[diffusion] chore: use native qwen2.5-vl generation (#34896)

docs/docs/sglang-diffusion/compatibility_matrix.mdx

@@ -37,6 +37,10 @@ Rows are grouped when a family shares the same runtime path or optimization supp
             <td>Qwen-Image</td>
             <td><div className="sgd-id-list"><code>Qwen/Qwen-Image</code><code>Qwen/Qwen-Image-2512</code><code>Qwen/Qwen-Image-Edit</code><code>Qwen/Qwen-Image-Edit-2509</code><code>Qwen/Qwen-Image-Edit-2511</code><code>Qwen/Qwen-Image-Layered</code></div></td>
           </tr>
+          <tr>
+            <td>LongCat-Image</td>
+            <td><div className="sgd-id-list"><code>meituan-longcat/LongCat-Image</code></div></td>
+          </tr>
           <tr>
             <td>SD3 / SD3.5</td>
             <td><div className="sgd-id-list"><code>stabilityai/stable-diffusion-3-medium</code><code>stabilityai/stable-diffusion-3-medium-diffusers</code><code>stabilityai/stable-diffusion-3.5-medium</code><code>stabilityai/stable-diffusion-3.5-medium-diffusers</code><code>stabilityai/stable-diffusion-3.5-large</code><code>stabilityai/stable-diffusion-3.5-large-diffusers</code></div></td>

docs/docs/sglang-diffusion/models_with_ar.mdx

@@ -1,11 +1,16 @@
 ---
-title: "Diffusion models with AR stage like GLM-Image"
+title: "Diffusion models with autoregressive stages"
-description: "Run diffusion pipelines that delegate an AR stage to a separate SGLang server, such as GLM-Image."
+description: "Run diffusion pipelines with in-process or separately deployed autoregressive encoders."
 ---
 
-## Quick Start
+SGLang Diffusion supports two AR execution paths. Qwen Image Layered and
+LongCat-Image use the native Qwen2.5-VL component in process. GLM-Image can use
+its bundled Transformers implementation or delegate AR inference to a separate
+SGLang server.
 
-Run model with transformers implementation for AR stage (default)
+## GLM-Image Quick Start
+
+Run GLM-Image with its bundled Transformers implementation (default):
 ```bash
 # Terminal 1 : launch server
 sglang serve --model-path zai-org/GLM-Image --port ${PORT}
@@ -20,7 +25,7 @@ curl http://${HOST}:${PORT}/v1/images/generations \
     "size": "widthxheight"
   }'
 ```
-Run model with SGLang srt implementation for AR stage (high performance)
+Run GLM-Image with a separate SGLang server for its AR stage:
 ```bash
 # Terminal 1 : launch server with AR model
 sglang serve --model-path /path/to/zai-org/GLM-Image/vision_language_encoder/ \
@@ -64,6 +69,16 @@ curl http://${HOST}:${PORT}/v1/images/generations \
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)", whiteSpace: "nowrap"}}>T2I, I2I, V2I</td>
       <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>T2I</td>
     </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>Qwen Image Layered</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Not used</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>I2I, in-process native Qwen2.5-VL</td>
+    </tr>
+    <tr>
+      <td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>LongCat-Image</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Not used</td>
+      <td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>T2I, in-process native Qwen2.5-VL</td>
+    </tr>
 </tbody>
 </table>
 

python/sglang/multimodal_gen/configs/models/encoders/base.py

@@ -86,6 +86,7 @@ class EncoderConfig(ModelConfig):
 @dataclass
 class TextEncoderConfig(EncoderConfig):
     arch_config: ArchConfig = field(default_factory=TextEncoderArchConfig)
+    generation_config: dict[str, Any] = field(default_factory=dict)
 
 
 @dataclass

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

@@ -8,6 +8,7 @@ from sglang.multimodal_gen.configs.models import DiTConfig, EncoderConfig, VAECo
 from sglang.multimodal_gen.configs.models.dits.longcat_image import (
     LongCatImageDitConfig,
 )
+from sglang.multimodal_gen.configs.models.encoders.qwen_image import Qwen2_5VLConfig
 from sglang.multimodal_gen.configs.models.vaes.longcat_image import (
     LongCatImageVAEConfig,
 )
@@ -16,6 +17,9 @@ from sglang.multimodal_gen.configs.pipeline_configs.base import (
     ModelTaskType,
     TextConditioningOutput,
 )
+from sglang.multimodal_gen.configs.pipeline_configs.model_deployment_config import (
+    ModelDeploymentConfig,
+)
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
 
 logger = init_logger(__name__)
@@ -219,18 +223,6 @@ def longcat_postprocess_text(outputs, text_inputs, pipeline_config):
     )
 
 
-@dataclass
-class LongCatImageEncoderConfig(EncoderConfig):
-    """Encoder config for the in-stage-loaded HF Qwen2.5-VL text encoder.
-
-    The encoder weights are loaded by `LongCatPromptRewriteStage` (not via
-    `TextEncoderLoader`), so this config only supplies the fields the standard
-    `TextEncodingStage` reads — primarily `tokenizer_kwargs`.
-    """
-
-    tokenizer_kwargs: dict = field(default_factory=lambda: {})
-
-
 @dataclass
 class LongCatImagePipelineConfig(ImagePipelineConfig):
     """Configuration for the LongCat-Image T2I pipeline."""
@@ -246,16 +238,20 @@ class LongCatImagePipelineConfig(ImagePipelineConfig):
     dit_config: DiTConfig = field(default_factory=LongCatImageDitConfig)
     vae_config: VAEConfig = field(default_factory=LongCatImageVAEConfig)
 
-    # The Qwen2.5-VL text encoder (~7B) is loaded in bf16; the encoder is loaded
-    # in-stage by LongCatPromptRewriteStage, not via TextEncoderLoader.
     text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16",))
     text_encoder_configs: tuple[EncoderConfig, ...] = field(
-        default_factory=lambda: (LongCatImageEncoderConfig(),)
+        default_factory=lambda: (Qwen2_5VLConfig(),)
     )
     postprocess_text_funcs: tuple[Callable, ...] = field(
         default_factory=lambda: (longcat_postprocess_text,)
     )
 
+    def get_model_deployment_config(self) -> ModelDeploymentConfig:
+        return ModelDeploymentConfig(
+            keep_resident_min_available_gb=70,
+            keep_resident_components=("text_encoder", "vae"),
+        )
+
     # --- LatentPreparationStage hooks ---
 
     def prepare_latent_shape(self, batch, batch_size, num_frames):

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

@@ -19,6 +19,9 @@ from sglang.multimodal_gen.configs.pipeline_configs.base import (
     pad_text_embeddings_with_mask,
     shard_rotary_emb_for_sp,
 )
+from sglang.multimodal_gen.configs.pipeline_configs.model_deployment_config import (
+    ModelDeploymentConfig,
+)
 from sglang.multimodal_gen.configs.post_training.pipeline_configs import (
     QwenImageRolloutPipelineMixin,
 )
@@ -753,6 +756,12 @@ class QwenImageLayeredPipelineConfig(QwenImageEditPipelineConfig):
     resolution: int = 640
     vae_precision: str = "bf16"
 
+    def get_model_deployment_config(self) -> ModelDeploymentConfig:
+        return ModelDeploymentConfig(
+            keep_resident_min_available_gb=70,
+            keep_resident_components=("text_encoder", "vae"),
+        )
+
     def postprocess_cfg_noise(
         self,
         batch,

python/sglang/multimodal_gen/runtime/loader/component_loaders/text_encoder_loader.py

@@ -52,6 +52,7 @@ from sglang.multimodal_gen.runtime.server_args import ServerArgs
 from sglang.multimodal_gen.runtime.utils.hf_diffusers_utils import (
     get_config,
     get_diffusers_component_config,
+    load_dict,
 )
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
 from sglang.multimodal_gen.runtime.utils.precision import precision_to_dtype
@@ -344,6 +345,9 @@ class TextEncoderLoader(ComponentLoader):
 
         encoder_config = server_args.pipeline_config.text_encoder_configs[encoder_index]
         encoder_config.update_model_arch(model_config)
+        encoder_config.generation_config = load_dict(
+            os.path.join(component_model_path, "generation_config.json")
+        )
 
         if encoder_index == 0:
             for key, value in diffusers_pretrained_config.__dict__.items():

python/sglang/multimodal_gen/runtime/models/encoders/qwen2_5vl.py

@@ -29,6 +29,9 @@ from sglang.multimodal_gen.runtime.layers.linear import (
 from sglang.multimodal_gen.runtime.layers.quantization import QuantizationConfig
 from sglang.multimodal_gen.runtime.loader.weight_utils import default_weight_loader
 from sglang.multimodal_gen.runtime.models.encoders.base import TextEncoder
+from sglang.multimodal_gen.runtime.models.encoders.qwen2_5vl_vision import (
+    Qwen2_5VLVisionTransformer,
+)
 from sglang.multimodal_gen.runtime.platforms import AttentionBackendEnum
 from sglang.multimodal_gen.runtime.utils.common import add_prefix
 
@@ -69,8 +72,6 @@ import torch
 import torch.nn as nn
 from transformers.activations import ACT2FN
 from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
-    Qwen2_5_VisionRotaryEmbedding,
-    Qwen2_5_VisionTransformerPretrainedModel,
     Qwen2_5_VLCausalLMOutputWithPast,
     Qwen2_5_VLModelOutputWithPast,
     Qwen2_5_VLRotaryEmbedding,
@@ -80,6 +81,53 @@ from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
 logger = logging.getLogger(__name__)
 
 
+def _apply_repetition_penalty(
+    logits: torch.Tensor,
+    input_ids: torch.LongTensor,
+    penalty: float,
+) -> torch.Tensor:
+    if penalty == 1.0:
+        return logits
+    selected_logits = torch.gather(logits, 1, input_ids)
+    selected_logits = torch.where(
+        selected_logits < 0,
+        selected_logits * penalty,
+        selected_logits / penalty,
+    )
+    return logits.scatter(1, input_ids, selected_logits)
+
+
+def _select_next_token(
+    logits: torch.Tensor,
+    input_ids: torch.LongTensor,
+    *,
+    do_sample: bool,
+    temperature: float,
+    top_k: int,
+    top_p: float,
+    repetition_penalty: float,
+) -> torch.LongTensor:
+    scores = _apply_repetition_penalty(logits.float(), input_ids, repetition_penalty)
+    if not do_sample or top_k == 1:
+        return torch.argmax(scores, dim=-1)
+
+    scores = scores / temperature
+    if 0 < top_k < scores.shape[-1]:
+        top_k_threshold = torch.topk(scores, top_k, dim=-1).values[..., -1, None]
+        scores = scores.masked_fill(scores < top_k_threshold, -torch.inf)
+    if top_p < 1.0:
+        sorted_scores, sorted_indices = torch.sort(scores, descending=True)
+        cumulative_probs = torch.softmax(sorted_scores, dim=-1).cumsum(dim=-1)
+        sorted_remove = cumulative_probs > top_p
+        sorted_remove[..., 1:] = sorted_remove[..., :-1].clone()
+        sorted_remove[..., 0] = False
+        remove = torch.zeros_like(sorted_remove).scatter(
+            1, sorted_indices, sorted_remove
+        )
+        scores = scores.masked_fill(remove, -torch.inf)
+    return torch.multinomial(torch.softmax(scores, dim=-1), num_samples=1).squeeze(1)
+
+
 def _tp_world_size() -> int:
     if not model_parallel_is_initialized():
         return 1
@@ -261,7 +309,15 @@ class Qwen2_5_VLAttention(nn.Module):
         query_states = query_states.transpose(1, 2)
         key_states = key_states.transpose(1, 2)
         value_states = value_states.transpose(1, 2)
-        attn_output = self.attn(query_states, key_states, value_states)
+        # Diffusion text encoding is cache-free and historically uses the native
+        # causal kernel; its trailing padding is removed during postprocessing.
+        # Cached generation still needs the explicit mask for padded batches.
+        attn_output = self.attn(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask if use_cache else None,
+        )
 
         attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
         attn_output = _linear_output(self.o_proj, attn_output)
@@ -597,28 +653,14 @@ class Qwen2_5_VLModel(nn.Module):
     _checkpoint_conversion_mapping = {"^model": "language_model"}
     # Reference: fix gemma3 grad acc #37208
     accepts_loss_kwargs = False
-    _no_split_modules = ["Qwen2_5_VLDecoderLayer", "Qwen2_5_VLVisionBlock"]
+    _no_split_modules = ["Qwen2_5_VLDecoderLayer", "Qwen2_5VLVisionBlock"]
 
     def __init__(self, config, enable_image_understanding: bool = False):
         super().__init__()
         self.language_model = Qwen2_5_VLTextModel(config.text_config)
 
         if enable_image_understanding:
-            self.visual = Qwen2_5_VisionTransformerPretrainedModel._from_config(
+            self.visual = Qwen2_5VLVisionTransformer(config.vision_config)
-                config.vision_config
-            )
-            self.visual.to(torch.get_default_dtype())
-            # keeps the vision rotary frequencies in fp32 even when weights are bf16 (as HF does)
-            head_dim = (
-                config.vision_config.hidden_size // config.vision_config.num_heads
-            )
-            rotary_dim = head_dim // 2
-            inv_freq = Qwen2_5_VisionRotaryEmbedding(rotary_dim).inv_freq
-            self.visual.rotary_pos_emb.register_buffer(
-                "inv_freq",
-                inv_freq,
-                persistent=False,
-            )
         self.rope_deltas = None  # cache rope_deltas here
         self.config = config
         # Initialize weights and apply final processing
@@ -902,11 +944,6 @@ class Qwen2_5_VLModel(nn.Module):
         """
         pixel_values = pixel_values.type(self.visual.dtype)
         image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
-        if not isinstance(image_embeds, torch.Tensor):
-            # In transformers v5, the visual encoder returns BaseModelOutputWithPooling.
-            # pooler_output contains the spatially merged embeddings (what we need),
-            # while last_hidden_state contains the raw unmerged output.
-            image_embeds = image_embeds.pooler_output
         split_sizes = (
             image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2
         ).tolist()
@@ -1106,6 +1143,9 @@ class Qwen2_5_VLModel(nn.Module):
 
 
 class Qwen2_5_VLForConditionalGeneration(TextEncoder):
+    layer_names = [*TextEncoder.layer_names, "model.visual.blocks"]
+    _fsdp_forward_methods = ("generate",)
+
     # BitandBytes specific attributes
     default_bitsandbytes_target_modules = [
         ".gate_up_proj.",
@@ -1132,6 +1172,7 @@ class Qwen2_5_VLForConditionalGeneration(TextEncoder):
     ) -> None:
         super().__init__(config)
         enable_image_understanding = config.enable_image_understanding
+        generation_config = config.generation_config
         config = config.arch_config
         self.model = Qwen2_5_VLModel(
             config, enable_image_understanding=enable_image_understanding
@@ -1141,6 +1182,7 @@ class Qwen2_5_VLForConditionalGeneration(TextEncoder):
         )
 
         self.enable_image_understanding = enable_image_understanding
+        self.generation_config = generation_config
 
         self.config = config
 
@@ -1225,6 +1267,152 @@ class Qwen2_5_VLForConditionalGeneration(TextEncoder):
             rope_deltas=outputs.rope_deltas,
         )
 
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        *,
+        max_new_tokens: int,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        second_per_grid_ts: Optional[torch.Tensor] = None,
+        mm_token_type_ids: Optional[torch.IntTensor] = None,
+        do_sample: Optional[bool] = None,
+        temperature: Optional[float] = None,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+        repetition_penalty: Optional[float] = None,
+        eos_token_id: Optional[Union[int, list[int]]] = None,
+        pad_token_id: Optional[int] = None,
+    ) -> torch.LongTensor:
+        """Generate tokens with Qwen2.5-VL's native decoder and KV cache."""
+        # Transformers 5 processors emit this field. The Qwen2.5-VL checkpoint
+        # derives modality positions from image/video placeholder token IDs.
+        del mm_token_type_ids
+
+        if input_ids.ndim != 2:
+            raise ValueError("input_ids must have shape [batch_size, sequence_length]")
+        if max_new_tokens < 0:
+            raise ValueError("max_new_tokens must be non-negative")
+        if max_new_tokens == 0:
+            return input_ids
+
+        generation_config = self.generation_config
+        do_sample = (
+            generation_config.get("do_sample", False)
+            if do_sample is None
+            else do_sample
+        )
+        temperature = (
+            generation_config.get("temperature", 1.0)
+            if temperature is None
+            else temperature
+        )
+        top_k = generation_config.get("top_k", 0) if top_k is None else top_k
+        top_p = generation_config.get("top_p", 1.0) if top_p is None else top_p
+        repetition_penalty = (
+            generation_config.get("repetition_penalty", 1.0)
+            if repetition_penalty is None
+            else repetition_penalty
+        )
+        eos_token_id = (
+            generation_config.get("eos_token_id", self.config.eos_token_id)
+            if eos_token_id is None
+            else eos_token_id
+        )
+        pad_token_id = (
+            generation_config.get("pad_token_id", self.config.pad_token_id)
+            if pad_token_id is None
+            else pad_token_id
+        )
+
+        if do_sample and temperature <= 0:
+            raise ValueError("temperature must be positive when sampling")
+        if top_k < 0:
+            raise ValueError("top_k must be non-negative")
+        if not 0 < top_p <= 1:
+            raise ValueError("top_p must be in (0, 1]")
+        if repetition_penalty <= 0:
+            raise ValueError("repetition_penalty must be positive")
+
+        eos_token_ids = (
+            []
+            if eos_token_id is None
+            else [eos_token_id] if isinstance(eos_token_id, int) else list(eos_token_id)
+        )
+        if pad_token_id is None:
+            raise ValueError("pad_token_id must be set for generation")
+        eos_tokens = torch.tensor(
+            eos_token_ids, dtype=input_ids.dtype, device=input_ids.device
+        )
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+
+        generated_ids = input_ids
+        unfinished = torch.ones(
+            input_ids.shape[0], dtype=torch.bool, device=input_ids.device
+        )
+        past_key_values = None
+        model_input_ids = input_ids
+        cache_position = torch.arange(input_ids.shape[1], device=input_ids.device)
+        self.model.rope_deltas = None
+
+        for _ in range(max_new_tokens):
+            outputs = self(
+                input_ids=model_input_ids,
+                attention_mask=attention_mask,
+                past_key_values=past_key_values,
+                use_cache=True,
+                cache_position=cache_position,
+                pixel_values=pixel_values,
+                pixel_values_videos=pixel_values_videos,
+                image_grid_thw=image_grid_thw,
+                video_grid_thw=video_grid_thw,
+                second_per_grid_ts=second_per_grid_ts,
+                logits_to_keep=1,
+            )
+            next_tokens = _select_next_token(
+                outputs.logits[:, -1, :],
+                generated_ids,
+                do_sample=do_sample,
+                temperature=temperature,
+                top_k=top_k,
+                top_p=top_p,
+                repetition_penalty=repetition_penalty,
+            )
+            next_tokens = torch.where(
+                unfinished,
+                next_tokens,
+                torch.full_like(next_tokens, pad_token_id),
+            )
+            generated_ids = torch.cat([generated_ids, next_tokens[:, None]], dim=-1)
+
+            if eos_tokens.numel() > 0:
+                reached_eos = (next_tokens[:, None] == eos_tokens[None, :]).any(dim=-1)
+                unfinished = unfinished & ~reached_eos
+                if not unfinished.any():
+                    break
+
+            past_key_values = outputs.past_key_values
+            model_input_ids = next_tokens[:, None]
+            attention_mask = torch.cat(
+                [attention_mask, attention_mask.new_ones((input_ids.shape[0], 1))],
+                dim=-1,
+            )
+            cache_position = torch.tensor(
+                [generated_ids.shape[1] - 1], device=input_ids.device
+            )
+            pixel_values = None
+            pixel_values_videos = None
+            image_grid_thw = None
+            video_grid_thw = None
+            second_per_grid_ts = None
+
+        return generated_ids
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         loaded_params: set[str] = set()
 

python/sglang/multimodal_gen/runtime/models/encoders/qwen2_5vl_vision.py

@@ -0,0 +1,461 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Native Qwen2.5-VL vision encoder."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from sglang.multimodal_gen.runtime.layers.attention.selector import get_attn_backend
+from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
+
+logger = init_logger(__name__)
+
+
+@dataclass(frozen=True)
+class _PackedSequenceMetadata:
+    cu_seqlens: torch.Tensor
+    cu_seqlens_host: tuple[int, ...]
+    max_seqlen: int
+
+    @classmethod
+    def from_cu_seqlens(cls, cu_seqlens: torch.Tensor) -> _PackedSequenceMetadata:
+        bounds = tuple(int(value) for value in cu_seqlens.tolist())
+        return cls(
+            cu_seqlens=cu_seqlens,
+            cu_seqlens_host=bounds,
+            max_seqlen=max(
+                stop - start for start, stop in zip(bounds[:-1], bounds[1:])
+            ),
+        )
+
+
+class Qwen2_5VLVisionRMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.float()
+        variance = hidden_states.square().mean(dim=-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class Qwen2_5VLVisionPatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size: int,
+        temporal_patch_size: int,
+        in_channels: int,
+        embed_dim: int,
+    ) -> None:
+        super().__init__()
+        self.patch_size = patch_size
+        self.temporal_patch_size = temporal_patch_size
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+        kernel_size = (temporal_patch_size, patch_size, patch_size)
+        self.proj = nn.Conv3d(
+            in_channels,
+            embed_dim,
+            kernel_size=kernel_size,
+            stride=kernel_size,
+            bias=False,
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = hidden_states.view(
+            -1,
+            self.in_channels,
+            self.temporal_patch_size,
+            self.patch_size,
+            self.patch_size,
+        )
+        return self.proj(hidden_states.to(self.proj.weight.dtype)).view(
+            -1, self.embed_dim
+        )
+
+
+class Qwen2_5VLVisionRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, position_ids: torch.Tensor) -> torch.Tensor:
+        return (position_ids.unsqueeze(-1) * self.inv_freq).flatten(1)
+
+
+def _rotate_half(hidden_states: torch.Tensor) -> torch.Tensor:
+    first, second = hidden_states.chunk(2, dim=-1)
+    return torch.cat((-second, first), dim=-1)
+
+
+def _apply_vision_rotary_embedding(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    query_dtype = query.dtype
+    key_dtype = key.dtype
+    query = query.float()
+    key = key.float()
+    cos = cos.unsqueeze(-2).float()
+    sin = sin.unsqueeze(-2).float()
+    query = query * cos + _rotate_half(query) * sin
+    key = key * cos + _rotate_half(key) * sin
+    return query.to(query_dtype), key.to(key_dtype)
+
+
+class Qwen2_5VLVisionAttention(nn.Module):
+    def __init__(self, config: Any, prefix: str) -> None:
+        super().__init__()
+        self.num_heads = config.num_heads
+        self.head_dim = config.hidden_size // config.num_heads
+        self.scaling = self.head_dim**-0.5
+        self.prefix = prefix
+        self.qkv = nn.Linear(config.hidden_size, config.hidden_size * 3, bias=True)
+        self.proj = nn.Linear(config.hidden_size, config.hidden_size)
+        self._attention_impl = None
+        self._initialize_attention(torch.get_default_dtype())
+
+    def _initialize_attention(self, dtype: torch.dtype) -> None:
+        backend = get_attn_backend(self.head_dim, dtype)
+        if backend.supports_packed_varlen():
+            self._attention_impl = backend.get_impl_cls()(
+                num_heads=self.num_heads,
+                head_size=self.head_dim,
+                num_kv_heads=self.num_heads,
+                softmax_scale=self.scaling,
+                causal=False,
+                prefix=self.prefix,
+            )
+        else:
+            logger.warning_once(
+                "Qwen2.5-VL vision attention uses torch SDPA because "
+                f"{backend.get_enum().name.lower()} does not support packed sequences"
+            )
+
+    def _packed_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        cu_seqlens_host: tuple[int, ...],
+        max_seqlen: int,
+    ) -> torch.Tensor:
+        if self._attention_impl is not None:
+            return self._attention_impl.forward_varlen(
+                query,
+                key,
+                value,
+                cu_seqlens=cu_seqlens,
+                cu_seqlens_host=cu_seqlens_host,
+                max_seqlen=max_seqlen,
+            )
+
+        output = torch.empty_like(query)
+        for start, stop in zip(cu_seqlens_host[:-1], cu_seqlens_host[1:]):
+            if start == stop:
+                continue
+            query_segment = query[start:stop].transpose(0, 1).unsqueeze(0)
+            key_segment = key[start:stop].transpose(0, 1).unsqueeze(0)
+            value_segment = value[start:stop].transpose(0, 1).unsqueeze(0)
+            segment = F.scaled_dot_product_attention(
+                query_segment,
+                key_segment,
+                value_segment,
+                dropout_p=0.0,
+                is_causal=False,
+                scale=self.scaling,
+            )
+            output[start:stop] = segment.squeeze(0).transpose(0, 1)
+        return output
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        *,
+        cu_seqlens: torch.Tensor,
+        cu_seqlens_host: tuple[int, ...],
+        max_seqlen: int,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+    ) -> torch.Tensor:
+        seq_len = hidden_states.shape[0]
+        query, key, value = (
+            self.qkv(hidden_states)
+            .reshape(seq_len, 3, self.num_heads, self.head_dim)
+            .permute(1, 0, 2, 3)
+            .unbind(0)
+        )
+        query, key = _apply_vision_rotary_embedding(query, key, *position_embeddings)
+        output = self._packed_attention(
+            query,
+            key,
+            value,
+            cu_seqlens,
+            cu_seqlens_host,
+            max_seqlen,
+        )
+        return self.proj(output.reshape(seq_len, -1).contiguous())
+
+
+class Qwen2_5VLVisionMLP(nn.Module):
+    def __init__(self, config: Any) -> None:
+        super().__init__()
+        if config.hidden_act != "silu":
+            raise ValueError(
+                f"Unsupported Qwen2.5-VL vision activation: {config.hidden_act}"
+            )
+        self.gate_proj = nn.Linear(
+            config.hidden_size, config.intermediate_size, bias=True
+        )
+        self.up_proj = nn.Linear(
+            config.hidden_size, config.intermediate_size, bias=True
+        )
+        self.down_proj = nn.Linear(
+            config.intermediate_size, config.hidden_size, bias=True
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(
+            F.silu(self.gate_proj(hidden_states)) * self.up_proj(hidden_states)
+        )
+
+
+class Qwen2_5VLVisionBlock(nn.Module):
+    def __init__(self, config: Any, layer_idx: int) -> None:
+        super().__init__()
+        self.norm1 = Qwen2_5VLVisionRMSNorm(config.hidden_size)
+        self.norm2 = Qwen2_5VLVisionRMSNorm(config.hidden_size)
+        self.attn = Qwen2_5VLVisionAttention(
+            config, prefix=f"visual.blocks.{layer_idx}.attn"
+        )
+        self.mlp = Qwen2_5VLVisionMLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        *,
+        cu_seqlens: torch.Tensor,
+        cu_seqlens_host: tuple[int, ...],
+        max_seqlen: int,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+    ) -> torch.Tensor:
+        hidden_states = hidden_states + self.attn(
+            self.norm1(hidden_states),
+            cu_seqlens=cu_seqlens,
+            cu_seqlens_host=cu_seqlens_host,
+            max_seqlen=max_seqlen,
+            position_embeddings=position_embeddings,
+        )
+        return hidden_states + self.mlp(self.norm2(hidden_states))
+
+
+class Qwen2_5VLVisionPatchMerger(nn.Module):
+    def __init__(
+        self,
+        output_dim: int,
+        context_dim: int,
+        spatial_merge_size: int,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = context_dim * spatial_merge_size**2
+        self.ln_q = Qwen2_5VLVisionRMSNorm(context_dim)
+        self.mlp = nn.Sequential(
+            nn.Linear(self.hidden_size, self.hidden_size),
+            nn.GELU(),
+            nn.Linear(self.hidden_size, output_dim),
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.ln_q(hidden_states).view(-1, self.hidden_size)
+        return self.mlp(hidden_states)
+
+
+def _vision_position_ids(
+    grid_thw: torch.Tensor, spatial_merge_size: int
+) -> torch.Tensor:
+    position_ids = []
+    for t, h, w in grid_thw.tolist():
+        h_positions = torch.arange(h, device=grid_thw.device)[:, None].expand(h, w)
+        h_positions = (
+            h_positions.reshape(
+                h // spatial_merge_size,
+                spatial_merge_size,
+                w // spatial_merge_size,
+                spatial_merge_size,
+            )
+            .transpose(1, 2)
+            .flatten()
+        )
+        w_positions = torch.arange(w, device=grid_thw.device)[None, :].expand(h, w)
+        w_positions = (
+            w_positions.reshape(
+                h // spatial_merge_size,
+                spatial_merge_size,
+                w // spatial_merge_size,
+                spatial_merge_size,
+            )
+            .transpose(1, 2)
+            .flatten()
+        )
+        positions = torch.stack((h_positions, w_positions), dim=-1)
+        position_ids.append(positions.repeat(t, 1))
+    return torch.cat(position_ids, dim=0)
+
+
+def _vision_window_index(
+    grid_thw: torch.Tensor,
+    *,
+    spatial_merge_size: int,
+    window_size: int,
+    patch_size: int,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    window_indices = []
+    cumulative_window_lengths = [0]
+    window_index_offset = 0
+    merger_window_size = window_size // spatial_merge_size // patch_size
+    spatial_merge_unit = spatial_merge_size**2
+
+    for grid_t, grid_h, grid_w in grid_thw.tolist():
+        merged_height = grid_h // spatial_merge_size
+        merged_width = grid_w // spatial_merge_size
+        index = torch.arange(
+            grid_t * merged_height * merged_width, device=grid_thw.device
+        ).reshape(grid_t, merged_height, merged_width)
+        pad_height = merger_window_size - merged_height % merger_window_size
+        pad_width = merger_window_size - merged_width % merger_window_size
+        num_windows_height = (merged_height + pad_height) // merger_window_size
+        num_windows_width = (merged_width + pad_width) // merger_window_size
+        index = F.pad(index, (0, pad_width, 0, pad_height), value=-100)
+        index = index.reshape(
+            grid_t,
+            num_windows_height,
+            merger_window_size,
+            num_windows_width,
+            merger_window_size,
+        )
+        index = index.permute(0, 1, 3, 2, 4).reshape(
+            grid_t,
+            num_windows_height * num_windows_width,
+            merger_window_size,
+            merger_window_size,
+        )
+        sequence_lengths = (index != -100).sum(dim=(2, 3)).reshape(-1)
+        index = index.flatten()
+        window_indices.append(index[index != -100] + window_index_offset)
+        cumulative = (
+            sequence_lengths.cumsum(0) * spatial_merge_unit
+            + cumulative_window_lengths[-1]
+        )
+        cumulative_window_lengths.extend(cumulative.tolist())
+        window_index_offset += grid_t * merged_height * merged_width
+
+    window_index = torch.cat(window_indices)
+    cu_window_seqlens = torch.tensor(
+        cumulative_window_lengths, device=grid_thw.device, dtype=torch.int32
+    )
+    return window_index, torch.unique_consecutive(cu_window_seqlens)
+
+
+class Qwen2_5VLVisionTransformer(nn.Module):
+    def __init__(self, config: Any) -> None:
+        super().__init__()
+        self.spatial_merge_size = config.spatial_merge_size
+        self.spatial_merge_unit = config.spatial_merge_size**2
+        self.patch_size = config.patch_size
+        self.window_size = config.window_size
+        self.full_attention_layers = frozenset(
+            int(layer_idx) for layer_idx in config.fullatt_block_indexes
+        )
+        self.patch_embed = Qwen2_5VLVisionPatchEmbed(
+            patch_size=config.patch_size,
+            temporal_patch_size=config.temporal_patch_size,
+            in_channels=config.in_channels,
+            embed_dim=config.hidden_size,
+        )
+        head_dim = config.hidden_size // config.num_heads
+        self.rotary_pos_emb = Qwen2_5VLVisionRotaryEmbedding(head_dim // 2)
+        self.blocks = nn.ModuleList(
+            Qwen2_5VLVisionBlock(config, layer_idx) for layer_idx in range(config.depth)
+        )
+        self.merger = Qwen2_5VLVisionPatchMerger(
+            output_dim=config.out_hidden_size,
+            context_dim=config.hidden_size,
+            spatial_merge_size=config.spatial_merge_size,
+        )
+
+    @property
+    def dtype(self) -> torch.dtype:
+        return self.patch_embed.proj.weight.dtype
+
+    @property
+    def device(self) -> torch.device:
+        return self.patch_embed.proj.weight.device
+
+    def forward(
+        self, hidden_states: torch.Tensor, grid_thw: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = hidden_states.to(device=self.device, dtype=self.dtype)
+        grid_thw = grid_thw.to(self.device)
+        hidden_states = self.patch_embed(hidden_states)
+
+        position_ids = _vision_position_ids(grid_thw, self.spatial_merge_size)
+        window_index, cu_window_seqlens = _vision_window_index(
+            grid_thw,
+            spatial_merge_size=self.spatial_merge_size,
+            window_size=self.window_size,
+            patch_size=self.patch_size,
+        )
+
+        seq_len = hidden_states.shape[0]
+        hidden_states = hidden_states.reshape(
+            seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1
+        )[window_index]
+        hidden_states = hidden_states.reshape(seq_len, -1)
+
+        rotary = self.rotary_pos_emb(position_ids)
+        rotary = rotary.reshape(
+            seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1
+        )[window_index]
+        rotary = rotary.reshape(seq_len, -1)
+        rotary = torch.cat((rotary, rotary), dim=-1)
+        position_embeddings = (
+            rotary.cos(),
+            rotary.sin(),
+        )
+
+        cu_seqlens = torch.repeat_interleave(
+            grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
+        ).cumsum(dim=0, dtype=torch.int32)
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        full_metadata = _PackedSequenceMetadata.from_cu_seqlens(cu_seqlens)
+        window_metadata = _PackedSequenceMetadata.from_cu_seqlens(cu_window_seqlens)
+
+        for layer_idx, block in enumerate(self.blocks):
+            metadata = (
+                full_metadata
+                if layer_idx in self.full_attention_layers
+                else window_metadata
+            )
+            hidden_states = block(
+                hidden_states,
+                cu_seqlens=metadata.cu_seqlens,
+                cu_seqlens_host=metadata.cu_seqlens_host,
+                max_seqlen=metadata.max_seqlen,
+                position_embeddings=position_embeddings,
+            )
+
+        merged = self.merger(hidden_states)
+        return merged[torch.argsort(window_index)]

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

@@ -27,11 +27,8 @@ class LongCatImagePipeline(LoRAPipeline, ComposedPipelineBase):
 
     pipeline_name = "LongCatImagePipeline"
 
-    # The Qwen2.5-VL text encoder is loaded in-stage by LongCatPromptRewriteStage
-    # (not via TextEncoderLoader), so "text_encoder" is intentionally absent;
-    # the stage registers the loaded module via add_module("text_encoder", ...)
-    # so the standard TextEncodingStage can fetch the same instance.
     _required_config_modules = [
+        "text_encoder",
         "tokenizer",
         "text_processor",
         "vae",
@@ -41,23 +38,17 @@ class LongCatImagePipeline(LoRAPipeline, ComposedPipelineBase):
 
     def create_pipeline_stages(self, server_args: ServerArgs):
         # 1. Prompt rewriting (optional) + request-level setup (generator, cfg renorm).
-        #    Loads the HF Qwen2.5-VL encoder and shares it with TextEncodingStage.
         rewrite_stage = LongCatPromptRewriteStage(
-            tokenizer=self.get_module("tokenizer"),
+            text_encoder=self.get_module("text_encoder"),
             text_processor=self.get_module("text_processor"),
-            model_path=self.model_path,
             text_encoder_dtype=PRECISION_TO_TYPE[
                 server_args.pipeline_config.text_encoder_precisions[0]
             ],
         )
         self.add_stage(rewrite_stage)
-        self.add_module("text_encoder", rewrite_stage.text_encoder)
 
         # 2. Text encoding via the standard stage (tokenize_prompt +
-        #    postprocess_text_funcs hooks on the pipeline config). Shares the
+        #    postprocess_text_funcs hooks on the pipeline config).
-        #    encoder instance registered above; both stages declare a
-        #    "text_encoder" ComponentUse so the residency manager keeps it
-        #    resident across rewrite->encode and offloads after the last use.
         self.add_standard_text_encoding_stage()
 
         # 3. Latent preparation (batch-size-aware via pipeline config hooks)

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

@@ -95,6 +95,7 @@ class QwenImageLayeredPipeline(QwenImageEditPipeline):
     pipeline_name = "QwenImageLayeredPipeline"
 
     _required_config_modules = [
+        "text_encoder",
         "vae",
         "tokenizer",
         "processor",
@@ -106,12 +107,11 @@ class QwenImageLayeredPipeline(QwenImageEditPipeline):
         def create_before_denoising_stage():
             return QwenImageLayeredBeforeDenoisingStage(
                 vae=self.get_module("vae"),
-                text_encoder=None,
+                text_encoder=self.get_module("text_encoder"),
                 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,
                 vae_dtype=PRECISION_TO_TYPE[server_args.pipeline_config.vae_precision],
                 text_encoder_dtype=PRECISION_TO_TYPE[
                     server_args.pipeline_config.text_encoder_precisions[0]

python/sglang/multimodal_gen/runtime/pipelines_core/stages/image_encoding.py

@@ -322,6 +322,7 @@ class ImageEncodingStage(PipelineStage):
                             pixel_values=image_inputs.pixel_values,
                             image_grid_thw=image_inputs.image_grid_thw,
                             output_hidden_states=True,
+                            use_cache=False,
                         )
                         if batch.do_classifier_free_guidance:
                             neg_outputs = self.text_encoder(
@@ -330,6 +331,7 @@ class ImageEncodingStage(PipelineStage):
                                 pixel_values=neg_image_inputs.pixel_values,
                                 image_grid_thw=neg_image_inputs.image_grid_thw,
                                 output_hidden_states=True,
+                                use_cache=False,
                             )
 
                 prompt_embeds, prompt_embeds_mask, prompt_seq_lens = (

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/longcat_image.py

@@ -1,7 +1,7 @@
 """Prompt-rewriting stage for LongCat-Image (T2I).
 
-`LongCatPromptRewriteStage` optionally rewrites the prompt via the HuggingFace
+`LongCatPromptRewriteStage` optionally rewrites the prompt via the native
-`Qwen2_5_VLForConditionalGeneration.generate()` and sets the CPU generator for
+Qwen2.5-VL encoder and sets the CPU generator for
 seed reproducibility. Text encoding, latent preparation, RoPE and denoising are
 all handled by the standard stages + `LongCatImagePipelineConfig` hooks, so this
 is the only model-specific stage.
@@ -181,15 +181,13 @@ def _get_prompt_language(prompt):
 class LongCatPromptRewriteStage(PipelineStage):
     """Optional prompt rewriting + request-level setup for LongCat-Image.
 
-    Loads the Qwen2.5-VL text encoder (HuggingFace) in-stage and, when
+    When `enable_prompt_rewrite` is set, rewrites the prompt via `.generate()`
-    `enable_prompt_rewrite` is set, rewrites the prompt via `.generate()`
     (using the checkpoint's generation_config.json sampling params). Always
     sets the CPU generator for seed reproducibility. CFG-renorm params are
     read directly from sampling_params in `postprocess_cfg_noise`, not set here.
 
-    The same encoder instance is shared with the standard `TextEncodingStage`
+    The same encoder instance is shared with the standard `TextEncodingStage`,
-    (the pipeline registers it via `add_module("text_encoder", ...)`), so
+    so rewrite and encode run on one set of weights. Both stages declare a
-    rewrite and encode run on one set of weights. Both stages declare a
     `text_encoder` ComponentUse; the residency manager keeps the encoder
     resident across the adjacent rewrite->encode uses and offloads it only
     after the last use (when `--text-encoder-cpu-offload` is enabled).
@@ -197,35 +195,19 @@ class LongCatPromptRewriteStage(PipelineStage):
 
     def __init__(
         self,
-        tokenizer,
+        text_encoder,
         text_processor,
-        model_path: str,
         text_encoder_dtype: torch.dtype,
     ):
         super().__init__()
         self.text_encoder_dtype = text_encoder_dtype
-        from transformers import Qwen2_5_VLForConditionalGeneration
+        self.text_encoder = text_encoder
-
-        cpu_offload = self.server_args.text_encoder_cpu_offload
-        init_device = torch.device("cpu") if cpu_offload else get_local_torch_device()
-        self.text_encoder = (
-            Qwen2_5_VLForConditionalGeneration.from_pretrained(
-                model_path, subfolder="text_encoder"
-            )
-            .to(init_device)
-            .to(dtype=self.text_encoder_dtype)
-        )
-        self.tokenizer = tokenizer
         self.text_processor = text_processor
 
     def component_uses(
         self, server_args: ServerArgs, stage_name: str | None = None
     ) -> list[ComponentUse]:
         stage_name = self._component_stage_name(stage_name)
-        # "text_encoder" matches is_text_encoder_component_name, so
-        # text_encoder_cpu_offload routes through VanillaD2HStrategy (the encoder
-        # is a plain HF nn.Module, not FSDP-sharded). memory_intensive triggers
-        # torch.cuda.empty_cache() after the encoder leaves CUDA.
         return [
             ComponentUse(
                 stage_name,

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/qwen_image_layered.py

@@ -164,7 +164,6 @@ class QwenImageLayeredBeforeDenoisingStage(PipelineStage):
         processor,
         transformer,
         scheduler,
-        model_path,
         vae_dtype: torch.dtype,
         text_encoder_dtype: torch.dtype,
     ) -> None:
@@ -172,15 +171,7 @@ class QwenImageLayeredBeforeDenoisingStage(PipelineStage):
         self.vae = vae.to(dtype=vae_dtype)
         self.vae_dtype = vae_dtype
         self.text_encoder_dtype = text_encoder_dtype
-        if text_encoder is None:
+        self.text_encoder = text_encoder
-            from transformers import Qwen2_5_VLForConditionalGeneration
-
-            text_encoder = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-                model_path, subfolder="text_encoder"
-            )
-        self.text_encoder = text_encoder.to(
-            device=get_local_torch_device(), dtype=self.text_encoder_dtype
-        )
         self.tokenizer = tokenizer
         self.processor = processor
         self.transformer = transformer
@@ -285,11 +276,12 @@ the image\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>as
             padding=True,
             return_tensors="pt",
         ).to(device)
-        encoder_hidden_states = self.text_encoder(
+        with set_forward_context(current_timestep=0, attn_metadata=None):
-            input_ids=txt_tokens.input_ids,
+            encoder_hidden_states = self.text_encoder(
-            attention_mask=txt_tokens.attention_mask,
+                input_ids=txt_tokens.input_ids,
-            output_hidden_states=True,
+                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

python/sglang/multimodal_gen/test/unit/test_disagg_roles.py

@@ -345,9 +345,6 @@ class TestPipelineSpecificExtraModules(unittest.TestCase):
             extra_allowed_modules=extras,
         )
         self.assertEqual(extras, {"vae", "transformer"})
-        self.assertNotIn(
-            "text_encoder", QwenImageLayeredPipeline._required_config_modules
-        )
         self.assertEqual(
             set(filtered),
             {
@@ -356,6 +353,7 @@ class TestPipelineSpecificExtraModules(unittest.TestCase):
                 "processor",
                 "transformer",
                 "scheduler",
+                "text_encoder",
             },
         )
 
@@ -464,21 +462,16 @@ class TestQwenImageLayeredDtype(_GlobalStageArgsMixin, unittest.TestCase):
             def to(self, *args, **kwargs):
                 return self
 
-        with patch(
+        stage = QwenImageLayeredBeforeDenoisingStage(
-            "sglang.multimodal_gen.runtime.pipelines_core.stages.model_specific_stages.qwen_image_layered.get_local_torch_device",
+            vae=_DummyVAE(),
-            return_value=torch.device("cpu"),
+            text_encoder=torch.nn.Linear(1, 1),
-        ):
+            tokenizer=object(),
-            stage = QwenImageLayeredBeforeDenoisingStage(
+            processor=object(),
-                vae=_DummyVAE(),
+            transformer=object(),
-                text_encoder=torch.nn.Linear(1, 1),
+            scheduler=object(),
-                tokenizer=object(),
+            vae_dtype=torch.float32,
-                processor=object(),
+            text_encoder_dtype=torch.float16,
-                transformer=object(),
+        )
-                scheduler=object(),
-                model_path="/unused",
-                vae_dtype=torch.float32,
-                text_encoder_dtype=torch.float16,
-            )
 
         uses = stage.component_uses(SimpleNamespace(), "qwen_layered")
         self.assertEqual(

python/sglang/multimodal_gen/test/unit/test_qwen2_5vl_generation.py

@@ -0,0 +1,226 @@
+from types import SimpleNamespace
+
+import torch
+from torch import nn
+
+import sglang.multimodal_gen.runtime.models.encoders.qwen2_5vl as qwen2_5vl
+from sglang.multimodal_gen.configs.models.encoders.qwen_image import Qwen2_5VLConfig
+from sglang.multimodal_gen.configs.pipeline_configs.longcat_image import (
+    LongCatImagePipelineConfig,
+)
+from sglang.multimodal_gen.configs.pipeline_configs.qwen_image import (
+    QwenImageLayeredPipelineConfig,
+)
+from sglang.multimodal_gen.runtime.models.encoders.qwen2_5vl import (
+    Qwen2_5_VLAttention,
+    Qwen2_5_VLForConditionalGeneration,
+    _apply_repetition_penalty,
+    _select_next_token,
+)
+from sglang.multimodal_gen.runtime.models.encoders.qwen2_5vl_vision import (
+    Qwen2_5VLVisionRotaryEmbedding,
+    Qwen2_5VLVisionTransformer,
+    _vision_position_ids,
+    _vision_window_index,
+)
+from sglang.multimodal_gen.runtime.pipelines.longcat_image import LongCatImagePipeline
+
+
+class _StubQwen2_5VL(Qwen2_5_VLForConditionalGeneration):
+    def __init__(self, next_tokens: list[list[int]], eos_token_id=5):
+        nn.Module.__init__(self)
+        self.model = nn.Module()
+        self.model.rope_deltas = torch.tensor([99])
+        self.config = SimpleNamespace(eos_token_id=eos_token_id, pad_token_id=0)
+        self.generation_config = {
+            "do_sample": True,
+            "temperature": 0.1,
+            "top_k": 1,
+            "top_p": 0.001,
+            "repetition_penalty": 1.05,
+            "eos_token_id": [eos_token_id, 6],
+            "pad_token_id": 0,
+        }
+        self.next_tokens = next_tokens
+        self.calls = []
+
+    def forward(self, input_ids, **kwargs):
+        call_index = len(self.calls)
+        self.calls.append((input_ids.clone(), kwargs))
+        logits = torch.full((input_ids.shape[0], 1, 8), -100.0)
+        for batch_index, token_id in enumerate(self.next_tokens[call_index]):
+            logits[batch_index, 0, token_id] = 100.0
+        return SimpleNamespace(
+            logits=logits,
+            past_key_values=f"cache-{call_index}",
+        )
+
+
+class _AttentionRecorder(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.masks = []
+
+    def forward(self, query, key, value, attn_mask=None):
+        self.masks.append(attn_mask)
+        return query
+
+
+def test_explicit_attention_mask_is_limited_to_cached_generation(monkeypatch):
+    attention = Qwen2_5_VLAttention.__new__(Qwen2_5_VLAttention)
+    nn.Module.__init__(attention)
+    attention.q_proj = nn.Identity()
+    attention.k_proj = nn.Identity()
+    attention.v_proj = nn.Identity()
+    attention.o_proj = nn.Identity()
+    attention.num_heads = 1
+    attention.num_key_value_heads = 1
+    attention.head_dim = 4
+    attention.rope_scaling = {"mrope_section": [1, 1, 0]}
+    attention.attn = _AttentionRecorder()
+    monkeypatch.setattr(
+        qwen2_5vl,
+        "apply_multimodal_rotary_pos_emb",
+        lambda query, key, *_args: (query, key),
+    )
+
+    hidden_states = torch.randn(1, 2, 4)
+    explicit_mask = torch.zeros(1, 1, 2, 2)
+    kwargs = {
+        "hidden_states": hidden_states,
+        "attention_mask": explicit_mask,
+        "position_embeddings": (torch.empty(0), torch.empty(0)),
+    }
+
+    attention(**kwargs, use_cache=False)
+    attention(**kwargs, use_cache=True)
+
+    assert attention.attn.masks[0] is None
+    assert attention.attn.masks[1] is explicit_mask
+
+
+def test_repetition_penalty_matches_sign_dependent_scaling():
+    logits = torch.tensor([[2.0, -3.0, 4.0]])
+    penalized = _apply_repetition_penalty(logits, torch.tensor([[0, 1]]), penalty=2.0)
+    torch.testing.assert_close(penalized, torch.tensor([[1.0, -6.0, 4.0]]))
+
+
+def test_top_k_one_sampling_is_deterministic():
+    token = _select_next_token(
+        torch.tensor([[1.0, 3.0, 2.0]]),
+        torch.tensor([[0]]),
+        do_sample=True,
+        temperature=0.1,
+        top_k=1,
+        top_p=0.001,
+        repetition_penalty=1.0,
+    )
+    assert token.tolist() == [1]
+
+
+def test_generate_reuses_cache_and_only_prefills_vision_once():
+    model = _StubQwen2_5VL([[4], [5]])
+    pixel_values = torch.ones(1, 3)
+    generated = model.generate(
+        torch.tensor([[1, 2]]),
+        attention_mask=torch.ones(1, 2, dtype=torch.long),
+        pixel_values=pixel_values,
+        image_grid_thw=torch.tensor([[1, 1, 1]]),
+        mm_token_type_ids=torch.tensor([[0, 1]]),
+        max_new_tokens=3,
+    )
+
+    assert generated.tolist() == [[1, 2, 4, 5]]
+    assert len(model.calls) == 2
+    assert model.calls[0][1]["pixel_values"] is pixel_values
+    assert model.calls[1][1]["pixel_values"] is None
+    assert model.calls[0][1]["past_key_values"] is None
+    assert model.calls[1][1]["past_key_values"] == "cache-0"
+    assert model.calls[0][1]["cache_position"].tolist() == [0, 1]
+    assert model.calls[1][1]["cache_position"].tolist() == [2]
+    assert model.model.rope_deltas is None
+
+
+def test_generate_pads_finished_rows_until_the_batch_stops():
+    model = _StubQwen2_5VL([[5, 4], [7, 6]])
+    generated = model.generate(
+        torch.tensor([[1], [2]]),
+        max_new_tokens=3,
+    )
+
+    assert generated.tolist() == [[1, 5, 0], [2, 4, 6]]
+
+
+def test_native_vision_indices_preserve_merged_token_groups():
+    grid_thw = torch.tensor([[1, 4, 4], [2, 2, 4]])
+    position_ids = _vision_position_ids(grid_thw, spatial_merge_size=2)
+    window_index, cu_window_seqlens = _vision_window_index(
+        grid_thw,
+        spatial_merge_size=2,
+        window_size=8,
+        patch_size=2,
+    )
+
+    assert position_ids.shape == (32, 2)
+    assert sorted(window_index.tolist()) == list(range(8))
+    assert cu_window_seqlens[0].item() == 0
+    assert cu_window_seqlens[-1].item() == 32
+
+
+def test_native_vision_keeps_rotary_trigonometry_in_fp32():
+    class PatchEmbed(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.proj = nn.Linear(1, 8, bias=False, dtype=torch.bfloat16)
+
+        def forward(self, hidden_states):
+            return self.proj(hidden_states)
+
+    class BlockRecorder(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.position_embedding_dtypes = None
+
+        def forward(self, hidden_states, *, position_embeddings, **_kwargs):
+            self.position_embedding_dtypes = tuple(
+                embedding.dtype for embedding in position_embeddings
+            )
+            return hidden_states
+
+    class Merger(nn.Module):
+        def forward(self, hidden_states):
+            return hidden_states.reshape(-1, 4, hidden_states.shape[-1])[:, 0]
+
+    model = Qwen2_5VLVisionTransformer.__new__(Qwen2_5VLVisionTransformer)
+    nn.Module.__init__(model)
+    model.spatial_merge_size = 2
+    model.spatial_merge_unit = 4
+    model.patch_size = 2
+    model.window_size = 8
+    model.full_attention_layers = frozenset({0})
+    model.patch_embed = PatchEmbed()
+    model.rotary_pos_emb = Qwen2_5VLVisionRotaryEmbedding(2)
+    block = BlockRecorder()
+    model.blocks = nn.ModuleList([block])
+    model.merger = Merger()
+    output = model(
+        torch.zeros(16, 1, dtype=torch.bfloat16),
+        grid_thw=torch.tensor([[1, 4, 4]]),
+    )
+
+    assert output.dtype == torch.bfloat16
+    assert block.position_embedding_dtypes == (torch.float32, torch.float32)
+
+
+def test_qwen_generation_pipelines_load_the_native_component():
+    longcat_config = LongCatImagePipelineConfig()
+
+    assert isinstance(longcat_config.text_encoder_configs[0], Qwen2_5VLConfig)
+    assert "text_encoder" in LongCatImagePipeline._required_config_modules
+    assert Qwen2_5_VLForConditionalGeneration._fsdp_forward_methods == ("generate",)
+    assert "model.visual.blocks" in Qwen2_5_VLForConditionalGeneration.layer_names
+
+    for pipeline_config in (longcat_config, QwenImageLayeredPipelineConfig()):
+        deployment = pipeline_config.get_model_deployment_config()
+        assert deployment.keep_resident_min_available_gb == 70
+        assert deployment.keep_resident_components == ("text_encoder", "vae")

python/sglang/multimodal_gen/test/unit/test_server_args.py

@@ -20,6 +20,9 @@ from sglang.multimodal_gen.configs.pipeline_configs.hunyuan import FastHunyuanCo
 from sglang.multimodal_gen.configs.pipeline_configs.lingbot_world import (
     LingBotWorldCausalDMDConfig,
 )
+from sglang.multimodal_gen.configs.pipeline_configs.longcat_image import (
+    LongCatImagePipelineConfig,
+)
 from sglang.multimodal_gen.configs.pipeline_configs.ltx_2 import (
     LTX2PipelineConfig,
     LTX23PipelineConfig,
@@ -32,6 +35,7 @@ from sglang.multimodal_gen.configs.pipeline_configs.model_deployment_config impo
 )
 from sglang.multimodal_gen.configs.pipeline_configs.mova import MOVAPipelineConfig
 from sglang.multimodal_gen.configs.pipeline_configs.qwen_image import (
+    QwenImageLayeredPipelineConfig,
     QwenImagePipelineConfig,
 )
 from sglang.multimodal_gen.configs.pipeline_configs.sana_wm import (
@@ -1146,6 +1150,32 @@ class TestOffloadDefaults(unittest.TestCase):
         self.assertEqual(qwen_deployment.keep_resident_components, ("vae",))
         self.assertIsNone(qwen_deployment.keep_resident_min_available_gb)
 
+    def test_qwen_ar_generation_residency_scales_with_available_memory(self):
+        pipeline_configs = (
+            QwenImageLayeredPipelineConfig(),
+            LongCatImagePipelineConfig(),
+        )
+
+        for pipeline_config in pipeline_configs:
+            high_memory_args = self._from_dict_with_pipeline_config(
+                pipeline_config,
+                memory_gb=80,
+                kwargs={"performance_mode": "auto"},
+            )
+            self.assertNotIn(
+                "text_encoder", high_memory_args.layerwise_offload_components or []
+            )
+            self.assertFalse(high_memory_args.text_encoder_cpu_offload)
+
+            constrained_args = self._from_dict_with_pipeline_config(
+                pipeline_config,
+                memory_gb=60,
+                kwargs={"performance_mode": "auto"},
+            )
+            self.assertIn(
+                "text_encoder", constrained_args.layerwise_offload_components or []
+            )
+
     def test_auto_multi_gpu_sana_wm_prefers_fsdp_and_cfg_parallel(self):
         args = self._from_dict_with_pipeline_config(
             SanaWMPipelineConfig(),