These files are kept for reference only. They are superseded by a maintained path and are not the recommended way to run the model.

Replaced by the production agent server at serving/qwen36_agent/ (see its README.md for the run instructions and measured performance).

The agent server supersedes this single-file example on every axis:

• committed-stream decode (streams only session-committed tokens) instead of the example's one-shot stream;
• session / exact-token-prefix reuse + message-append across turns;
• execution-state capsules (snapshot / restore / fork) for shared prefixes;
• /v1/sessions management, OpenAI tool-call streaming, true SSE;
• auto-scaled CUDA-graph cache, access-log-off-by-default, per-completion metric line, explicit MTP/speculative status, and hardened request parsing.

Use python -m serving.qwen36_agent.server --checkpoint <ckpt> instead.