Issue evidence · Serve · sgl-project/sglang

Client disconnect during active request crashes entire engine (uncaught CancelledError)

GitHub issue: https://github.com/sgl-project/sglang/issues/39216

Not reproduced

A client that disconnects during a long prefill crashes the whole SGLang engine.

What Badgr ran

Ran lmsysorg/sglang:v0.5.21 with Qwen/Qwen2.5-0.5B-Instruct on an RTX 4090 through Badgr (context 32768), then aborted a ~28k-token chat request client-side six times.

What came back

  • The server stayed healthy after all six aborts.

Not the reporter's setup: they use DeepSeek-V4.1-Flash on 4x RTX 6000D with a 200,000-token prompt and the dev-dsv41 image. Only the first abort clearly landed mid-prefill; later ones returned within 0.1 s, so they may not have exercised that path.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence