serve: type annotations and B007 for ruff

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LFPG29daR1FhchJBTjiEKr

serve/openai_server.py

@@ -26,9 +26,13 @@ import threading
 import time
 import uuid
 from pathlib import Path
+from typing import TYPE_CHECKING, Any
 
 # Module level so FastAPI can resolve the string annotation `request: Request`.
-from fastapi import Request
+from fastapi import FastAPI, Request
+
+if TYPE_CHECKING:
+    from deepinfer._C.engine import Sequence
 
 REPO_ROOT = Path(__file__).resolve().parent.parent
 sys.path.insert(0, str(REPO_ROOT / "examples"))
@@ -123,7 +127,7 @@ def worker_main(args: argparse.Namespace, dp_rank: int, in_q: mp.Queue, out_q: m
     def mm_data():
         return dict(image_bytes=[], image_sizes=[]) if engine.multi_modal_processor is not None else None
 
-    def enqueue(rid, seq, num_cached_tokens):
+    def enqueue(rid: str, seq: Sequence, num_cached_tokens: int) -> None:
         # Same as LLMEngine.add_request's _enqueue: the processor must get the sequence's own multimodal dict
         # (overwrite_prompt wrote the text/image segments into it), plus the cached prefix length.
         engine.scheduler.add_sequence(seq)
@@ -228,7 +232,7 @@ def worker_main(args: argparse.Namespace, dp_rank: int, in_q: mp.Queue, out_q: m
 class IncrementalDetokenizer:
     """vLLM-style prefix/read offsets so each delta decodes only a short tail."""
 
-    def __init__(self, tokenizer) -> None:
+    def __init__(self, tokenizer: Any) -> None:
         self.tok = tokenizer
         self.ids: list[int] = []
         self.num_tokens = 0
@@ -300,12 +304,12 @@ class Frontend:
             else:
                 self.loop.call_soon_threadsafe(self._deliver_batch, payload)
 
-    def _deliver(self, rid: str, item) -> None:
+    def _deliver(self, rid: str, item: tuple) -> None:
         q = self.streams.get(rid)
         if q is not None:
             q.put_nowait(item)
 
-    def _deliver_batch(self, batch) -> None:
+    def _deliver_batch(self, batch: list[tuple]) -> None:
         for rid, token_id, finished, reason, prompt_len in batch:
             q = self.streams.get(rid)
             if q is not None:
@@ -375,7 +379,7 @@ class Frontend:
             self.stats["finished"] += 1
 
 
-def build_app(fe: Frontend):
+def build_app(fe: Frontend) -> FastAPI:
     from fastapi import FastAPI
     from fastapi.responses import JSONResponse, PlainTextResponse, StreamingResponse
 
@@ -469,7 +473,7 @@ def build_app(fe: Frontend):
                 done = False
                 try:
                     yield chunk({"role": "assistant", "content": ""})
-                    async for text, finish in tokens():
+                    async for text, finish in tokens():  # noqa: B007  (the last finish reason is reported below)
                         if text or finish:
                             yield chunk({"content": text} if text else {}, finish)
                     done = True
@@ -484,7 +488,7 @@ def build_app(fe: Frontend):
         done = False
         try:
             parts, finish = [], None
-            async for text, finish in tokens():
+            async for text, finish in tokens():  # noqa: B007  (the last finish reason is reported below)
                 parts.append(text)
             done = True
         finally: