Feature/Qwen3_VL

python/pyproject.toml

@@ -66,6 +66,9 @@ multimodal = [
     "torchvision",
 ]
 
+[tool.setuptools.package-data]
+sgl_jax = ["**/*.yaml"]
+
 [tool.setuptools.packages.find]
 exclude = [
     "assets*",

python/sgl_jax/srt/configs/model_config.py

@@ -59,7 +59,6 @@ def __init__(
         moe_backend: str | MoEBackend = MoEBackend.AUTO,
         model_sub_dir: str | None = None,
     ) -> None:
-
         self.model_path = model_path
         self.model_sub_dir = model_sub_dir
         self.revision = revision
@@ -696,6 +695,7 @@ def is_generation_model(model_architectures: list[str], is_embedding: bool = Fal
     "Qwen2AudioForConditionalGeneration",
     "Qwen2VLForConditionalGeneration",
     "Qwen2_5_VLForConditionalGeneration",
+    "Qwen3VLForConditionalGeneration",
     "KimiVLForConditionalGeneration",
     "InternVLChatModel",
     "Phi4MMForCausalLM",

python/sgl_jax/srt/models/registry.py

@@ -84,14 +84,14 @@ def import_model_classes():
                 entry = module.EntryClass
                 if isinstance(entry, list):  # To support multiple model classes in one module
                     for tmp in entry:
-                        assert (
-                            tmp.__name__ not in model_arch_name_to_cls
-                        ), f"Duplicated model implementation for {tmp.__name__}"
+                        assert tmp.__name__ not in model_arch_name_to_cls, (
+                            f"Duplicated model implementation for {tmp.__name__}"
+                        )
                         model_arch_name_to_cls[tmp.__name__] = tmp
                 else:
-                    assert (
-                        entry.__name__ not in model_arch_name_to_cls
-                    ), f"Duplicated model implementation for {entry.__name__}"
+                    assert entry.__name__ not in model_arch_name_to_cls, (
+                        f"Duplicated model implementation for {entry.__name__}"
+                    )
                     model_arch_name_to_cls[entry.__name__] = entry
 
     return model_arch_name_to_cls

python/sgl_jax/srt/multimodal/configs/qwen_vl/qwen3_vl_config.py

@@ -0,0 +1,176 @@
+from dataclasses import dataclass, field
+
+from sgl_jax.srt.multimodal.configs.multimodal_base_config import MultiModalModelConfigs
+
+@dataclass
+class Qwen3VLVisionConfig:
+    """Vision encoder configuration for Qwen3-VL."""
+
+    depth: int = 24
+    hidden_size: int = 1024
+    intermediate_size: int = 4096
+    num_heads: int = 16
+    in_channels: int = 3
+    patch_size: int = 16
+    temporal_patch_size: int = 2
+    spatial_merge_size: int = 2
+    out_hidden_size: int = 2048
+    num_position_embeddings: int = 2304
+    deepstack_visual_indexes: tuple = (5, 11, 17)
+    hidden_act: str = "gelu"
+    layer_norm_eps: float = 1e-6
+    rope_theta: float = 10000.0
+
+    @property
+    def head_dim(self) -> int:
+        return self.hidden_size // self.num_heads
+
+    @classmethod
+    def qwen3vl_2b(cls):
+        return cls(
+            depth=24,
+            hidden_size=1024,
+            intermediate_size=4096,
+            num_heads=16,
+            out_hidden_size=2048,
+            deepstack_visual_indexes=(5, 11, 17),
+        )
+
+    @classmethod
+    def qwen3vl_4b(cls):
+        return cls(
+            depth=24,
+            hidden_size=1024,
+            intermediate_size=4096,
+            num_heads=16,
+            out_hidden_size=2560,
+            deepstack_visual_indexes=(5, 11, 17),
+        )
+
+    @classmethod
+    def qwen3vl_8b(cls):
+        return cls(
+            depth=27,
+            hidden_size=1152,
+            intermediate_size=4304,
+            num_heads=16,
+            out_hidden_size=4096,
+            deepstack_visual_indexes=(8, 16, 24),
+        )
+
+    @classmethod
+    def qwen3vl_32b(cls):
+        return cls(
+            depth=27,
+            hidden_size=1152,
+            intermediate_size=4304,
+            num_heads=16,
+            out_hidden_size=5120,
+            deepstack_visual_indexes=(8, 16, 24),
+        )
+
+
+@dataclass
+class Qwen3VLTextConfig:
+    """Text decoder configuration for Qwen3-VL."""
+
+    vocab_size: int = 151936
+    hidden_size: int = 2048
+    intermediate_size: int = 6144
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    head_dim: int = 128
+    hidden_act: str = "silu"
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 5_000_000
+    mrope_section: tuple = (24, 20, 20)  # T, H, W partitions of head_dim
+    attention_bias: bool = False
+    tie_word_embeddings: bool = True
+
+    @classmethod
+    def qwen3vl_2b(cls):
+        return cls(
+            hidden_size=2048,
+            intermediate_size=6144,
+            num_hidden_layers=28,
+            num_attention_heads=16,
+            num_key_value_heads=8,
+            tie_word_embeddings=True,
+        )
+
+    @classmethod
+    def qwen3vl_4b(cls):
+        return cls(
+            hidden_size=2560,
+            intermediate_size=9728,
+            num_hidden_layers=36,
+            num_attention_heads=32,
+            num_key_value_heads=8,
+            tie_word_embeddings=True,
+        )
+
+    @classmethod
+    def qwen3vl_8b(cls):
+        return cls(
+            hidden_size=4096,
+            intermediate_size=12288,
+            num_hidden_layers=36,
+            num_attention_heads=32,
+            num_key_value_heads=8,
+            tie_word_embeddings=False,
+        )
+
+    @classmethod
+    def qwen3vl_32b(cls):
+        return cls(
+            hidden_size=5120,
+            intermediate_size=25600,
+            num_hidden_layers=64,
+            num_attention_heads=64,
+            num_key_value_heads=8,
+            tie_word_embeddings=False,
+        )
+
+
+@dataclass
+class Qwen3VLConfig(MultiModalModelConfigs):
+    """Combined configuration for Qwen3-VL model."""
+
+    vision_config: Qwen3VLVisionConfig = field(default_factory=Qwen3VLVisionConfig)
+    text_config: Qwen3VLTextConfig = field(default_factory=Qwen3VLTextConfig)
+    image_token_id: int = 151655
+    video_token_id: int = 151656
+    vision_start_token_id: int = 151652
+    vision_end_token_id: int = 151653
+
+    @classmethod
+    def qwen3vl_2b(cls):
+        """Qwen3-VL 2B configuration."""
+        return cls(
+            vision_config=Qwen3VLVisionConfig.qwen3vl_2b(),
+            text_config=Qwen3VLTextConfig.qwen3vl_2b(),
+        )
+
+    @classmethod
+    def qwen3vl_4b(cls):
+        """Qwen3-VL 4B configuration."""
+        return cls(
+            vision_config=Qwen3VLVisionConfig.qwen3vl_4b(),
+            text_config=Qwen3VLTextConfig.qwen3vl_4b(),
+        )
+
+    @classmethod
+    def qwen3vl_8b(cls):
+        """Qwen3-VL 8B configuration."""
+        return cls(
+            vision_config=Qwen3VLVisionConfig.qwen3vl_8b(),
+            text_config=Qwen3VLTextConfig.qwen3vl_8b(),
+        )
+
+    @classmethod
+    def qwen3vl_32b(cls):
+        return cls(
+            vision_config=Qwen3VLVisionConfig.qwen3vl_32b(),
+            text_config=Qwen3VLTextConfig.qwen3vl_32b(),
+        )

python/sgl_jax/srt/multimodal/entrypoint/http_server.py

@@ -421,10 +421,11 @@ def _execute_multimodal_server_warmup(
 
     # Send a warmup request
     # For Wan models, send an image generation request
-    if "Qwen2.5-VL" in server_args.model_path:
+    if "Qwen3-VL" in server_args.model_path or "Qwen2.5-VL" in server_args.model_path:
+        model_name = "Qwen/Qwen3-VL" if "Qwen3-VL" in server_args.model_path else "Qwen/Qwen2.5-VL"
         request_endpoint = "/v1/chat/completions"
         json_data = {
-            "model": "Qwen/Qwen2.5-VL",
+            "model": model_name,
             "messages": [
                 {
                     "role": "user",

python/sgl_jax/srt/multimodal/manager/stage.py

@@ -20,6 +20,10 @@
 from sgl_jax.srt.multimodal.models.qwen2_5VL.qwen2_5_vl_generation import (
     Qwen2_5_VL_Generation,
 )
+from sgl_jax.srt.multimodal.models.qwen3_VL.qwen3_vl_vit import Qwen3_VL_VisionModel
+from sgl_jax.srt.multimodal.models.qwen3_VL.qwen3_vl_generation import (
+    Qwen3_VL_Generation,
+)
 from sgl_jax.srt.multimodal.models.wan.diffusion.wan_dit import (
     WanDualTransformer3DModel,
     WanTransformer3DModel,
@@ -189,6 +193,10 @@ def get_model_class(name: str):
         return Qwen2_5_VL_Generation
     elif name == "Qwen2_5_VL_VisionModel":
         return Qwen2_5_VL_VisionModel
+    elif name == "Qwen3_VL_Generation":
+        return Qwen3_VL_Generation
+    elif name == "Qwen3_VL_VisionModel":
+        return Qwen3_VL_VisionModel
     elif name == "Qwen2ForCausalLM":
         return Qwen2ForCausalLM
     else:

python/sgl_jax/srt/multimodal/models/qwen3_VL/__init__.py

@@ -0,0 +1,43 @@
+"""Qwen3-VL model implementation for SGLang-JAX.
+
+This module provides the Qwen3-VL multimodal model for high-performance
+distributed inference on TPUs.
+
+Components:
+- Qwen3_VL_VisionModel: Vision encoder with DeepStack feature extraction
+- Qwen3_VL_Generation: Text decoder with M-RoPE for conditional generation
+"""
+
+from sgl_jax.srt.multimodal.models.qwen3_VL.qwen3_vl_generation import (
+    MRotaryEmbedding,
+    Qwen3_VL_Generation,
+    Qwen3_VL_Model,
+)
+from sgl_jax.srt.multimodal.models.qwen3_VL.qwen3_vl_vit import (
+    Qwen3_VL_VisionModel,
+    Qwen3_VL_VisionTransformer,
+    Qwen3_VLImageInputs,
+    Qwen3_VLVisionAttention,
+    Qwen3_VLVisionBlock,
+    Qwen3_VLVisionMLP,
+    Qwen3_VLVisionPatchEmbed,
+    Qwen3_VLVisionPatchMerger,
+    Qwen3_VLVisionRotaryEmbedding,
+)
+
+__all__ = [
+    # Vision components
+    "Qwen3_VL_VisionModel",
+    "Qwen3_VL_VisionTransformer",
+    "Qwen3_VLVisionPatchEmbed",
+    "Qwen3_VLVisionRotaryEmbedding",
+    "Qwen3_VLVisionMLP",
+    "Qwen3_VLVisionAttention",
+    "Qwen3_VLVisionBlock",
+    "Qwen3_VLVisionPatchMerger",
+    "Qwen3_VLImageInputs",
+    # Generation components
+    "MRotaryEmbedding",
+    "Qwen3_VL_Model",
+    "Qwen3_VL_Generation",
+]