[Bug] legacy /v1/completions disconnect-abort calls nonexistent AsyncEngine.stop_session

[AmirF194]

Describe the bug

The non-streaming /v1/completions endpoint (completions_v1 in lmdeploy/serve/openai/api_server.py) crashes on client disconnect instead of returning a clean 400.

_inner_call's disconnect branch calls VariableInterface.async_engine.stop_session(request.session_id):

async for res in cleanup_generator:
    if await raw_request.is_disconnected():
        # Abort the request if the client disconnects.
        await VariableInterface.async_engine.stop_session(request.session_id)
        return create_error_response(HTTPStatus.BAD_REQUEST, 'Client disconnected')

AsyncEngine has no stop_session method (checked by enumerating every method on the class), so this raises AttributeError on every disconnect, instead of aborting the session and returning the intended 400. The three sibling disconnect handlers in the same file (chat_completions_v1, the streaming /generate non-streaming branch, and /end_session) all use the correct session.async_abort() call.

Separately, even after routing through session.async_abort(), the disconnect response is lost: _inner_call's return create_error_response(...) is the return value of a coroutine passed to asyncio.gather(...), and that return value is discarded:

await asyncio.gather(*[_inner_call(i, generators[i], sessions[i]) for i in range(len(generators))])

response = CompletionResponse(
    id=request_id,
    ...
    choices=choices,
    ...
).model_dump()

choices[i] is never set for the disconnected generator (it stays None), so CompletionResponse construction raises a pydantic ValidationError instead of ever returning the 400.

Reproduction

Docker (python:3.11-slim, CPU-only torch/transformers/xgrammar/etc., no GPU needed), current main (2aae7d77). Imported the real, unmodified completions_v1, VariableInterface, CompletionRequest, SessionManager, and GenOut, and drove the endpoint directly with a fake AsyncEngine.generate() that yields one item and a fake raw_request.is_disconnected() that returns True:

AttributeError: 'FakeAsyncEngine' object has no attribute 'stop_session'
  File "lmdeploy/serve/openai/api_server.py", line 935, in _inner_call
    await VariableInterface.async_engine.stop_session(request.session_id)

Environment

Reproduced against InternLM/lmdeploy at commit 2aae7d77b73640bc916923cd90cb8ea8015b5d02 (2026-07-22), not tied to a specific GPU/OS environment (the fault is in request-handling logic, not a backend).

Error traceback

Traceback (most recent call last):
  File "repro.py", line 41, in main
    result = await completions_v1(request, raw_request)
  File "lmdeploy/serve/openai/api_server.py", line 963, in completions_v1
    await asyncio.gather(*[_inner_call(i, generators[i], sessions[i]) for i in range(len(generators))])
  File "lmdeploy/serve/openai/api_server.py", line 935, in _inner_call
    await VariableInterface.async_engine.stop_session(request.session_id)
AttributeError: 'FakeAsyncEngine' object has no attribute 'stop_session'

I have a fix ready (route the disconnect through the already-in-scope session.async_abort(), matching the sibling handlers, and propagate _inner_call's result out of asyncio.gather so the 400 actually reaches the caller) and will open a PR referencing this issue.