vLLM issue #14628: two RTX 5090s cannot run tensor parallelism because P2P is driver-disabled; fixed by an NCCL commit that falls back to shared memory

Reporter shahizat, 2x RTX 5090 32GB, driver 570.124.06, CUDA 12.8.61, PyTorch 2.7.0a0+ecf3bae40a, vLLM built from source. Independent corroboration in-thread: xycjscs on 2x 5090D with QwQ-32B hit the same failure; two other commenters ask about P2P support for 5090. Fix verified by reporter against NCCL commit f44ac759fee12ecb3cc6891e9e739a000f66fd70. Explicitly reports NCCL_P2P_DISABLE=1 did NOT fix it.