BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Allsikt//Article Deadline//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
UID:dc7b62eaee0c4940e65e0bdc25ef805c42741ed7@allsikt.tech
DTSTAMP:20260728T114632Z
DTSTART;VALUE=DATE:20260728
DTEND;VALUE=DATE:20260729
SUMMARY:PipelineRL Migrates vLLM from 0.8.5 to 0.18.1\, Fixing Train‑Inference Mismatch
DESCRIPTION:Patch vLLM to 0.18.1\, enable processed_logprobs\, disable prefix caching and async scheduling\, configure weight‑update handling with pause_generation(mode='keep'\, clear_cache=False)\, and enable fp32 lm_head to restore training parity.\n\nSource: Hugging Face Blog\nOpen: https://www.allsikt.se/article/pipelinerl-migrates-vllm-from-0-8-5-to-0-18-1-fixing-train-inference-mismatch-501a76b2
URL:https://www.allsikt.se/article/pipelinerl-migrates-vllm-from-0-8-5-to-0-18-1-fixing-train-inference-mismatch-501a76b2
STATUS:CONFIRMED
TRANSP:TRANSPARENT
END:VEVENT
END:VCALENDAR