[diffusion] model: support stable-diffusion-3-medium-diffusers (#19225)

Co-authored-by: zhaochenyang20 <zhaochen20@outlook.com> Co-authored-by: Kangrui Du <kangruidu@gmail.com> Co-authored-by: Xiaole Guo <gxlvera@gmail.com>
2026-04-13 16:07:06 +08:00
parent 9e6d1c066e
commit 4df60434d7
17 changed files with 800 additions and 29 deletions
@@ -5,6 +5,9 @@ from sglang.multimodal_gen.configs.models.dits.hunyuan3d import Hunyuan3DDiTConf
 from sglang.multimodal_gen.configs.models.dits.hunyuanvideo import HunyuanVideoConfig
 from sglang.multimodal_gen.configs.models.dits.mova_audio import MOVAAudioConfig
 from sglang.multimodal_gen.configs.models.dits.mova_video import MOVAVideoConfig
+from sglang.multimodal_gen.configs.models.dits.stablediffusion3 import (
+    StableDiffusion3TransformerConfig,
+)
 from sglang.multimodal_gen.configs.models.dits.wanvideo import WanVideoConfig

 __all__ = [
@@ -14,4 +17,5 @@ __all__ = [
    "Hunyuan3DDiTConfig",
    "MOVAAudioConfig",
    "MOVAVideoConfig",
+    "StableDiffusion3TransformerConfig",
 ]
@@ -0,0 +1,37 @@
+# SPDX-License-Identifier: Apache-2.0
+"""StableDiffusion3 Transformer model configuration"""
+
+from dataclasses import dataclass, field
+
+from sglang.multimodal_gen.configs.models.dits.base import DiTArchConfig, DiTConfig
+
+
+@dataclass
+class StableDiffusion3TransformerArchConfig(DiTArchConfig):
+    """Architecture configuration for StableDiffusion3 Transformer, applicable to SD3-medium, SD3.5-medium, SD3.5-large."""
+
+    sample_size: int = 128
+    patch_size: int = 2
+    in_channels: int = 16
+    out_channels: int = 16
+    num_layers: int = 18
+    attention_head_dim: int = 64
+    num_attention_heads: int = 18
+    cross_attention_dim: int = 4096
+    joint_attention_dim: int = 4096
+    caption_projection_dim: int = 1152
+    pooled_projection_dim: int = 2048
+    pos_embed_max_size: int = 96
+    dual_attention_layers: tuple[int, ...] = ()
+    qk_norm: str | None = None
+
+    _class_name: str = "SD3Transformer2DModel"
+
+
+@dataclass
+class StableDiffusion3TransformerConfig(DiTConfig):
+    """Configuration for StableDiffusion3 Transformer model."""
+
+    arch_config: StableDiffusion3TransformerArchConfig = field(
+        default_factory=StableDiffusion3TransformerArchConfig
+    )
@@ -3,11 +3,15 @@
 from sglang.multimodal_gen.configs.models.vaes.dac import DacVAEConfig
 from sglang.multimodal_gen.configs.models.vaes.hunyuan3d import Hunyuan3DVAEConfig
 from sglang.multimodal_gen.configs.models.vaes.hunyuanvae import HunyuanVAEConfig
+from sglang.multimodal_gen.configs.models.vaes.stablediffusion3 import (
+    StableDiffusion3VAEConfig,
+)
 from sglang.multimodal_gen.configs.models.vaes.wanvae import WanVAEConfig

 __all__ = [
    "DacVAEConfig",
    "HunyuanVAEConfig",
+    "StableDiffusion3VAEConfig",
    "WanVAEConfig",
    "Hunyuan3DVAEConfig",
 ]
@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+"""StableDiffusion3 VAE configuration."""
+
+from dataclasses import dataclass, field
+
+from sglang.multimodal_gen.configs.models.vaes.base import VAEArchConfig, VAEConfig
+
+
+@dataclass
+class StableDiffusion3VAEArchConfig(VAEArchConfig):
+    """Architecture configuration for StableDiffusion3 VAE."""
+
+    scaling_factor: float = 1.5305
+    shift_factor: float = 0.0609
+
+    spatial_compression_ratio: int = 8
+    temporal_compression_ratio: int = 1
+
+    in_channels: int = 3
+    out_channels: int = 3
+    latent_channels: int = 16
+    sample_size: int = 128
+
+    block_out_channels: tuple[int, ...] = (128, 256, 512, 512)
+    layers_per_block: int = 2
+    act_fn: str = "silu"
+    norm_num_groups: int = 32
+
+    down_block_types: tuple[str, ...] = (
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D",
+    )
+    up_block_types: tuple[str, ...] = (
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D",
+    )
+
+    attention_head_dim: int = 8
+    mid_block_add_attention: bool = True
+    use_quant_conv: bool = False
+    use_post_quant_conv: bool = False
+
+
+@dataclass
+class StableDiffusion3VAEConfig(VAEConfig):
+    """Configuration for StableDiffusion3 VAE."""
+
+    arch_config: StableDiffusion3VAEArchConfig = field(
+        default_factory=StableDiffusion3VAEArchConfig
+    )
+
+    tile_sample_min_height: int = 512
+    tile_sample_min_width: int = 512
+    tile_sample_min_num_frames: int = 1
+    tile_sample_stride_height: int = 448
+    tile_sample_stride_width: int = 448
+    tile_sample_stride_num_frames: int = 1
+
+    use_tiling: bool = True
+    use_temporal_tiling: bool = False
+    use_parallel_tiling: bool = True
+    use_temporal_scaling_frames: bool = False
+
+    def __post_init__(self) -> None:
+        """Post initialization for SD3 VAE specific setup."""
+        super().__post_init__()
+        self.update_model_arch({"_class_name": "AutoencoderKL"})
+        self.blend_num_frames = 0
@@ -30,6 +30,9 @@ from sglang.multimodal_gen.configs.pipeline_configs.hunyuan3d import (
 from sglang.multimodal_gen.configs.pipeline_configs.ltx_2 import LTX2PipelineConfig
 from sglang.multimodal_gen.configs.pipeline_configs.mova import MOVAPipelineConfig
 from sglang.multimodal_gen.configs.pipeline_configs.sana import SanaPipelineConfig
+from sglang.multimodal_gen.configs.pipeline_configs.stablediffusion3 import (
+    StableDiffusion3PipelineConfig,
+)
 from sglang.multimodal_gen.configs.pipeline_configs.wan import (
    SelfForcingWanT2V480PConfig,
    WanI2V480PConfig,
@@ -55,6 +58,7 @@ __all__ = [
    "SanaPipelineConfig",
    "SlidingTileAttnConfig",
    "MOVAPipelineConfig",
+    "StableDiffusion3PipelineConfig",
    "WanT2V480PConfig",
    "WanI2V480PConfig",
    "WanT2V720PConfig",
@@ -459,6 +459,33 @@ class PipelineConfig:
        sharded_tensor = sharded_tensor[:, :, rank_in_sp_group, :, :, :]
        return sharded_tensor, True

+    def get_text_encoder_attention_mask(
+        self, text_inputs: dict, encoder_index: int
+    ) -> "torch.Tensor | None":
+        """Return the attention mask for the given text encoder.
+
+        Override to suppress (return None) or modify the mask per model.
+        """
+        return text_inputs.get("attention_mask")
+
+    def get_text_encoder_pooler_output(
+        self, outputs: "BaseEncoderOutput", encoder_index: int
+    ) -> "torch.Tensor | None":
+        """Return the pooler output for the given text encoder, or None to skip.
+
+        Override for models that need pooled embeddings (e.g. FLUX v1, SD3).
+        """
+        return None
+
+    def select_vae_weight_files(
+        self,
+        safetensors_list: list[str],
+        component_model_path: str,
+        component_name: str,
+        vae_precision: str,
+    ) -> list[str]:
+        return safetensors_list
+
    def get_pos_prompt_embeds(self, batch):
        return batch.prompt_embeds

@@ -81,6 +81,13 @@ class FluxPipelineConfig(ImagePipelineConfig):
        ]
    )

+    def get_text_encoder_attention_mask(self, text_inputs, encoder_index):
+        # Flux v1 does not use attention masks for text encoders.
+        return None
+
+    def get_text_encoder_pooler_output(self, outputs, encoder_index):
+        return outputs.pooler_output
+
    def prepare_sigmas(self, sigmas, num_inference_steps):
        return self._prepare_sigmas(sigmas, num_inference_steps)

@@ -391,6 +398,14 @@ class Flux2PipelineConfig(FluxPipelineConfig):
        ]
    )

+    def get_text_encoder_attention_mask(self, text_inputs, encoder_index):
+        # Flux2 uses standard attention masks (unlike Flux v1).
+        return text_inputs.get("attention_mask")
+
+    def get_text_encoder_pooler_output(self, outputs, encoder_index):
+        # Flux2 does not use pooler output.
+        return None
+
    def tokenize_prompt(self, prompts: list[str], tokenizer, tok_kwargs) -> dict:
        messages = build_flux2_text_messages(prompts)
        inputs = tokenizer.apply_chat_template(
@@ -0,0 +1,202 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Stable Diffusion 3 pipeline configuration."""
+
+import os
+from dataclasses import dataclass, field
+from typing import Callable
+
+import torch
+
+from sglang.multimodal_gen.configs.models import DiTConfig, EncoderConfig, VAEConfig
+from sglang.multimodal_gen.configs.models.dits import StableDiffusion3TransformerConfig
+from sglang.multimodal_gen.configs.models.encoders import BaseEncoderOutput
+from sglang.multimodal_gen.configs.models.encoders.base import TextEncoderArchConfig
+from sglang.multimodal_gen.configs.models.encoders.clip import (
+    CLIPTextArchConfig,
+    CLIPTextConfig,
+)
+from sglang.multimodal_gen.configs.models.encoders.t5 import (
+    T5ArchConfig,
+    T5Config,
+)
+from sglang.multimodal_gen.configs.models.vaes.stablediffusion3 import (
+    StableDiffusion3VAEConfig,
+)
+from sglang.multimodal_gen.configs.pipeline_configs.base import (
+    ModelTaskType,
+    SpatialImagePipelineConfig,
+)
+
+
+def sd3_clip_postprocess_text(outputs: BaseEncoderOutput, _text_inputs) -> torch.Tensor:
+    """Extract pre-final hidden state for SD3 CLIP encoders."""
+    if outputs.hidden_states is None:
+        raise ValueError(
+            "SD3 CLIP postprocessing requires hidden_states from encoder output."
+        )
+    return outputs.hidden_states[-2]
+
+
+def t5_postprocess_text(outputs: BaseEncoderOutput, _text_inputs) -> torch.Tensor:
+    return outputs.last_hidden_state
+
+
+def select_sd3_vae_weight_files(
+    safetensors_list: list[str],
+    component_model_path: str,
+    component_name: str,
+    vae_precision: str,
+) -> list[str]:
+    """Select SD3 VAE checkpoint file candidates with minimal policy."""
+    if component_name not in ("vae", "video_vae"):
+        return safetensors_list
+
+    base_name = "diffusion_pytorch_model"
+    if vae_precision == "fp16":
+        fp16_path = os.path.join(component_model_path, f"{base_name}.fp16.safetensors")
+        if os.path.exists(fp16_path):
+            return [fp16_path]
+
+    full_path = os.path.join(component_model_path, f"{base_name}.safetensors")
+    if os.path.exists(full_path):
+        return [full_path]
+    return safetensors_list
+
+
+@dataclass
+class SD3CLIPTextArchConfig(CLIPTextArchConfig):
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        self.tokenizer_kwargs.update(
+            {
+                "max_length": self.text_len,
+                "padding": "max_length",
+            }
+        )
+
+
+@dataclass
+class SD3CLIPTextConfig(CLIPTextConfig):
+    arch_config: TextEncoderArchConfig = field(default_factory=SD3CLIPTextArchConfig)
+
+
+@dataclass
+class SD3T5ArchConfig(T5ArchConfig):
+    def __post_init__(self) -> None:
+        super().__post_init__()
+        self.tokenizer_kwargs.update({"max_length": 256})
+
+
+@dataclass
+class SD3T5Config(T5Config):
+    arch_config: TextEncoderArchConfig = field(default_factory=SD3T5ArchConfig)
+
+
+@dataclass
+class StableDiffusion3PipelineConfig(SpatialImagePipelineConfig):
+    """Configuration for SD3 image generation pipeline.
+
+    This config intentionally relies on SD3-specific encoder configs to provide
+    tokenizer kwargs, instead of stage-level tokenizer overrides.
+    """
+
+    task_type: ModelTaskType = ModelTaskType.T2I
+
+    dit_config: DiTConfig = field(default_factory=StableDiffusion3TransformerConfig)
+    vae_config: VAEConfig = field(default_factory=StableDiffusion3VAEConfig)
+
+    text_encoder_configs: tuple[EncoderConfig, ...] = field(
+        default_factory=lambda: (
+            SD3CLIPTextConfig(),
+            SD3CLIPTextConfig(),
+            SD3T5Config(),
+        )
+    )
+
+    text_encoder_precisions: tuple[str, ...] = field(
+        default_factory=lambda: ("fp16", "fp16", "fp32")
+    )
+
+    preprocess_text_funcs: tuple[Callable[[str], str] | None, ...] = field(
+        default_factory=lambda: (
+            None,
+            None,
+            None,
+        )
+    )
+
+    postprocess_text_funcs: tuple[
+        Callable[[BaseEncoderOutput, dict], torch.Tensor], ...
+    ] = field(
+        default_factory=lambda: (
+            sd3_clip_postprocess_text,
+            sd3_clip_postprocess_text,
+            t5_postprocess_text,
+        )
+    )
+
+    should_use_guidance: bool = False
+    guidance_scale: float = 7.0
+
+    def __post_init__(self) -> None:
+        configs = list(self.text_encoder_configs)
+        configs[0].update_model_arch({"_class_name": "CLIPTextModelWithProjection"})
+        configs[1].update_model_arch({"_class_name": "CLIPTextModelWithProjection"})
+        configs[2].update_model_arch({"_class_name": "T5EncoderModel"})
+        self.text_encoder_configs = tuple(configs)
+
+    def get_text_encoder_pooler_output(self, outputs, encoder_index):
+        # SD3 uses pooled embeddings only from the two CLIP encoders (indices 0 and 1).
+        if encoder_index <= 1:
+            return outputs.pooler_output
+        return None
+
+    def select_vae_weight_files(
+        self,
+        safetensors_list: list[str],
+        component_model_path: str,
+        component_name: str,
+        vae_precision: str,
+    ) -> list[str]:
+        return select_sd3_vae_weight_files(
+            safetensors_list=safetensors_list,
+            component_model_path=component_model_path,
+            component_name=component_name,
+            vae_precision=vae_precision,
+        )
+
+    def tokenize_prompt(self, prompt: list[str], tokenizer, tok_kwargs) -> dict:
+        text_inputs = tokenizer(prompt, **tok_kwargs)
+        text_inputs["attention_mask"] = None
+        return text_inputs
+
+    def get_pos_prompt_embeds(self, batch):
+        return batch.prompt_embeds[0]
+
+    def get_neg_prompt_embeds(self, batch):
+        return batch.negative_prompt_embeds[0]
+
+    def prepare_pos_cond_kwargs(self, batch, device, rotary_emb, dtype):
+        return {
+            "pooled_projections": (
+                batch.pooled_embeds[0] if batch.pooled_embeds else None
+            )
+        }
+
+    def prepare_neg_cond_kwargs(self, batch, device, rotary_emb, dtype):
+        return {
+            "pooled_projections": (
+                batch.neg_pooled_embeds[0] if batch.neg_pooled_embeds else None
+            )
+        }
+
+    # SD3 image latents are spatial (B, C, H, W), not video-like (B, C, T, H, W).
+    def prepare_latent_shape(self, batch, batch_size, num_frames):  # noqa: ARG002
+        spatial_ratio = self.vae_config.arch_config.spatial_compression_ratio
+        in_channels = self.dit_config.arch_config.in_channels
+        return (
+            batch_size,
+            in_channels,
+            batch.height // spatial_ratio,
+            batch.width // spatial_ratio,
+        )
@@ -0,0 +1,20 @@
+# Copied and adapted from: https://github.com/hao-ai-lab/FastVideo
+
+# SPDX-License-Identifier: Apache-2.0
+"""StableDiffusion3 sampling parameters configuration."""
+
+from dataclasses import dataclass
+
+from sglang.multimodal_gen.configs.sample.sampling_params import SamplingParams
+
+
+@dataclass
+class StableDiffusion3SamplingParams(SamplingParams):
+    """Sampling parameters for StableDiffusion3."""
+
+    # A single space ensures tokenizers produce valid (non-empty) input for CFG.
+    negative_prompt: str = " "
+    num_frames: int = 1
+    num_inference_steps: int = 50
+    guidance_scale: float = 7.0
+    guidance_rescale: float = 0.0
@@ -68,6 +68,9 @@ from sglang.multimodal_gen.configs.pipeline_configs.qwen_image import (
    QwenImagePipelineConfig,
 )
 from sglang.multimodal_gen.configs.pipeline_configs.sana import SanaPipelineConfig
+from sglang.multimodal_gen.configs.pipeline_configs.stablediffusion3 import (
+    StableDiffusion3PipelineConfig,
+)
 from sglang.multimodal_gen.configs.pipeline_configs.wan import (
    FastWan2_1_T2V_480P_Config,
    FastWan2_2_TI2V_5B_Config,
@@ -109,6 +112,9 @@ from sglang.multimodal_gen.configs.sample.qwenimage import (
    QwenImageSamplingParams,
 )
 from sglang.multimodal_gen.configs.sample.sana import SanaSamplingParams
+from sglang.multimodal_gen.configs.sample.stablediffusion3 import (
+    StableDiffusion3SamplingParams,
+)
 from sglang.multimodal_gen.configs.sample.wan import (
    FastWanT2V480PConfig,
    Turbo_Wan2_2_I2V_A14B_SamplingParam,
@@ -849,6 +855,26 @@ def _register_configs():
        hf_model_paths=["Qwen/Qwen-Image-Layered"],
        model_detectors=[lambda hf_id: "qwen-image-layered" in hf_id.lower()],
    )
+    register_configs(
+        sampling_param_cls=StableDiffusion3SamplingParams,
+        pipeline_config_cls=StableDiffusion3PipelineConfig,
+        hf_model_paths=[
+            "stabilityai/stable-diffusion-3-medium",
+            "stabilityai/stable-diffusion-3-medium-diffusers",
+            "stabilityai/stable-diffusion-3.5-medium",
+            "stabilityai/stable-diffusion-3.5-medium-diffusers",
+            "stabilityai/stable-diffusion-3.5-large",
+            "stabilityai/stable-diffusion-3.5-large-diffusers",
+        ],
+        model_detectors=[
+            lambda hf_id: "stable-diffusion-3-medium" in hf_id.lower()
+            or "stable-diffusion-3.5-medium" in hf_id.lower()
+            or "stable-diffusion-3.5-large" in hf_id.lower()
+            or "sd3-medium" in hf_id.lower()
+            or "sd3.5-medium" in hf_id.lower()
+            or "sd3.5-large" in hf_id.lower()
+        ],
+    )

    register_configs(
        sampling_param_cls=GlmImageSamplingParams,
@@ -1,6 +1,7 @@
 import dataclasses
 import glob
 import os
+import re
 from collections.abc import Generator, Iterable
 from typing import cast

@@ -218,21 +219,21 @@ class TextEncoderLoader(ComponentLoader):
            component_path=component_model_path
        )

-        def is_not_first_encoder(module_name):
-            return "2" in module_name
-
        # TODO(mick): had to throw an exception for different text-encoder arch
-        if not is_not_first_encoder(component_name):
-            encoder_config = server_args.pipeline_config.text_encoder_configs[0]
-            encoder_config.update_model_arch(model_config)
+        encoder_index = self._extract_encoder_index(component_name)
+        assert encoder_index < len(
+            server_args.pipeline_config.text_encoder_configs
+        ) and encoder_index < len(server_args.pipeline_config.text_encoder_precisions)
+
+        encoder_config = server_args.pipeline_config.text_encoder_configs[encoder_index]
+        encoder_config.update_model_arch(model_config)
+
+        if encoder_index == 0:
            for key, value in diffusers_pretrained_config.__dict__.items():
                setattr(encoder_config.arch_config, key, value)
-            encoder_dtype = server_args.pipeline_config.text_encoder_precisions[0]
-        else:
-            assert len(server_args.pipeline_config.text_encoder_configs) == 2
-            encoder_config = server_args.pipeline_config.text_encoder_configs[1]
-            encoder_config.update_model_arch(model_config)
-            encoder_dtype = server_args.pipeline_config.text_encoder_precisions[1]
+        encoder_dtype = server_args.pipeline_config.text_encoder_precisions[
+            encoder_index
+        ]
        # TODO(will): add support for other dtypes
        return self.load_model(
            component_model_path,
@@ -242,6 +243,28 @@ class TextEncoderLoader(ComponentLoader):
            cpu_offload_flag=cpu_offload_flag,
        )

+    @staticmethod
+    def _extract_encoder_index(component_name: str) -> int:
+        """
+        Map text encoder component names to zero-based indices.
+
+        Examples:
+        - text_encoder -> 0
+        - text_encoder_2 -> 1
+        - text_encoder_3 -> 2
+        """
+        match = re.search(r"_(\d+)$", component_name)
+        if match is None:
+            return 0
+
+        suffix_num = int(match.group(1))
+        if suffix_num <= 0:
+            raise ValueError(
+                f"Invalid text encoder component name '{component_name}': "
+                "numeric suffix must be >= 1."
+            )
+        return suffix_num - 1
+
    def load_model(
        self,
        model_path: str,
@@ -129,6 +129,13 @@ class VAELoader(ComponentLoader):
            vae = vae_cls(vae_config).to(target_device)

        safetensors_list = _list_safetensors_files(component_model_path)
+        safetensors_list = server_args.pipeline_config.select_vae_weight_files(
+            safetensors_list=safetensors_list,
+            component_model_path=component_model_path,
+            component_name=component_name,
+            vae_precision=vae_precision,
+        )
+
        assert (
            len(safetensors_list) >= 1
        ), f"Found no safetensors files in {component_model_path}"
@@ -181,9 +181,7 @@ class skip_init_modules:

 def _normalize_component_type(module_type: str) -> str:
    """Normalize module types like 'text_encoder_2' -> 'text_encoder'."""
-    if module_type.endswith("_2"):
-        return module_type[:-2]
-    return module_type
+    return re.sub(r"_\d+$", "", module_type)


 def _clean_hf_config_inplace(model_config: dict) -> None:
@@ -0,0 +1,180 @@
+# SPDX-License-Identifier: Apache-2.0
+"""StableDiffusion3 Transformer model implementation.
+
+NOTE: This initial implementation uses diffusers' JointTransformerBlock directly.
+A native SGLang attention implementation is needed for FlashAttention, TP/SP,
+quantization, and LoRA support.
+"""
+
+from typing import Any
+
+import torch
+import torch.nn as nn
+from diffusers.models.attention import JointTransformerBlock
+from diffusers.models.embeddings import CombinedTimestepTextProjEmbeddings, PatchEmbed
+from diffusers.models.normalization import AdaLayerNormContinuous
+
+from sglang.multimodal_gen.configs.models.dits.stablediffusion3 import (
+    StableDiffusion3TransformerConfig,
+)
+from sglang.multimodal_gen.runtime.models.dits.base import CachableDiT
+from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+
+logger = init_logger(__name__)
+
+
+class SD3Transformer2DModel(CachableDiT):
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["JointTransformerBlock"]
+    _skip_layerwise_casting_patterns = ["pos_embed", "norm"]
+
+    def __init__(
+        self,
+        config: StableDiffusion3TransformerConfig,
+        hf_config: dict[str, Any] | None = None,
+        quant_config=None,
+    ):
+        super().__init__(config=config, hf_config=hf_config)
+        self.config = config
+        arch_config = config.arch_config
+        sample_size = arch_config.sample_size
+        patch_size = arch_config.patch_size
+        in_channels = arch_config.in_channels
+        num_layers = arch_config.num_layers
+        attention_head_dim = arch_config.attention_head_dim
+        num_attention_heads = arch_config.num_attention_heads
+        joint_attention_dim = arch_config.joint_attention_dim
+        caption_projection_dim = arch_config.caption_projection_dim
+        pooled_projection_dim = arch_config.pooled_projection_dim
+        out_channels = arch_config.out_channels
+        pos_embed_max_size = arch_config.pos_embed_max_size
+        dual_attention_layers = arch_config.dual_attention_layers
+        qk_norm = arch_config.qk_norm
+
+        self.out_channels = out_channels if out_channels is not None else in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+        self.patch_size = patch_size
+
+        self.pos_embed = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=self.inner_dim,
+            pos_embed_max_size=pos_embed_max_size,
+        )
+        self.time_text_embed = CombinedTimestepTextProjEmbeddings(
+            embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim
+        )
+        self.context_embedder = nn.Linear(joint_attention_dim, caption_projection_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                JointTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    context_pre_only=i == num_layers - 1,
+                    qk_norm=qk_norm,
+                    use_dual_attention=i in dual_attention_layers,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        self.norm_out = AdaLayerNormContinuous(
+            self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6
+        )
+        self.proj_out = nn.Linear(
+            self.inner_dim, patch_size * patch_size * self.out_channels, bias=True
+        )
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor | None = None,
+        pooled_projections: torch.Tensor | None = None,
+        timestep: torch.LongTensor | None = None,
+        block_controlnet_hidden_states: list | None = None,
+        guidance: torch.Tensor | None = None,
+        joint_attention_kwargs: dict[str, Any] | None = None,
+        skip_layers: list[int] | None = None,
+    ) -> torch.Tensor:
+        if encoder_hidden_states is None:
+            raise ValueError("encoder_hidden_states must be provided.")
+        if pooled_projections is None:
+            raise ValueError("pooled_projections must be provided.")
+
+        encoder_embeddings = encoder_hidden_states
+
+        height, width = hidden_states.shape[-2:]
+
+        hidden_states = self.pos_embed(hidden_states)
+        temb = self.time_text_embed(timestep, pooled_projections)
+        encoder_embeddings = self.context_embedder(encoder_embeddings)
+
+        skip_layer_set = set(skip_layers) if skip_layers else set()
+
+        if block_controlnet_hidden_states is not None:
+            interval_control = len(self.transformer_blocks) / len(
+                block_controlnet_hidden_states
+            )
+        else:
+            interval_control = 0
+
+        for index_block, block in enumerate(self.transformer_blocks):
+            if index_block not in skip_layer_set:
+                encoder_embeddings, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_embeddings,
+                    temb=temb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+
+            # controlnet residual
+            if (
+                block_controlnet_hidden_states is not None
+                and block.context_pre_only is False
+            ):
+                hidden_states = (
+                    hidden_states
+                    + block_controlnet_hidden_states[
+                        int(index_block / interval_control)
+                    ]
+                )
+
+        hidden_states = self.norm_out(hidden_states, temb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # unpatchify
+        patch_size = self.patch_size
+        height = height // patch_size
+        width = width // patch_size
+
+        hidden_states = hidden_states.reshape(
+            shape=(
+                hidden_states.shape[0],
+                height,
+                width,
+                patch_size,
+                patch_size,
+                self.out_channels,
+            )
+        )
+        hidden_states = hidden_states.permute(0, 5, 1, 3, 2, 4)
+        output = hidden_states.reshape(
+            shape=(
+                hidden_states.shape[0],
+                self.out_channels,
+                height * patch_size,
+                width * patch_size,
+            )
+        )
+
+        return output
+
+
+# Entry class for registry
+EntryClass = SD3Transformer2DModel
@@ -594,6 +594,48 @@ class CLIPTextModel(TextEncoder):
        return loaded_params


+class CLIPTextModelWithProjection(CLIPTextModel):
+    """
+    CLIP text encoder with projection head for pooled_output.
+    """
+
+    def __init__(
+        self,
+        config: CLIPTextConfig,
+    ) -> None:
+        super().__init__(config)
+        self.text_projection = nn.Linear(
+            config.hidden_size, config.projection_dim, bias=False
+        )
+
+    def forward(
+        self,
+        input_ids: torch.Tensor | None,
+        position_ids: torch.Tensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        output_hidden_states: bool | None = None,
+        **kwargs,
+    ) -> BaseEncoderOutput:
+        outputs: BaseEncoderOutput = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_hidden_states=output_hidden_states,
+        )
+
+        pooled_output = outputs.pooler_output
+        if pooled_output is not None:
+            pooled_output = self.text_projection(pooled_output)
+
+        return BaseEncoderOutput(
+            last_hidden_state=outputs.last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
 class CLIPVisionTransformer(nn.Module):

    def __init__(
@@ -759,4 +801,4 @@ class BertModel(CLIPTextModel):
    pass


-EntryClass = [CLIPTextModel, CLIPVisionModel]
+EntryClass = [CLIPTextModel, CLIPTextModelWithProjection, CLIPVisionModel]
@@ -0,0 +1,111 @@
+# Copied and adapted from: https://github.com/hao-ai-lab/FastVideo
+
+# SPDX-License-Identifier: Apache-2.0
+"""StableDiffusion3 pipeline implementation."""
+
+import torch
+
+from sglang.multimodal_gen.runtime.pipelines_core.composed_pipeline_base import (
+    ComposedPipelineBase,
+)
+from sglang.multimodal_gen.runtime.pipelines_core.schedule_batch import Req
+from sglang.multimodal_gen.runtime.pipelines_core.stages import (
+    InputValidationStage,
+    PipelineStage,
+    TextEncodingStage,
+)
+from sglang.multimodal_gen.runtime.server_args import ServerArgs
+from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+
+logger = init_logger(__name__)
+
+
+class SD3ConditioningStage(PipelineStage):
+    """Merge CLIP-T, CLIP-G and T5 embeddings into unified prompt/pooled tensors."""
+
+    @torch.no_grad()
+    def forward(self, batch: Req, server_args: ServerArgs) -> Req:
+        batch.prompt_embeds, batch.pooled_embeds = self._merge(
+            batch.prompt_embeds, batch.pooled_embeds
+        )
+        if batch.do_classifier_free_guidance:
+            batch.negative_prompt_embeds, batch.neg_pooled_embeds = self._merge(
+                batch.negative_prompt_embeds, batch.neg_pooled_embeds
+            )
+        return batch
+
+    @staticmethod
+    def _merge(
+        embeds_list: list[torch.Tensor],
+        pooled_list: list[torch.Tensor],
+    ) -> tuple[list[torch.Tensor], list[torch.Tensor]]:
+        """Merge 3 encoder outputs into unified prompt/pooled tensors.
+
+        SD3-medium uses exactly 3 text encoders (CLIP-L, CLIP-G, T5).
+        Returns single-element lists to match the batch field format expected
+        by downstream stages (get_pos_prompt_embeds accesses index [0]).
+        """
+        if len(embeds_list) != 3:
+            raise ValueError(
+                f"SD3 requires exactly 3 prompt embedding tensors, got {len(embeds_list)}."
+            )
+        if len(pooled_list) < 2:
+            raise ValueError(
+                f"SD3 requires at least 2 pooled embedding tensors, got {len(pooled_list)}."
+            )
+
+        clipt, clipg, t5 = embeds_list
+        clip_merged = torch.cat([clipt, clipg], dim=-1)
+        clip_merged = torch.nn.functional.pad(
+            clip_merged, (0, t5.shape[-1] - clip_merged.shape[-1])
+        )
+        merged_embeds = [torch.cat([clip_merged, t5], dim=-2)]
+        merged_pooled = [torch.cat([pooled_list[0], pooled_list[1]], dim=-1)]
+        return merged_embeds, merged_pooled
+
+
+class StableDiffusion3Pipeline(ComposedPipelineBase):
+    """StableDiffusion3 pipeline implementation."""
+
+    pipeline_name = "StableDiffusion3Pipeline"
+
+    _required_config_modules = [
+        "text_encoder",
+        "text_encoder_2",
+        "text_encoder_3",
+        "tokenizer",
+        "tokenizer_2",
+        "tokenizer_3",
+        "vae",
+        "transformer",
+        "scheduler",
+    ]
+
+    def create_pipeline_stages(self, server_args: ServerArgs):
+        self.add_stage(InputValidationStage())
+
+        self.add_stage(
+            TextEncodingStage(
+                text_encoders=[
+                    self.get_module("text_encoder"),
+                    self.get_module("text_encoder_2"),
+                    self.get_module("text_encoder_3"),
+                ],
+                tokenizers=[
+                    self.get_module("tokenizer"),
+                    self.get_module("tokenizer_2"),
+                    self.get_module("tokenizer_3"),
+                ],
+            ),
+            "prompt_encoding_stage_primary",
+        )
+
+        self.add_stage(SD3ConditioningStage())
+
+        self.add_standard_timestep_preparation_stage()
+        self.add_standard_latent_preparation_stage()
+        self.add_standard_denoising_stage()
+        self.add_standard_decoding_stage()
+
+
+EntryClass = StableDiffusion3Pipeline
@@ -12,8 +12,6 @@ import inspect
 import torch

 from sglang.multimodal_gen.configs.models.encoders import BaseEncoderOutput
-from sglang.multimodal_gen.configs.pipeline_configs import FluxPipelineConfig
-from sglang.multimodal_gen.configs.pipeline_configs.flux import Flux2PipelineConfig
 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.pipelines_core.schedule_batch import Req
@@ -263,11 +261,11 @@ class TextEncodingStage(PipelineStage):
            ).to(target_device)

            input_ids = text_inputs["input_ids"]
-            is_flux_v1 = isinstance(
-                server_args.pipeline_config, FluxPipelineConfig
-            ) and not isinstance(server_args.pipeline_config, Flux2PipelineConfig)
-
-            attention_mask = None if is_flux_v1 else text_inputs["attention_mask"]
+            attention_mask = (
+                server_args.pipeline_config.get_text_encoder_attention_mask(
+                    text_inputs, i
+                )
+            )
            encoder_forward_kwargs = {
                "input_ids": input_ids,
                "output_hidden_states": True,
@@ -292,12 +290,13 @@ class TextEncodingStage(PipelineStage):
                prompt_embeds = prompt_embeds.to(device=target_device)

            embeds_list.append(prompt_embeds)
-            if is_flux_v1 and outputs.pooler_output is not None:
-                # FLUX.1 only consumes the pooled CLIP projection. The T5
-                # encoder in the same pipeline has no pooler output.
-                pooled_embeds_list.append(
-                    outputs.pooler_output.to(device=target_device)
-                )
+
+            pooled_output = server_args.pipeline_config.get_text_encoder_pooler_output(
+                outputs, i
+            )
+            if pooled_output is not None:
+                pooled_embeds_list.append(pooled_output.to(device=target_device))
+
            if return_attention_mask:
                mask_to_store = (
                    attention_mask.to(device=target_device)