatch_attention_fn
+from torch.nn.utils.rnn import pad_sequence
+from vllm.model_executor.layers.layernorm import RMSNorm
+
+ADALN_EMBED_DIM = 256
+SEQ_MULTI_OF = 32
+
+
+class TimestepEmbedder(nn.Module):
+    def __init__(self, out_size, mid_size=None, frequency_embedding_size=256):
+        super().__init__()
+        if mid_size is None:
+            mid_size = out_size
+        self.mlp = nn.Sequential(
+            nn.Linear(
+                frequency_embedding_size,
+                mid_size,
+                bias=True,
+            ),
+            nn.SiLU(),
+            nn.Linear(
+                mid_size,
+                out_size,
+                bias=True,
+            ),
+        )
+
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=t.device) / half
+        )
+        args = t[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, t):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        weight_dtype = self.mlp[0].weight.dtype
+        if weight_dtype.is_floating_point:
+            t_freq = t_freq.to(weight_dtype)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+class ZImageAttention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        num_kv_heads: int,
+        qk_norm: bool = True,
+        eps: float = 1e-6,
+    ) -> None:
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.head_dim = dim // num_heads
+        self.qk_norm = qk_norm
+
+        self.to_q = nn.Linear(dim, dim, bias=False)
+        self.to_k = nn.Linear(dim, self.head_dim * num_kv_heads, bias=False)
+        self.to_v = nn.Linear(dim, self.head_dim * num_kv_heads, bias=False)
+        # TODO enable this when we refactor weight loader
+        # self.to_qkv = QKVParallelLinear(
+        #     hidden_size=dim,
+        #     head_size=self.head_dim,
+        #     total_num_heads=num_heads,
+        #     disable_tp=True,
+        # )
+
+        assert qk_norm is True
+        self.norm_q = RMSNorm(self.head_dim, eps=eps)
+        self.norm_k = RMSNorm(self.head_dim, eps=eps)
+
+        self.to_out = nn.ModuleList([nn.Linear(dim, dim, bias=False)])
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        freqs_cis: tuple[torch.Tensor, torch.Tensor] | None = None,
+    ):
+        query = self.to_q(hidden_states)
+        key = self.to_k(hidden_states)
+        value = self.to_v(hidden_states)
+        # qkv, _ = self.to_qkv(hidden_states)
+        # query, key, value = qkv.chunk(3, dim=-1)
+
+        query = query.unflatten(-1, (self.num_heads, -1))
+        key = key.unflatten(-1, (self.num_heads, -1))
+        value = value.unflatten(-1, (self.num_heads, -1))
+
+        query = self.norm_q(query)
+        key = self.norm_k(key)
+
+        def apply_rotary_emb(x_in: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
+            x = torch.view_as_complex(x_in.float().reshape(*x_in.shape[:-1], -1, 2))
+            freqs_cis = freqs_cis.unsqueeze(2)
+            x_out = torch.view_as_real(x * freqs_cis).flatten(3)
+            return x_out.type_as(x_in)
+
+        if freqs_cis is not None:
+            query = apply_rotary_emb(query, freqs_cis)
+            key = apply_rotary_emb(key, freqs_cis)
+
+        # Cast to correct dtype
+        dtype = query.dtype
+        query, key = query.to(dtype), key.to(dtype)
+
+        # From [batch, seq_len] to [batch, 1, 1, seq_len] -> broadcast to [batch, heads, seq_len, seq_len]
+        if attention_mask is not None and attention_mask.ndim == 2:
+            attention_mask = attention_mask[:, None, None, :]
+
+        # Compute joint attention
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+        )
+
+        # Reshape back
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(dtype)
+
+        hidden_states = self.to_out[0](hidden_states)
+
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim: int, hidden_dim: int):
+        super().__init__()
+        self.w1 = nn.Linear(dim, hidden_dim, bias=False)
+        self.w2 = nn.Linear(hidden_dim, dim, bias=False)
+        self.w3 = nn.Linear(dim, hidden_dim, bias=False)
+
+    def _forward_silu_gating(self, x1, x3):
+        return F.silu(x1) * x3
+
+    def forward(self, x):
+        return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
+
+
+class ZImageTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        norm_eps: float,
+        qk_norm: bool,
+        modulation=True,
+    ):
+        super().__init__()
+        self.dim = dim
+
+        self.attention = ZImageAttention(
+            dim=dim,
+            num_heads=n_heads,
+            num_kv_heads=n_kv_heads,
+            qk_norm=qk_norm,
+            eps=1e-5,
+        )
+
+        self.feed_forward = FeedForward(dim=dim, hidden_dim=int(dim / 3 * 8))
+        self.layer_id = layer_id
+
+        self.attention_norm1 = RMSNorm(dim, eps=norm_eps)
+        self.ffn_norm1 = RMSNorm(dim, eps=norm_eps)
+
+        self.attention_norm2 = RMSNorm(dim, eps=norm_eps)
+        self.ffn_norm2 = RMSNorm(dim, eps=norm_eps)
+
+        self.modulation = modulation
+        if modulation:
+            self.adaLN_modulation = nn.Sequential(
+                nn.Linear(min(dim, ADALN_EMBED_DIM), 4 * dim, bias=True),
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attn_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        adaln_input: Optional[torch.Tensor] = None,
+    ):
+        if self.modulation:
+            assert adaln_input is not None
+            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).unsqueeze(1).chunk(4, dim=2)
+            gate_msa, gate_mlp = gate_msa.tanh(), gate_mlp.tanh()
+            scale_msa, scale_mlp = 1.0 + scale_msa, 1.0 + scale_mlp
+
+            # Attention block
+            attn_out = self.attention(
+                self.attention_norm1(x) * scale_msa,
+                attention_mask=attn_mask,
+                freqs_cis=freqs_cis,
+            )
+            x = x + gate_msa * self.attention_norm2(attn_out)
+
+            # FFN block
+            x = x + gate_mlp * self.ffn_norm2(
+                self.feed_forward(
+                    self.ffn_norm1(x) * scale_mlp,
+                )
+            )
+        else:
+            # Attention block
+            attn_out = self.attention(
+                self.attention_norm1(x),
+                attention_mask=attn_mask,
+                freqs_cis=freqs_cis,
+            )
+            x = x + self.attention_norm2(attn_out)
+
+            # FFN block
+            x = x + self.ffn_norm2(
+                self.feed_forward(
+                    self.ffn_norm1(x),
+                )
+            )
+
+        return x
+
+
+class FinalLayer(nn.Module):
+    def __init__(self, hidden_size, out_channels):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, out_channels, bias=True)
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(min(hidden_size, ADALN_EMBED_DIM), hidden_size, bias=True),
+        )
+
+    def forward(self, x, c):
+        scale = 1.0 + self.adaLN_modulation(c)
+        x = self.norm_final(x) * scale.unsqueeze(1)
+        x = self.linear(x)
+        return x
+
+
+class RopeEmbedder:
+    def __init__(
+        self,
+        theta: float = 256.0,
+        axes_dims: list[int] = (16, 56, 56),
+        axes_lens: list[int] = (64, 128, 128),
+    ):
+        self.theta = theta
+        self.axes_dims = axes_dims
+        self.axes_lens = axes_lens
+        assert len(axes_dims) == len(axes_lens), "axes_dims and axes_lens must have the same length"
+        self.freqs_cis = None
+
+    @staticmethod
+    def precompute_freqs_cis(dim: list[int], end: list[int], theta: float = 256.0):
+        with torch.device("cpu"):
+            freqs_cis = []
+            for i, (d, e) in enumerate(zip(dim, end)):
+                freqs = 1.0 / (theta ** (torch.arange(0, d, 2, dtype=torch.float64, device="cpu") / d))
+                timestep = torch.arange(e, device=freqs.device, dtype=torch.float64)
+                freqs = torch.outer(timestep, freqs).float()
+                freqs_cis_i = torch.polar(torch.ones_like(freqs), freqs).to(torch.complex64)  # complex64
+                freqs_cis.append(freqs_cis_i)
+
+            return freqs_cis
+
+    def __call__(self, ids: torch.Tensor):
+        assert ids.ndim == 2
+        assert ids.shape[-1] == len(self.axes_dims)
+        device = ids.device
+
+        if self.freqs_cis is None:
+            self.freqs_cis = self.precompute_freqs_cis(self.axes_dims, self.axes_lens, theta=self.theta)
+            self.freqs_cis = [freqs_cis.to(device) for freqs_cis in self.freqs_cis]
+        else:
+            # Ensure freqs_cis are on the same device as ids
+            if self.freqs_cis[0].device != device:
+                self.freqs_cis = [freqs_cis.to(device) for freqs_cis in self.freqs_cis]
+
+        result = []
+        for i in range(len(self.axes_dims)):
+            index = ids[:, i]
+            result.append(self.freqs_cis[i][index])
+        return torch.cat(result, dim=-1)
+
+
+class ZImageTransformer2DModel(nn.Module):
+    def __init__(
+        self,
+        all_patch_size=(2,),
+        all_f_patch_size=(1,),
+        in_channels=16,
+        dim=3840,
+        n_layers=30,
+        n_refiner_layers=2,
+        n_heads=30,
+        n_kv_heads=30,
+        norm_eps=1e-5,
+        qk_norm=True,
+        cap_feat_dim=2560,
+        rope_theta=256.0,
+        t_scale=1000.0,
+        axes_dims=[32, 48, 48],
+        axes_lens=[1024, 512, 512],
+    ) -> None:
+        super().__init__()
+        self.dtype = torch.bfloat16
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.all_patch_size = all_patch_size
+        self.all_f_patch_size = all_f_patch_size
+        self.dim = dim
+        self.n_heads = n_heads
+
+        self.rope_theta = rope_theta
+        self.t_scale = t_scale
+        self.gradient_checkpointing = False
+
+        assert len(all_patch_size) == len(all_f_patch_size)
+
+        all_x_embedder = {}
+        all_final_layer = {}
+        for patch_idx, (patch_size, f_patch_size) in enumerate(zip(all_patch_size, all_f_patch_size)):
+            x_embedder = nn.Linear(f_patch_size * patch_size * patch_size * in_channels, dim, bias=True)
+            all_x_embedder[f"{patch_size}-{f_patch_size}"] = x_embedder
+
+            final_layer = FinalLayer(dim, patch_size * patch_size * f_patch_size * self.out_channels)
+            all_final_layer[f"{patch_size}-{f_patch_size}"] = final_layer
+
+        self.all_x_embedder = nn.ModuleDict(all_x_embedder)
+        self.all_final_layer = nn.ModuleDict(all_final_layer)
+        self.noise_refiner = nn.ModuleList(
+            [
+                ZImageTransformerBlock(
+                    1000 + layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    norm_eps,
+                    qk_norm,
+                    modulation=True,
+                )
+                for layer_id in range(n_refiner_layers)
+            ]
+        )
+        self.context_refiner = nn.ModuleList(
+            [
+                ZImageTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    norm_eps,
+                    qk_norm,
+                    modulation=False,
+                )
+                for layer_id in range(n_refiner_layers)
+            ]
+        )
+        self.t_embedder = TimestepEmbedder(min(dim, ADALN_EMBED_DIM), mid_size=1024)
+        self.cap_embedder = nn.Sequential(
+            RMSNorm(cap_feat_dim, eps=norm_eps),
+            nn.Linear(cap_feat_dim, dim, bias=True),
+        )
+
+        self.x_pad_token = nn.Parameter(torch.empty((1, dim)))
+        self.cap_pad_token = nn.Parameter(torch.empty((1, dim)))
+
+        self.layers = nn.ModuleList(
+            [
+                ZImageTransformerBlock(layer_id, dim, n_heads, n_kv_heads, norm_eps, qk_norm)
+                for layer_id in range(n_layers)
+            ]
+        )
+        self.axes_dims = axes_dims
+        self.axes_lens = axes_lens
+
+        self.rope_embedder = RopeEmbedder(theta=rope_theta, axes_dims=axes_dims, axes_lens=axes_lens)
+
+    def unpatchify(self, x: list[torch.Tensor], size: list[tuple], patch_size, f_patch_size) -> list[torch.Tensor]:
+        pH = pW = patch_size
+        pF = f_patch_size
+        bsz = len(x)
+        assert len(size) == bsz
+        for i in range(bsz):
+            F, H, W = size[i]
+            ori_len = (F // pF) * (H // pH) * (W // pW)
+            # "f h w pf ph pw c -> c (f pf) (h ph) (w pw)"
+            x[i] = (
+                x[i][:ori_len]
+                .view(F // pF, H // pH, W // pW, pF, pH, pW, self.out_channels)
+                .permute(6, 0, 3, 1, 4, 2, 5)
+                .reshape(self.out_channels, F, H, W)
+            )
+        return x
+
+    @staticmethod
+    def create_coordinate_grid(size, start=None, device=None):
+        if start is None:
+            start = (0 for _ in size)
+
+        axes = [torch.arange(x0, x0 + span, dtype=torch.int32, device=device) for x0, span in zip(start, size)]
+        grids = torch.meshgrid(axes, indexing="ij")
+        return torch.stack(grids, dim=-1)
+
+    def patchify_and_embed(
+        self,
+        all_image: list[torch.Tensor],
+        all_cap_feats: list[torch.Tensor],
+        patch_size: int,
+        f_patch_size: int,
+    ):
+        pH = pW = patch_size
+        pF = f_patch_size
+        device = all_image[0].device
+
+        all_image_out = []
+        all_image_size = []
+        all_image_pos_ids = []
+        all_image_pad_mask = []
+        all_cap_pos_ids = []
+        all_cap_pad_mask = []
+        all_cap_feats_out = []
+
+        for i, (image, cap_feat) in enumerate(zip(all_image, all_cap_feats)):
+            ### Process Caption
+            cap_ori_len = len(cap_feat)
+            cap_padding_len = (-cap_ori_len) % SEQ_MULTI_OF
+            # padded position ids
+            cap_padded_pos_ids = self.create_coordinate_grid(
+                size=(cap_ori_len + cap_padding_len, 1, 1),
+                start=(1, 0, 0),
+                device=device,
+            ).flatten(0, 2)
+            all_cap_pos_ids.append(cap_padded_pos_ids)
+            # pad mask
+            all_cap_pad_mask.append(
+                torch.cat(
+                    [
+                        torch.zeros((cap_ori_len,), dtype=torch.bool, device=device),
+                        torch.ones((cap_padding_len,), dtype=torch.bool, device=device),
+                    ],
+                    dim=0,
+                )
+            )
+            # padded feature
+            cap_padded_feat = torch.cat(
+                [cap_feat, cap_feat[-1:].repeat(cap_padding_len, 1)],
+                dim=0,
+            )
+            all_cap_feats_out.append(cap_padded_feat)
+
+            ### Process Image
+            C, F, H, W = image.size()
+            all_image_size.append((F, H, W))
+            F_tokens, H_tokens, W_tokens = F // pF, H // pH, W // pW
+
+            image = image.view(C, F_tokens, pF, H_tokens, pH, W_tokens, pW)
+            # "c f pf h ph w pw -> (f h w) (pf ph pw c)"
+            image = image.permute(1, 3, 5, 2, 4, 6, 0).reshape(F_tokens * H_tokens * W_tokens, pF * pH * pW * C)
+
+            image_ori_len = len(image)
+            image_padding_len = (-image_ori_len) % SEQ_MULTI_OF
+
+            image_ori_pos_ids = self.create_coordinate_grid(
+                size=(F_tokens, H_tokens, W_tokens),
+                start=(cap_ori_len + cap_padding_len + 1, 0, 0),
+                device=device,
+            ).flatten(0, 2)
+            image_padding_pos_ids = (
+                self.create_coordinate_grid(
+                    size=(1, 1, 1),
+                    start=(0, 0, 0),
+                    device=device,
+                )
+                .flatten(0, 2)
+                .repeat(image_padding_len, 1)
+            )
+            image_padded_pos_ids = torch.cat([image_ori_pos_ids, image_padding_pos_ids], dim=0)
+            all_image_pos_ids.append(image_padded_pos_ids)
+            # pad mask
+            all_image_pad_mask.append(
+                torch.cat(
+                    [
+                        torch.zeros((image_ori_len,), dtype=torch.bool, device=device),
+                        torch.ones((image_padding_len,), dtype=torch.bool, device=device),
+                    ],
+                    dim=0,
+                )
+            )
+            # padded feature
+            image_padded_feat = torch.cat([image, image[-1:].repeat(image_padding_len, 1)], dim=0)
+            all_image_out.append(image_padded_feat)
+
+        return (
+            all_image_out,
+            all_cap_feats_out,
+            all_image_size,
+            all_image_pos_ids,
+            all_cap_pos_ids,
+            all_image_pad_mask,
+            all_cap_pad_mask,
+        )
+
+    def forward(
+        self,
+        x: list[torch.Tensor],
+        t,
+        cap_feats: list[torch.Tensor],
+        patch_size=2,
+        f_patch_size=1,
+    ):
+        assert patch_size in self.all_patch_size
+        assert f_patch_size in self.all_f_patch_size
+
+        bsz = len(x)
+        device = x[0].device
+        t = t * self.t_scale
+        t = self.t_embedder(t)
+
+        (
+            x,
+            cap_feats,
+            x_size,
+            x_pos_ids,
+            cap_pos_ids,
+            x_inner_pad_mask,
+            cap_inner_pad_mask,
+        ) = self.patchify_and_embed(x, cap_feats, patch_size, f_patch_size)
+
+        # x embed & refine
+        x_item_seqlens = [len(_) for _ in x]
+        assert all(_ % SEQ_MULTI_OF == 0 for _ in x_item_seqlens)
+        x_max_item_seqlen = max(x_item_seqlens)
+
+        x = torch.cat(x, dim=0)
+        x = self.all_x_embedder[f"{patch_size}-{f_patch_size}"](x)
+
+        # Match t_embedder output dtype to x for layerwise casting compatibility
+        adaln_input = t.type_as(x)
+        x[torch.cat(x_inner_pad_mask)] = self.x_pad_token
+        x = list(x.split(x_item_seqlens, dim=0))
+        x_freqs_cis = list(self.rope_embedder(torch.cat(x_pos_ids, dim=0)).split(x_item_seqlens, dim=0))
+
+        x = pad_sequence(x, batch_first=True, padding_value=0.0)
+        x_freqs_cis = pad_sequence(x_freqs_cis, batch_first=True, padding_value=0.0)
+        x_attn_mask = torch.zeros((bsz, x_max_item_seqlen), dtype=torch.bool, device=device)
+        for i, seq_len in enumerate(x_item_seqlens):
+            x_attn_mask[i, :seq_len] = 1
+
+        for layer in self.noise_refiner:
+            x = layer(x, x_attn_mask, x_freqs_cis, adaln_input)
+
+        # cap embed & refine
+        cap_item_seqlens = [len(_) for _ in cap_feats]
+        assert all(_ % SEQ_MULTI_OF == 0 for _ in cap_item_seqlens)
+        cap_max_item_seqlen = max(cap_item_seqlens)
+
+        cap_feats = torch.cat(cap_feats, dim=0)
+        cap_feats = self.cap_embedder(cap_feats)
+        cap_feats[torch.cat(cap_inner_pad_mask)] = self.cap_pad_token
+        cap_feats = list(cap_feats.split(cap_item_seqlens, dim=0))
+        cap_freqs_cis = list(self.rope_embedder(torch.cat(cap_pos_ids, dim=0)).split(cap_item_seqlens, dim=0))
+
+        cap_feats = pad_sequence(cap_feats, batch_first=True, padding_value=0.0)
+        cap_freqs_cis = pad_sequence(cap_freqs_cis, batch_first=True, padding_value=0.0)
+        cap_attn_mask = torch.zeros((bsz, cap_max_item_seqlen), dtype=torch.bool, device=device)
+        for i, seq_len in enumerate(cap_item_seqlens):
+            cap_attn_mask[i, :seq_len] = 1
+
+        for layer in self.context_refiner:
+            cap_feats = layer(cap_feats, cap_attn_mask, cap_freqs_cis)
+
+        # unified
+        unified = []
+        unified_freqs_cis = []
+        for i in range(bsz):
+            x_len = x_item_seqlens[i]
+            cap_len = cap_item_seqlens[i]
+            unified.append(torch.cat([x[i][:x_len], cap_feats[i][:cap_len]]))
+            unified_freqs_cis.append(torch.cat([x_freqs_cis[i][:x_len], cap_freqs_cis[i][:cap_len]]))
+        unified_item_seqlens = [a + b for a, b in zip(cap_item_seqlens, x_item_seqlens)]
+        assert unified_item_seqlens == [len(_) for _ in unified]
+        unified_max_item_seqlen = max(unified_item_seqlens)
+
+        unified = pad_sequence(unified, batch_first=True, padding_value=0.0)
+        unified_freqs_cis = pad_sequence(unified_freqs_cis, batch_first=True, padding_value=0.0)
+        unified_attn_mask = torch.zeros((bsz, unified_max_item_seqlen), dtype=torch.bool, device=device)
+        for i, seq_len in enumerate(unified_item_seqlens):
+            unified_attn_mask[i, :seq_len] = 1
+
+        for layer in self.layers:
+            unified = layer(unified, unified_attn_mask, unified_freqs_cis, adaln_input)
+
+        unified = self.all_final_layer[f"{patch_size}-{f_patch_size}"](unified, adaln_input)
+        unified = list(unified.unbind(dim=0))
+        x = self.unpatchify(unified, x_size, patch_size, f_patch_size)
+
+        return x, {}
+
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
+        weights_map: dict[str, torch.Tensor] = {name: tensor for name, tensor in weights}
+        params_dict = dict(self.named_parameters())
+
+        for name, loaded_weight in weights_map.items():
+            # In ZImageAttention, to_q, to_k, to_v are separate.
+            # The weight names are like '...attention.to_q.weight'.
+            # The logic from qwen_image_transformer for stacking is not needed here.
+            # We can do a direct mapping.
+
+            # The names in the state dict seem to match the model structure.
+            # Let's try a direct load.
+            if name in params_dict:
+                param = params_dict[name]
+                with torch.no_grad():
+                    param.copy_(loaded_weight)
+            else:
+                # The provided weight list might use a different naming convention.
+                # For example, `layers.0.attention.to_out.0.weight` vs `...`
+                # Let's try to handle simple cases.
+
+                # Replace dots with hyphens for ModuleDict keys
+                # e.g. all_x_embedder.2-1.weight -> all_x_embedder.2-1.weight
+                # The names seem to match already.
+
+                # If there are any specific mismatches, they can be handled here.
+                # For now, we assume a direct mapping and warn if a weight is not found.
+                import logging
+
+                logging.warning(f"Weight {name} not found in model parameters.")

vllm_omni/diffusion/registry.py

@@ -12,6 +12,11 @@ _DIFFUSION_MODELS = {
         "pipeline_qwen_image",
         "QwenImagePipeline",
     ),
+    "ZImagePipeline": (
+        "z_image",
+        "pipeline_z_image",
+        "ZImagePipeline",
+    ),
 }
 
 
@@ -41,6 +46,7 @@ _DIFFUSION_POST_PROCESS_FUNCS = {
     # `post_process_func` function must be placed in {mod_folder}/{mod_relname}.py,
     # where mod_folder and mod_relname are  defined and mapped using `_DIFFUSION_MODELS` via the `arch` key
     "QwenImagePipeline": "get_qwen_image_post_process_func",
+    "ZImagePipeline": "get_post_process_func",
 }
 
 

vllm_omni/diffusion/worker/gpu_worker.py

@@ -12,6 +12,7 @@ from vllm.distributed.parallel_state import (
     initialize_model_parallel,
 )
 from vllm.logger import init_logger
+from vllm.model_executor.model_loader.utils import set_default_torch_dtype
 
 from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
 from vllm_omni.diffusion.registry import initialize_model
@@ -61,9 +62,10 @@ class GPUWorker:
         initialize_model_parallel(tensor_model_parallel_size=world_size)
 
         with device:
-            self.pipeline = initialize_model(self.od_config)
-            self.pipeline.load_weights()
-            self.pipeline.eval()
+            with set_default_torch_dtype(self.od_config.dtype):
+                self.pipeline = initialize_model(self.od_config)
+                self.pipeline.load_weights()
+                self.pipeline.eval()
         logger.info(f"Worker {self.rank}: Initialized device, model, and distributed environment.")
         logger.info(f"Worker {self.rank}: Model loaded successfully.")
 

vllm_omni/diffusion/worker/npu/npu_worker.py

@@ -13,6 +13,7 @@ from vllm.distributed.parallel_state import (
     initialize_model_parallel,
 )
 from vllm.logger import init_logger
+from vllm.model_executor.model_loader.utils import set_default_torch_dtype
 
 from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
 from vllm_omni.diffusion.registry import initialize_model
@@ -67,9 +68,10 @@ class NPUWorker:
         initialize_model_parallel(tensor_model_parallel_size=world_size)
 
         with device:
-            self.pipeline = initialize_model(self.od_config)
-            self.pipeline.load_weights()
-            self.pipeline.eval()
+            with set_default_torch_dtype(self.od_config.dtype):
+                self.pipeline = initialize_model(self.od_config)
+                self.pipeline.load_weights()
+                self.pipeline.eval()
         logger.info(f"Worker {self.rank}: Initialized device, model, and distributed environment.")
         logger.info(f"Worker {self.rank}: Model loaded successfully.")