FlashAttention-2: Faster attention with better parallelism and work partitioning | Hacker News Reader