refactor(service): extract inference bridge backends into sglang/vllm submodules (#1221)

Break monolithic data_proxy/backend.py into properly-separated
submodules with a protocol-based backend abstraction, preparing
the inference_service package structure for weight update integration.

Key changes:
- Add InfBridgeBackend protocol in inference_service/backend.py
- Extract SGLangBridgeBackend into sglang/bridge.py
- Extract VLLMBridgeBackend into vllm/bridge.py
- Move inf_bridge.py from data_proxy/ to inference_service/
- Create sglang/ and vllm/ as proper Python packages
- Update all consumer imports (app.py, 9 test files)

Author: garrett4wade

areal/experimental/inference_service/backend.py

@@ -0,0 +1,86 @@
+# SPDX-License-Identifier: Apache-2.0
+
+"""Inference bridge backend protocol.
+
+Defines the :class:`InfBridgeBackend` protocol that concrete backends
+(SGLang, vLLM, …) must satisfy.  See ``sglang/bridge.py`` and
+``vllm/bridge.py`` for implementations.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any, Protocol, runtime_checkable
+
+from areal.api.io_struct import HttpGenerationResult, HttpRequest
+
+if TYPE_CHECKING:
+    from areal.api.io_struct import ModelRequest
+
+
+# ---------------------------------------------------------------------------
+# Backend protocol
+# ---------------------------------------------------------------------------
+
+
+@runtime_checkable
+class InfBridgeBackend(Protocol):
+    """Protocol for inference-server backends used by :class:`InfBridge`.
+
+    Each method converts between AReaL domain objects and HTTP payloads
+    specific to a particular inference server (SGLang, vLLM, …).
+    """
+
+    def build_generation_request(
+        self,
+        req: ModelRequest,
+        with_lora: bool,
+        version: int = -1,
+    ) -> HttpRequest:
+        """Translate a :class:`ModelRequest` into a backend-specific HTTP request.
+
+        Parameters
+        ----------
+        req:
+            The model-level generation request.
+        with_lora:
+            Whether to include LoRA adapter info in the payload.
+        version:
+            Current weight version (used for LoRA versioning).
+
+        Returns
+        -------
+        HttpRequest
+            An endpoint + JSON payload ready for :pymethod:`InfBridge._send_request`.
+        """
+        ...
+
+    def parse_generation_response(
+        self,
+        response: dict[str, Any],
+    ) -> HttpGenerationResult:
+        """Parse a raw JSON response from the backend into an
+        :class:`HttpGenerationResult`.
+        """
+        ...
+
+    def get_pause_request(self) -> HttpRequest:
+        """Return the HTTP request that pauses generation on the backend."""
+        ...
+
+    def get_resume_request(self) -> HttpRequest:
+        """Return the HTTP request that resumes generation on the backend."""
+        ...
+
+    def get_generation_max_new_tokens(self, http_req: HttpRequest) -> int:
+        """Return the current generation budget encoded in ``http_req``."""
+        ...
+
+    def patch_generation_request(
+        self,
+        http_req: HttpRequest,
+        req: ModelRequest,
+        accumulated_tokens: list[int],
+        remaining_tokens: int,
+    ) -> None:
+        """Mutate ``http_req`` for an abort/resubmit iteration."""
+        ...

areal/experimental/inference_service/data_proxy/app.py

@@ -15,12 +15,7 @@
 from fastapi.responses import StreamingResponse
 from flask import Flask
 
-from areal.experimental.inference_service.data_proxy.backend import (
-    SGLangBridgeBackend,
-    VLLMBridgeBackend,
-)
 from areal.experimental.inference_service.data_proxy.config import DataProxyConfig
-from areal.experimental.inference_service.data_proxy.inf_bridge import InfBridge
 from areal.experimental.inference_service.data_proxy.pause import PauseState
 from areal.experimental.inference_service.data_proxy.session import (
     ExportTrajectoriesRequest,
@@ -35,6 +30,9 @@
 from areal.experimental.inference_service.data_proxy.tokenizer_proxy import (
     TokenizerProxy,
 )
+from areal.experimental.inference_service.inf_bridge import InfBridge
+from areal.experimental.inference_service.sglang.bridge import SGLangBridgeBackend
+from areal.experimental.inference_service.vllm.bridge import VLLMBridgeBackend
 from areal.experimental.openai.client import ArealOpenAI
 from areal.experimental.openai.proxy.server import serialize_interactions
 from areal.infra.rpc.guard.data_blueprint import (