Eval bug: Responses API (/v1/responses) can`t cancel a stream to stop generation

[wqerrewetw]

Name and Version

after b7793

Operating systems

Windows

GGML backends

CUDA

Hardware

irrelevant

Models

irrelevant

Problem description & steps to reproduce

When starting an llama-server, if the HTTP connection is interrupted on the client side during the generation process, the previous Chat Completions API (v1/chat/completions) can cancel the stream to stop generation, but the new Responses API (/v1/responses) cannot.

First Bad Commit

#18486

Relevant log output

Logs

[o7si]

I can't reproduce this on macOS. It looks like it might be a Windows-specific issue.