Practical fixes that actually work, drawn from the same NVIDIA forum threads, Reddit reports, GitHub issues, and dual-Spark recipes people are using right now.
Add these two environment variables to every vLLM launch (or docker-compose / start script):
UCX_MEM_MMAP_HOOK_MODE=none
UCX_RCACHE_MAX_UNRELEASED=1024