diffusion: support Qwen-Image-Layered (#15817)

Co-authored-by: yhyang201 <yhyang201@gmail.com>

python/sglang/multimodal_gen/configs/models/vaes/qwenimage.py

@@ -20,7 +20,7 @@ class QwenImageVAEArchConfig(VAEArchConfig):
     dropout: float = 0.0
 
     is_residual: bool = False
-    in_channels: int = 3
+    input_channels: int = 3
     out_channels: int = 3
     patch_size: int | None = None
     scale_factor_temporal: int = 4

python/sglang/multimodal_gen/configs/pipeline_configs/base.py

@@ -299,6 +299,9 @@ class PipelineConfig:
 
         return shape
 
+    def allow_set_num_frames(self):
+        return False
+
     def get_decode_scale_and_shift(self, device, dtype, vae):
         vae_arch_config = self.vae_config.arch_config
         scaling_factor = getattr(vae_arch_config, "scaling_factor", None)

python/sglang/multimodal_gen/configs/pipeline_configs/qwen_image.py

@@ -12,6 +12,7 @@ from sglang.multimodal_gen.configs.models.vaes.qwenimage import QwenImageVAEConf
 from sglang.multimodal_gen.configs.pipeline_configs.base import (
     ImagePipelineConfig,
     ModelTaskType,
+    maybe_unpad_latents,
     shard_rotary_emb_for_sp,
 )
 from sglang.multimodal_gen.runtime.models.vision_utils import resize
@@ -466,3 +467,90 @@ class QwenImageEditPlusPipelineConfig(QwenImageEditPipelineConfig):
             "freqs_cis": ((img_cos, img_sin), (txt_cos, txt_sin)),
             "img_shapes": img_shapes,
         }
+
+
+@dataclass
+class QwenImageLayeredPipelineConfig(QwenImageEditPipelineConfig):
+    resolution: int = 640  # TODO: allow user to set resolution
+    vae_precision: str = "bf16"
+
+    def _prepare_edit_cond_kwargs(
+        self, batch, prompt_embeds, rotary_emb, device, dtype
+    ):
+        batch_size = batch.latents.shape[0]
+        assert batch_size == 1
+        height = batch.height
+        width = batch.width
+        image_size = batch.original_condition_image_size
+
+        vae_scale_factor = self.get_vae_scale_factor()
+
+        img_shapes = batch.img_shapes
+        txt_seq_lens = batch.txt_seq_lens
+
+        (img_cos, img_sin), (txt_cos, txt_sin) = (
+            QwenImageEditPlusPipelineConfig.get_freqs_cis(
+                img_shapes, txt_seq_lens, rotary_emb, device, dtype
+            )
+        )
+
+        # perform sp shard on noisy image tokens
+        noisy_img_seq_len = (
+            1 * (height // vae_scale_factor // 2) * (width // vae_scale_factor // 2)
+        )
+        noisy_img_cos = shard_rotary_emb_for_sp(img_cos[:noisy_img_seq_len, :])
+        noisy_img_sin = shard_rotary_emb_for_sp(img_sin[:noisy_img_seq_len, :])
+
+        # concat back the img_cos for input image (since it is not sp-shared later)
+        img_cos = torch.cat([noisy_img_cos, img_cos[noisy_img_seq_len:, :]], dim=0).to(
+            device=device
+        )
+        img_sin = torch.cat([noisy_img_sin, img_sin[noisy_img_seq_len:, :]], dim=0).to(
+            device=device
+        )
+
+        return {
+            "txt_seq_lens": txt_seq_lens,
+            "img_shapes": img_shapes,
+            "freqs_cis": ((img_cos, img_sin), (txt_cos, txt_sin)),
+            "additional_t_cond": torch.tensor([0], device=device, dtype=torch.long),
+        }
+
+    def _unpad_and_unpack_latents(self, latents, batch):
+        vae_scale_factor = self.vae_config.arch_config.vae_scale_factor
+        channels = self.dit_config.arch_config.in_channels
+        batch_size = latents.shape[0]
+        layers = batch.num_frames
+
+        height = 2 * (int(batch.height) // (vae_scale_factor * 2))
+        width = 2 * (int(batch.width) // (vae_scale_factor * 2))
+
+        latents = maybe_unpad_latents(latents, batch)
+        latents = latents.view(
+            batch_size, layers + 1, height // 2, width // 2, channels // 4, 2, 2
+        )
+        latents = latents.permute(0, 1, 4, 2, 5, 3, 6)
+
+        latents = latents.reshape(
+            batch_size, layers + 1, channels // (2 * 2), height, width
+        )
+        latents = latents.permute(0, 2, 1, 3, 4)  # (b, c, f, h, w)
+        return latents, batch_size, channels, height, width
+
+    def allow_set_num_frames(self):
+        return True
+
+    def post_denoising_loop(self, latents, batch):
+        # unpack latents for qwen-image
+        (
+            latents,
+            batch_size,
+            channels,
+            height,
+            width,
+        ) = self._unpad_and_unpack_latents(latents, batch)
+        b, c, f, h, w = latents.shape
+        latents = latents[:, :, 1:]  # remove the first frame as it is the origin input
+        latents = latents.permute(0, 2, 1, 3, 4).view(-1, c, 1, h, w)
+        # latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width)
+        return latents

python/sglang/multimodal_gen/configs/sample/qwenimage.py

@@ -21,3 +21,17 @@ class QwenImageEditPlusSamplingParams(QwenImageSamplingParams):
     guidance_scale: float = 4.0
     # true_cfg_scale: float = 4.0
     num_inference_steps: int = 40
+
+
+@dataclass
+class QwenImageLayeredSamplingParams(QwenImageSamplingParams):
+    # num_frames: int = 4
+    height: int = 640
+    width: int = 640
+    prompt: str = " "
+    negative_prompt: str = " "
+
+    guidance_scale: float = 4.0
+    num_inference_steps: int = 50
+    cfg_normalize: bool = True
+    use_en_prompt: bool = True

python/sglang/multimodal_gen/configs/sample/sampling_params.py

@@ -69,7 +69,7 @@ class DataType(Enum):
 
     def get_default_extension(self) -> str:
         if self == DataType.IMAGE:
-            return "jpg"
+            return "png"
         else:
             return "mp4"
 
@@ -279,8 +279,10 @@ class SamplingParams:
 
         if pipeline_config.task_type.is_image_gen():
             # settle num_frames
+            if not server_args.pipeline_config.allow_set_num_frames():
                 logger.debug(f"num_frames set to 1 for image generation model")
                 self.num_frames = 1
+
         elif self.adjust_frames:
             # NOTE: We must apply adjust_num_frames BEFORE the SP alignment logic below.
             # If we apply it after, adjust_num_frames might modify the frame count

python/sglang/multimodal_gen/registry.py

@@ -30,6 +30,7 @@ from sglang.multimodal_gen.configs.pipeline_configs.flux import Flux2PipelineCon
 from sglang.multimodal_gen.configs.pipeline_configs.qwen_image import (
     QwenImageEditPipelineConfig,
     QwenImageEditPlusPipelineConfig,
+    QwenImageLayeredPipelineConfig,
     QwenImagePipelineConfig,
 )
 from sglang.multimodal_gen.configs.pipeline_configs.wan import (
@@ -46,6 +47,7 @@ from sglang.multimodal_gen.configs.sample.hunyuan import (
 )
 from sglang.multimodal_gen.configs.sample.qwenimage import (
     QwenImageEditPlusSamplingParams,
+    QwenImageLayeredSamplingParams,
     QwenImageSamplingParams,
 )
 from sglang.multimodal_gen.configs.sample.stepvideo import StepVideoT2VSamplingParams
@@ -443,5 +445,11 @@ def _register_configs():
         hf_model_paths=["Qwen/Qwen-Image-Edit-2511"],
     )
 
+    register_configs(
+        sampling_param_cls=QwenImageLayeredSamplingParams,
+        pipeline_config_cls=QwenImageLayeredPipelineConfig,
+        hf_model_paths=["Qwen/Qwen-Image-Layered"],
+    )
+
 
 _register_configs()

python/sglang/multimodal_gen/runtime/entrypoints/diffusion_generator.py

@@ -183,14 +183,13 @@ class DiffGenerator:
                 raise ValueError(f"No prompts found in file: {prompt_txt_path}")
 
             logger.info("Found %d prompts in %s", len(prompts), prompt_txt_path)
-        elif prompt is not None:
+        else:
+            if prompt is None:
+                prompt = " "
             if isinstance(prompt, str):
                 prompts.append(prompt)
             elif isinstance(prompt, list):
                 prompts.extend(prompt)
-        else:
-            raise ValueError("Either prompt or prompt_txt must be provided")
-
         sampling_params = SamplingParams.from_user_sampling_params_args(
             self.server_args.model_path,
             server_args=self.server_args,

python/sglang/multimodal_gen/runtime/entrypoints/utils.py

@@ -93,6 +93,15 @@ def post_process_sample(
                 )
             else:
                 quality = 75
+                if len(frames) > 1:
+                    for i, image in enumerate(frames):
+                        parts = save_file_path.rsplit(".", 1)
+                        if len(parts) == 2:
+                            indexed_path = f"{parts[0]}_{i}.{parts[1]}"
+                        else:
+                            indexed_path = f"{save_file_path}_{i}"
+                        imageio.imwrite(indexed_path, image, quality=quality)
+                else:
                     imageio.imwrite(save_file_path, frames[0], quality=quality)
             logger.info(f"Output saved to {CYAN}{save_file_path}{RESET}")
         else:

python/sglang/multimodal_gen/runtime/models/dits/qwen_image.py

@@ -44,7 +44,7 @@ def _get_qkv_projections(
 
 
 class QwenTimestepProjEmbeddings(nn.Module):
-    def __init__(self, embedding_dim):
+    def __init__(self, embedding_dim, use_additional_t_cond=False):
         super().__init__()
 
         self.time_proj = Timesteps(
@@ -53,14 +53,25 @@ class QwenTimestepProjEmbeddings(nn.Module):
         self.timestep_embedder = TimestepEmbedding(
             in_channels=256, time_embed_dim=embedding_dim
         )
+        self.use_additional_t_cond = use_additional_t_cond
+        if use_additional_t_cond:
+            self.addition_t_embedding = nn.Embedding(2, embedding_dim)
 
-    def forward(self, timestep, hidden_states):
+    def forward(self, timestep, hidden_states, addition_t_cond=None):
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(
             timesteps_proj.to(dtype=hidden_states.dtype)
         )  # (N, D)
 
         conditioning = timesteps_emb
+        if self.use_additional_t_cond:
+            if addition_t_cond is None:
+                raise ValueError(
+                    "When additional_t_cond is True, addition_t_cond must be provided."
+                )
+            addition_t_emb = self.addition_t_embedding(addition_t_cond)
+            addition_t_emb = addition_t_emb.to(dtype=hidden_states.dtype)
+            conditioning = conditioning + addition_t_emb
 
         return conditioning
 
@@ -232,6 +243,200 @@ class QwenEmbedRope(nn.Module):
         return freqs.clone().contiguous()
 
 
+class QwenEmbedLayer3DRope(nn.Module):
+    def __init__(self, theta: int, axes_dim: List[int], scale_rope=False):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+        pos_index = torch.arange(4096)
+        neg_index = torch.arange(4096).flip(0) * -1 - 1
+        self.pos_freqs = torch.cat(
+            [
+                self.rope_params(pos_index, self.axes_dim[0], self.theta),
+                self.rope_params(pos_index, self.axes_dim[1], self.theta),
+                self.rope_params(pos_index, self.axes_dim[2], self.theta),
+            ],
+            dim=1,
+        )
+        self.neg_freqs = torch.cat(
+            [
+                self.rope_params(neg_index, self.axes_dim[0], self.theta),
+                self.rope_params(neg_index, self.axes_dim[1], self.theta),
+                self.rope_params(neg_index, self.axes_dim[2], self.theta),
+            ],
+            dim=1,
+        )
+
+        self.scale_rope = scale_rope
+
+    def rope_params(self, index, dim, theta=10000):
+        """
+        Args:
+            index: [0, 1, 2, 3] 1D Tensor representing the position index of the token
+        """
+        device = index.device
+        assert dim % 2 == 0
+        freqs = torch.outer(
+            index,
+            (
+                1.0
+                / torch.pow(
+                    theta,
+                    torch.arange(0, dim, 2, device=device).to(torch.float32).div(dim),
+                )
+            ).to(device=device),
+        )
+        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
+        """
+
+        # When models are initialized under a "meta" device context (e.g. init_empty_weights),
+        # tensors created during __init__ become meta tensors. Calling .to(...) on a meta tensor
+        # raises "Cannot copy out of meta tensor". Rebuild the frequencies on the target device
+        # in that case; otherwise move them if just on a different device.
+        if getattr(self.pos_freqs, "device", torch.device("meta")).type == "meta":
+            pos_index = torch.arange(4096, device=device)
+            neg_index = torch.arange(4096, device=device).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,
+            ).to(device=device)
+            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,
+            ).to(device=device)
+        elif self.pos_freqs.device != device:
+            self.pos_freqs = self.pos_freqs.to(device)
+            self.neg_freqs = self.neg_freqs.to(device)
+
+        if isinstance(video_fhw, list):
+            video_fhw = video_fhw[0]
+        if not isinstance(video_fhw, list):
+            video_fhw = [video_fhw]
+
+        vid_freqs = []
+        max_vid_index = 0
+        layer_num = len(video_fhw) - 1
+        for idx, fhw in enumerate(video_fhw):
+            frame, height, width = fhw
+            if idx != layer_num:
+                video_freq = self._compute_video_freqs(frame, height, width, idx)
+            else:
+                ### For the condition image, we set the layer index to -1
+                video_freq = self._compute_condition_freqs(frame, height, width)
+            video_freq = video_freq.to(device)
+            vid_freqs.append(video_freq)
+
+            if self.scale_rope:
+                max_vid_index = max(height // 2, width // 2, max_vid_index)
+            else:
+                max_vid_index = max(height, width, max_vid_index)
+
+        max_vid_index = max(max_vid_index, layer_num)
+        max_len = max(txt_seq_lens)
+        txt_freqs = self.pos_freqs[max_vid_index : max_vid_index + max_len, ...]
+        vid_freqs = torch.cat(vid_freqs, dim=0)
+
+        return vid_freqs, txt_freqs
+
+    @functools.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()
+
+    @functools.lru_cache(maxsize=None)
+    def _compute_condition_freqs(self, frame, height, width):
+        seq_lens = frame * height * width
+        freqs_pos = self.pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+        freqs_neg = self.neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
+
+        freqs_frame = (
+            freqs_neg[0][-1:].view(frame, 1, 1, -1).expand(frame, height, width, -1)
+        )
+        if self.scale_rope:
+            freqs_height = torch.cat(
+                [freqs_neg[1][-(height - height // 2) :], freqs_pos[1][: height // 2]],
+                dim=0,
+            )
+            freqs_height = freqs_height.view(1, height, 1, -1).expand(
+                frame, height, width, -1
+            )
+            freqs_width = torch.cat(
+                [freqs_neg[2][-(width - width // 2) :], freqs_pos[2][: width // 2]],
+                dim=0,
+            )
+            freqs_width = freqs_width.view(1, 1, width, -1).expand(
+                frame, height, width, -1
+            )
+        else:
+            freqs_height = (
+                freqs_pos[1][:height]
+                .view(1, height, 1, -1)
+                .expand(frame, height, width, -1)
+            )
+            freqs_width = (
+                freqs_pos[2][:width]
+                .view(1, 1, width, -1)
+                .expand(frame, height, width, -1)
+            )
+
+        freqs = torch.cat([freqs_frame, freqs_height, freqs_width], dim=-1).reshape(
+            seq_lens, -1
+        )
+        return freqs.clone().contiguous()
+
+
 class QwenImageCrossAttention(nn.Module):
     def __init__(
         self,
@@ -553,16 +758,30 @@ class QwenImageTransformer2DModel(CachableDiT):
         num_attention_heads = config.arch_config.num_attention_heads
         joint_attention_dim = config.arch_config.joint_attention_dim
         axes_dims_rope = config.arch_config.axes_dims_rope
-        zero_cond_t = getattr(config.arch_config, "zero_cond_t", False)
+        self.zero_cond_t = getattr(config.arch_config, "zero_cond_t", False)
         self.out_channels = out_channels or in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
-        self.zero_cond_t = zero_cond_t
 
+        self.use_additional_t_cond: bool = getattr(
+            config.arch_config, "use_additional_t_cond", False
+        )  # For qwen-image-layered now
+        self.use_layer3d_rope: bool = getattr(
+            config.arch_config, "use_layer3d_rope", False
+        )  # For qwen-image-layered now
+
+        if not self.use_layer3d_rope:
             self.rotary_emb = QwenEmbedRope(
                 theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True
             )
+        else:
+            self.rotary_emb = QwenEmbedLayer3DRope(
+                theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True
+            )
 
-        self.time_text_embed = QwenTimestepProjEmbeddings(embedding_dim=self.inner_dim)
+        self.time_text_embed = QwenTimestepProjEmbeddings(
+            embedding_dim=self.inner_dim,
+            use_additional_t_cond=self.use_additional_t_cond,
+        )
 
         self.txt_norm = RMSNorm(joint_attention_dim, eps=1e-6)
 
@@ -575,7 +794,7 @@ class QwenImageTransformer2DModel(CachableDiT):
                     dim=self.inner_dim,
                     num_attention_heads=num_attention_heads,
                     attention_head_dim=attention_head_dim,
-                    zero_cond_t=zero_cond_t,
+                    zero_cond_t=self.zero_cond_t,
                 )
                 for _ in range(num_layers)
             ]
@@ -597,6 +816,7 @@ class QwenImageTransformer2DModel(CachableDiT):
         img_shapes: Optional[List[Tuple[int, int, int]]] = None,
         txt_seq_lens: Optional[List[int]] = None,
         freqs_cis: tuple[torch.Tensor, torch.Tensor] = None,
+        additional_t_cond: Optional[torch.Tensor] = None,
         guidance: torch.Tensor = None,  # TODO: this should probably be removed
         attention_kwargs: Optional[Dict[str, Any]] = None,
         controlnet_block_samples=None,
@@ -658,7 +878,7 @@ class QwenImageTransformer2DModel(CachableDiT):
         encoder_hidden_states = self.txt_norm(encoder_hidden_states)
         encoder_hidden_states = self.txt_in(encoder_hidden_states)
 
-        temb = self.time_text_embed(timestep, hidden_states)
+        temb = self.time_text_embed(timestep, hidden_states, additional_t_cond)
 
         image_rotary_emb = freqs_cis
         for index_block, block in enumerate(self.transformer_blocks):

python/sglang/multimodal_gen/runtime/models/model_stages/qwen_image_layered.py

@@ -0,0 +1,528 @@
+import inspect
+import math
+from typing import List, Optional, Union
+
+import numpy as np
+import torch
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.utils.torch_utils import randn_tensor
+
+from sglang.multimodal_gen.runtime.distributed import get_local_torch_device
+from sglang.multimodal_gen.runtime.managers.forward_context import set_forward_context
+from sglang.multimodal_gen.runtime.models.vision_utils import load_image
+from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import Req
+from sglang.multimodal_gen.runtime.pipelines_core.stages.base import PipelineStage
+from sglang.multimodal_gen.runtime.server_args import ServerArgs
+from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+
+logger = init_logger(__name__)
+
+
+# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit_plus.calculate_dimensions
+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
+
+
+# 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 sample_mode == "sample":
+        return encoder_output.sample(generator)
+    elif sample_mode == "argmax":
+        return encoder_output.mode()
+    else:
+        return encoder_output
+
+
+# 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
+
+
+class QwenImageLayeredBeforeDenoisingStage(PipelineStage):
+    def __init__(
+        self, vae, tokenizer, processor, transformer, scheduler, model_path
+    ) -> None:
+        self.vae = vae.to(torch.bfloat16)
+        from transformers import Qwen2_5_VLForConditionalGeneration
+
+        self.text_encoder = (
+            Qwen2_5_VLForConditionalGeneration.from_pretrained(
+                model_path, subfolder="text_encoder"
+            )
+            .to(get_local_torch_device())
+            .to(torch.bfloat16)
+        )
+        self.tokenizer = tokenizer
+        self.processor = processor
+        self.transformer = transformer
+        self.scheduler = scheduler
+
+        self.vae_scale_factor = (
+            2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        )
+        self.image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor * 2
+        )
+        self.vl_processor = processor
+        self.tokenizer_max_length = 1024
+        self.latent_channels = self.vae.z_dim if getattr(self, "vae", None) else 16
+
+        self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        self.prompt_template_encode_start_idx = 34
+        self.image_caption_prompt_cn = """<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n# 图像标注器\n你是一个专业的图像标注器。请基于输入图像，撰写图注:\n1.
+使用自然、描述性的语言撰写图注，不要使用结构化形式或富文本形式。\n2. 通过加入以下内容，丰富图注细节：\n - 对象的属性：如数量、颜色、形状、大小、位置、材质、状态、动作等\n -
+对象间的视觉关系：如空间关系、功能关系、动作关系、从属关系、比较关系、因果关系等\n - 环境细节：例如天气、光照、颜色、纹理、气氛等\n - 文字内容：识别图像中清晰可见的文字，不做翻译和解释，用引号在图注中强调\n3.
+保持真实性与准确性：\n - 不要使用笼统的描述\n -
+描述图像中所有可见的信息，但不要加入没有在图像中出现的内容\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>assistant\n"""
+        self.image_caption_prompt_en = """<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n# Image Annotator\nYou are a professional
+image annotator. Please write an image caption based on the input image:\n1. Write the caption using natural,
+descriptive language without structured formats or rich text.\n2. Enrich caption details by including: \n - Object
+attributes, such as quantity, color, shape, size, material, state, position, actions, and so on\n - Vision Relations
+between objects, such as spatial relations, functional relations, possessive relations, attachment relations, action
+relations, comparative relations, causal relations, and so on\n - Environmental details, such as weather, lighting,
+colors, textures, atmosphere, and so on\n - Identify the text clearly visible in the image, without translation or
+explanation, and highlight it in the caption with quotation marks\n3. Maintain authenticity and accuracy:\n - Avoid
+generalizations\n - Describe all visible information in the image, while do not add information not explicitly shown in
+the image\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>assistant\n"""
+        self.default_sample_size = 128
+
+    # 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_image_caption(self, prompt_image, use_en_prompt=True, device=None):
+        if use_en_prompt:
+            prompt = self.image_caption_prompt_en
+        else:
+            prompt = self.image_caption_prompt_cn
+        model_inputs = self.vl_processor(
+            text=prompt,
+            images=prompt_image,
+            padding=True,
+            return_tensors="pt",
+        ).to(device)
+        with set_forward_context(current_timestep=0, attn_metadata=None):
+            generated_ids = self.text_encoder.generate(
+                **model_inputs, max_new_tokens=512
+            )
+            generated_ids_trimmed = [
+                out_ids[len(in_ids) :]
+                for in_ids, out_ids in zip(model_inputs.input_ids, generated_ids)
+            ]
+            output_text = self.vl_processor.batch_decode(
+                generated_ids_trimmed,
+                skip_special_tokens=True,
+                clean_up_tokenization_spaces=False,
+            )[0]
+            return output_text.strip()
+
+    def _get_qwen_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        template = self.prompt_template_encode
+        drop_idx = self.prompt_template_encode_start_idx
+        txt = [template.format(e) for e in prompt]
+        txt_tokens = self.tokenizer(
+            txt,
+            padding=True,
+            return_tensors="pt",
+        ).to(device)
+        encoder_hidden_states = self.text_encoder(
+            input_ids=txt_tokens.input_ids,
+            attention_mask=txt_tokens.attention_mask,
+            output_hidden_states=True,
+        )
+        hidden_states = encoder_hidden_states.hidden_states[-1]
+        split_hidden_states = self._extract_masked_hidden(
+            hidden_states, txt_tokens.attention_mask
+        )
+        split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
+        attn_mask_list = [
+            torch.ones(e.size(0), dtype=torch.long, device=e.device)
+            for e in split_hidden_states
+        ]
+        max_seq_len = max([e.size(0) for e in split_hidden_states])
+        prompt_embeds = torch.stack(
+            [
+                torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))])
+                for u in split_hidden_states
+            ]
+        )
+        encoder_attention_mask = torch.stack(
+            [
+                torch.cat([u, u.new_zeros(max_seq_len - u.size(0))])
+                for u in attn_mask_list
+            ]
+        )
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        return prompt_embeds, encoder_attention_mask
+
+    @staticmethod
+    def _pack_latents(latents, batch_size, num_channels_latents, height, width, layers):
+        latents = latents.view(
+            batch_size, layers, num_channels_latents, height // 2, 2, width // 2, 2
+        )
+        latents = latents.permute(0, 1, 3, 5, 2, 4, 6)
+        latents = latents.reshape(
+            batch_size, layers * (height // 2) * (width // 2), num_channels_latents * 4
+        )
+
+        return latents
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        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
+            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.
+        """
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(
+                prompt, device
+            )
+
+        prompt_embeds = prompt_embeds[:, :max_sequence_length]
+        prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length]
+
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(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(num_images_per_prompt, seq_len)
+
+        return prompt_embeds, prompt_embeds_mask
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline._encode_vae_image
+    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+        self.vae = self.vae.to(get_local_torch_device())
+        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
+        self.vae.to("cpu")
+        return image_latents
+
+    def prepare_latents(
+        self,
+        image,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        layers,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
+        shape = (
+            batch_size,
+            layers + 1,
+            num_channels_latents,
+            height,
+            width,
+        )  ### the generated first image is combined image
+
+        image_latents = None
+        if image is not None:
+            image = image.to(device=device, dtype=dtype)
+            if image.shape[1] != self.latent_channels:
+                image_latents = self._encode_vae_image(image=image, generator=generator)
+            else:
+                image_latents = image
+            if (
+                batch_size > image_latents.shape[0]
+                and batch_size % image_latents.shape[0] == 0
+            ):
+                # expand init_latents for batch_size
+                additional_image_per_prompt = batch_size // image_latents.shape[0]
+                image_latents = torch.cat(
+                    [image_latents] * additional_image_per_prompt, dim=0
+                )
+            elif (
+                batch_size > image_latents.shape[0]
+                and batch_size % image_latents.shape[0] != 0
+            ):
+                raise ValueError(
+                    f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+                )
+            else:
+                image_latents = torch.cat([image_latents], dim=0)
+
+            image_latent_height, image_latent_width = image_latents.shape[3:]
+            image_latents = image_latents.permute(
+                0, 2, 1, 3, 4
+            )  # (b, c, f, h, w) -> (b, f, c, h, w)
+            image_latents = self._pack_latents(
+                image_latents,
+                batch_size,
+                num_channels_latents,
+                image_latent_height,
+                image_latent_width,
+                1,
+            )
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+            latents = self._pack_latents(
+                latents, batch_size, num_channels_latents, height, width, layers + 1
+            )
+        else:
+            latents = latents.to(device=device, dtype=dtype)
+
+        return latents, image_latents
+
+    def forward(
+        self,
+        batch: Req,
+        server_args: ServerArgs,
+    ) -> Req:
+        use_en_prompt = True
+        device = get_local_torch_device()
+        layers = batch.num_frames
+        num_inference_steps = batch.num_inference_steps
+        generator = batch.generator
+
+        assert batch.image_path is not None
+        image = load_image(batch.image_path[0])
+        image = image.convert("RGBA")
+        image_size = image.size
+        resolution = 640  # TODO: support user-specified resolution
+        calculated_width, calculated_height = calculate_dimensions(
+            resolution * resolution, image_size[0] / image_size[1]
+        )
+
+        height = calculated_height
+        width = calculated_width
+
+        multiple_of = self.vae_scale_factor * 2
+        width = width // multiple_of * multiple_of
+        height = height // multiple_of * multiple_of
+
+        # if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels):
+        image = self.image_processor.resize(image, calculated_height, calculated_width)
+        prompt_image = image
+        image = self.image_processor.preprocess(
+            image, calculated_height, calculated_width
+        )
+        image = image.unsqueeze(2)
+        image = image.to(dtype=torch.bfloat16)
+
+        prompt = self.get_image_caption(
+            prompt_image, use_en_prompt=use_en_prompt, device=device
+        )
+
+        prompt_embeds, prompt_embeds_mask = self.encode_prompt(
+            prompt=prompt,
+            device=device,
+        )
+
+        negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(
+            prompt=batch.negative_prompt,
+            device=device,
+        )
+
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents, image_latents = self.prepare_latents(
+            image,
+            1,
+            num_channels_latents,
+            height,
+            width,
+            layers,
+            prompt_embeds.dtype,
+            device,
+            generator,
+        )
+        img_shapes = [
+            [
+                *[
+                    (
+                        1,
+                        height // self.vae_scale_factor // 2,
+                        width // self.vae_scale_factor // 2,
+                    )
+                    for _ in range(layers + 1)
+                ],
+                (
+                    1,
+                    calculated_height // self.vae_scale_factor // 2,
+                    calculated_width // self.vae_scale_factor // 2,
+                ),
+            ]
+        ]
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 0, num_inference_steps + 1)[:-1]
+        image_seq_len = latents.shape[1]
+        base_seqlen = 256 * 256 / 16 / 16
+        mu = (image_latents.shape[1] / base_seqlen) ** 0.5
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            mu=mu,
+        )
+
+        txt_seq_lens = (
+            prompt_embeds_mask.sum(dim=1).tolist()
+            if prompt_embeds_mask is not None
+            else None
+        )
+        negative_txt_seq_lens = (
+            negative_prompt_embeds_mask.sum(dim=1).tolist()
+            if negative_prompt_embeds_mask is not None
+            else None
+        )
+        is_rgb = torch.tensor([0]).to(device=device, dtype=torch.long)
+
+        batch.prompt_embeds = [prompt_embeds]
+        batch.prompt_embeds_mask = [prompt_embeds_mask]
+        batch.negative_prompt_embeds = [negative_prompt_embeds]
+        batch.negative_prompt_embeds_mask = [negative_prompt_embeds_mask]
+        batch.latents = latents
+        batch.image_latent = image_latents
+        batch.num_inference_steps = num_inference_steps
+        batch.sigmas = sigmas.tolist()  # Convert numpy array to list for validation
+        batch.generator = torch.manual_seed(0)
+        batch.original_condition_image_size = image_size
+        batch.raw_latent_shape = latents.shape
+        batch.txt_seq_lens = txt_seq_lens
+        batch.img_shapes = img_shapes
+
+        return batch

python/sglang/multimodal_gen/runtime/models/vaes/autoencoder_kl_qwenimage.py

@@ -445,6 +445,7 @@ class QwenImageEncoder3d(nn.Module):
         temperal_downsample=[True, True, False],
         dropout=0.0,
         non_linearity: str = "silu",
+        input_channels: int = 3,
     ):
         super().__init__()
         # dim = config.arch_config.dim
@@ -468,7 +469,7 @@ class QwenImageEncoder3d(nn.Module):
         scale = 1.0
 
         # init block
-        self.conv_in = QwenImageCausalConv3d(3, dims[0], 3, padding=1)
+        self.conv_in = QwenImageCausalConv3d(input_channels, dims[0], 3, padding=1)
 
         # downsample blocks
         self.down_blocks = nn.ModuleList([])
@@ -649,6 +650,7 @@ class QwenImageDecoder3d(nn.Module):
         temperal_upsample=[False, True, True],
         dropout=0.0,
         non_linearity: str = "silu",
+        input_channels=3,
     ):
         super().__init__()
         self.dim = dim
@@ -701,7 +703,7 @@ class QwenImageDecoder3d(nn.Module):
 
         # output blocks
         self.norm_out = QwenImageRMS_norm(out_dim, images=False)
-        self.conv_out = QwenImageCausalConv3d(out_dim, 3, 3, padding=1)
+        self.conv_out = QwenImageCausalConv3d(out_dim, input_channels, 3, padding=1)
 
         self.gradient_checkpointing = False
 
@@ -783,15 +785,19 @@ class AutoencoderKLQwenImage(nn.Module):
         self.z_dim = z_dim
         self.temperal_downsample = temperal_downsample
         self.temperal_upsample = temperal_downsample[::-1]
+        self.input_channels = config.arch_config.input_channels
+        self.latents_mean = config.arch_config.latents_mean
+        self.config = config.arch_config
+
 
         self.encoder = QwenImageEncoder3d(
-            base_dim, z_dim * 2, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout
+            base_dim, z_dim * 2, dim_mult, num_res_blocks, attn_scales, self.temperal_downsample, dropout, input_channels=self.input_channels
         )
         self.quant_conv = QwenImageCausalConv3d(z_dim * 2, z_dim * 2, 1)
         self.post_quant_conv = QwenImageCausalConv3d(z_dim, z_dim, 1)
 
         self.decoder = QwenImageDecoder3d(
-            base_dim, z_dim, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout
+            base_dim, z_dim, dim_mult, num_res_blocks, attn_scales, self.temperal_upsample, dropout, input_channels=self.input_channels
         )
 
         self.spatial_compression_ratio = 2 ** len(self.temperal_downsample)

python/sglang/multimodal_gen/runtime/pipelines/qwen_image.py

@@ -3,6 +3,9 @@
 # SPDX-License-Identifier: Apache-2.0
 from diffusers.image_processor import VaeImageProcessor
 
+from sglang.multimodal_gen.runtime.models.model_stages.qwen_image_layered import (
+    QwenImageLayeredBeforeDenoisingStage,
+)
 from sglang.multimodal_gen.runtime.pipelines_core import LoRAPipeline
 from sglang.multimodal_gen.runtime.pipelines_core.composed_pipeline_base import (
     ComposedPipelineBase,
@@ -196,4 +199,62 @@ class QwenImageEditPlusPipeline(QwenImageEditPipeline):
     pipeline_name = "QwenImageEditPlusPipeline"
 
 
-EntryClass = [QwenImagePipeline, QwenImageEditPipeline, QwenImageEditPlusPipeline]
+def prepare_mu_layered(batch: Req, server_args: ServerArgs):
+    base_seqlen = 256 * 256 / 16 / 16
+    mu = (batch.image_latent.shape[1] / base_seqlen) ** 0.5
+    return "mu", mu
+
+
+class QwenImageLayeredPipeline(QwenImageEditPipeline):
+    pipeline_name = "QwenImageLayeredPipeline"
+
+    _required_config_modules = [
+        "vae",
+        "tokenizer",
+        "processor",
+        "transformer",
+        "scheduler",
+    ]
+
+    def create_pipeline_stages(self, server_args: ServerArgs):
+        """Set up pipeline stages with proper dependency injection."""
+
+        self.add_stage(
+            stage_name="QwenImageLayeredBeforeDenoisingStage",
+            stage=QwenImageLayeredBeforeDenoisingStage(
+                vae=self.get_module("vae"),
+                tokenizer=self.get_module("tokenizer"),
+                processor=self.get_module("processor"),
+                transformer=self.get_module("transformer"),
+                scheduler=self.get_module("scheduler"),
+                model_path=self.model_path,
+            ),
+        )
+
+        self.add_stage(
+            stage_name="timestep_preparation_stage",
+            stage=TimestepPreparationStage(
+                scheduler=self.get_module("scheduler"),
+                prepare_extra_set_timesteps_kwargs=[prepare_mu_layered],
+            ),
+        )
+
+        self.add_stage(
+            stage_name="denoising_stage",
+            stage=DenoisingStage(
+                transformer=self.get_module("transformer"),
+                scheduler=self.get_module("scheduler"),
+            ),
+        )
+
+        self.add_stage(
+            stage_name="decoding_stage", stage=DecodingStage(vae=self.get_module("vae"))
+        )
+
+
+EntryClass = [
+    QwenImagePipeline,
+    QwenImageEditPipeline,
+    QwenImageEditPlusPipeline,
+    QwenImageLayeredPipeline,
+]