fix(serve): auto-generate batch completion sessions to avoid id collision (#4773)

[Anai-Guo]

Summary

Fixes #4773.

The legacy /v1/completions endpoint assigns one session per prompt. When request.prompt is a list, completions_v1 created sessions with hardcoded user ids 1, 2, 3...:

elif isinstance(request.prompt, list):
    for i in range(len(request.prompt)):
        sessions.append(VariableInterface.create_session(i + 1))

create_session(i + 1) routes through SessionManager.map_user_session_id, which raises when a user id is already mapped. So two concurrent list-prompt requests — even single-item lists like ["hello"] with the default session_id — both request internal id 1, and the second crashes with an unhandled ValueError: User session id 1 already exists, surfacing as a 500. This branch is the only one of the four create_session call sites (:440, :830, :986) that ignores request.session_id and bypasses the user→internal mapping semantics added in #4523.

Fix

Create batch sessions the way the str branch already does for the common unset session_id=-1 case: auto-generate unique internal ids (via create_session() / session_mgr.get()), which never touch the user map and therefore cannot collide across concurrent requests.

• A single-item list now mirrors the str path exactly — honors an explicit session_id, or auto-generates when it's the default -1.
• A multi-prompt batch gets one auto-generated session per prompt, keeping every prompt distinct from each other and from other concurrent requests.

Downstream code only uses each session as an opaque handle (generate(prompt, session, ...), cleaned up by _with_request_cleanup), so the internal ids need not be contiguous.

Test

Added test_batch_completions_sessions_auto_generate_without_collision (extends the existing SessionManager-level suite, no GPU/model needed). It reproduces the reporter's minimal repro — mapping a fixed user id twice raises — and asserts that auto-generation hands out distinct, non-colliding ids for a batch of prompts.

🤖 Generated with Claude Code