Test sglang router per service implementation

Author: Bihan Rana

src/dstack/_internal/core/models/routers.py

@@ -9,17 +9,10 @@
 
 class RouterType(str, Enum):
     SGLANG = "sglang"
-    SGLANG_DEPRECATED = "sglang_deprecated"
-    SGLANG_NEW = "sglang_new"
     VLLM = "vllm"
 
 
 class SGLangRouterConfig(CoreModel):
-    type: Literal["sglang_deprecated"] = "sglang_deprecated"
-    policy: str = "cache_aware"
-
-
-class SGLangNewRouterConfig(CoreModel):
     type: Literal["sglang"] = "sglang"
     policy: str = "cache_aware"
 
@@ -30,5 +23,5 @@ class VLLMRouterConfig(CoreModel):
 
 
 AnyRouterConfig = Annotated[
-    Union[SGLangRouterConfig, SGLangNewRouterConfig, VLLMRouterConfig], Field(discriminator="type")
+    Union[SGLangRouterConfig, VLLMRouterConfig], Field(discriminator="type")
 ]

src/dstack/_internal/proxy/gateway/model_routers/__init__.py

@@ -2,25 +2,19 @@
 
 from dstack._internal.core.models.routers import AnyRouterConfig
 from dstack._internal.proxy.gateway.model_routers.sglang import SglangRouter
-from dstack._internal.proxy.gateway.model_routers.sglang_new import SglangRouterNew
 
-from .base import Replica, Router, RouterContext
+from .base import Router, RouterContext
 
 
 def get_router(router: AnyRouterConfig, context: Optional[RouterContext] = None) -> Router:
     """Factory function to create a router instance from router configuration."""
     if router.type == "sglang":
-        return SglangRouterNew(router=router, context=context)
-    if router.type == "sglang_deprecated":
         return SglangRouter(router=router, context=context)
-    if router.type == "sglang_new":
-        return SglangRouterNew(router=router, context=context)
     raise ValueError(f"Router type '{router.type}' is not available")
 
 
 __all__ = [
     "Router",
     "RouterContext",
-    "Replica",
     "get_router",
 ]

src/dstack/_internal/proxy/gateway/model_routers/base.py

@@ -19,17 +19,6 @@ class Config:
     log_level: Literal["debug", "info", "warning", "error"] = "info"
 
 
-class Replica(BaseModel):
-    """Represents a single replica (worker) endpoint managed by the router.
-
-    The model field identifies which model this replica serves.
-    In SGLang, model = model_id (e.g., "meta-llama/Meta-Llama-3.1-8B-Instruct").
-    """
-
-    url: str  # HTTP URL where the replica is accessible (e.g., "http://127.0.0.1:10001")
-    model: str  # (e.g., "meta-llama/Meta-Llama-3.1-8B-Instruct")
-
-
 class Router(ABC):
     """Abstract base class for router implementations (e.g., SGLang, vLLM).
 
@@ -79,67 +68,35 @@ def is_running(self) -> bool:
         ...
 
     @abstractmethod
-    def register_replicas(
-        self, domain: str, num_replicas: int, model_id: Optional[str] = None
-    ) -> List[Replica]:
-        """Register replicas to a domain (allocate ports/URLs for workers).
-
-        Args:
-            domain: The domain name for this service.
-            num_replicas: The number of replicas to allocate for this domain.
-            model_id: Optional model identifier (e.g., "meta-llama/Meta-Llama-3.1-8B-Instruct").
-                Required only for routers that support IGW (Inference Gateway) mode for multi-model serving.
-
-        Returns:
-            List of Replica objects with allocated URLs and model_id set (if provided).
-
-        Raises:
-            Exception: If allocation fails.
-        """
-        ...
-
-    @abstractmethod
-    def unregister_replicas(self, domain: str) -> None:
-        """Unregister replicas for a domain (remove model and unassign all its replicas).
-
-        Args:
-            domain: The domain name for this service.
-
-        Raises:
-            Exception: If removal fails or domain is not found.
-        """
-        ...
-
-    @abstractmethod
-    def add_replicas(self, replicas: List[Replica]) -> None:
+    def add_replicas(self, replica_urls: List[str]) -> None:
         """Register replicas with the router (actual API calls to add workers).
 
         Args:
-            replicas: The list of replicas to add to router.
+            replica_urls: The list of replica URLs to add to router.
 
         Raises:
             Exception: If adding replicas fails.
         """
         ...
 
     @abstractmethod
-    def remove_replicas(self, replicas: List[Replica]) -> None:
+    def remove_replicas(self, replica_urls: List[str]) -> None:
         """Unregister replicas from the router (actual API calls to remove workers).
 
         Args:
-            replicas: The list of replicas to remove from router.
+            replica_urls: The list of replica URLs to remove from router.
 
         Raises:
             Exception: If removing replicas fails.
         """
         ...
 
     @abstractmethod
-    def update_replicas(self, replicas: List[Replica]) -> None:
+    def update_replicas(self, replica_urls: List[str]) -> None:
         """Update replicas for service, replacing the current set.
 
         Args:
-            replicas: The new list of replicas for this service.
+            replica_urls: The new list of replica URLs for this service.
 
         Raises:
             Exception: If updating replicas fails.