Announcing Llama 3.1 Support in vLLM

vLLM launch post for Llama 3.1 with two measured configurations: Meta-Llama-3.1-405B-Instruct-FP8 on 8xH100 (TP=8) at 2.82 req/s, and the BF16 405B on 16xH100 with PP2xTP8 vs TP16 (6.6x improvement on the non-InfiniBand cluster). FP8 GSM8K 8-shot CoT exact match 95.38% vs 96.8% BF16. Post explicitly calls the performance 'preliminary'.