June 2026 monthly summary for volcengine/verl: Focused on stability, security, and measurable business value. Key work included hardening resource management in NaiveRouter to prevent phantom load during retries and solidifying safe parameter handling in the Qwen3 tool parser. The changes improve reliability under load, reduce risk of load skew, and prevent code execution vulnerabilities from untrusted model outputs. Overall, delivered more predictable performance, safer automation, and a foundation for scalable tool usage in production.

Month 2026-05: Documented two major repos to improve product clarity and reduce post-release support. In jeejeelee/vllm, removed pooling multitask support in v0.21 and updated the token embedding/classification configuration docs; fixed stale version numbers in token_embed.md and token_classify.md. In NVIDIA/TensorRT-LLM, added a deprecation notice to the legacy support matrix and clarified that Torch Sampler is the default for all requests, with a corrected auto sampler behavior description for beam search. These changes align docs with current capabilities, improve user guidance, and reduce misconfigurations.