Sarvam 30B runs efficiently on mid-tier accelerators such as L40S, enabling production deployments without relying on premium GPUs. Under tighter compute and memory bandwidth constraints, the optimized kernels and scheduling strategies deliver 1.5x to 3x throughput improvements at typical operating points. The improvements are more pronounced at longer input and output sequence lengths (28K / 4K), where most real-world inference requests fall.
专属人物小传:给每个核心 NPC 写一段详细的经历,包括性格、过往遭遇、对女主的初始态度——比如大夫 NPC 温柔但看重道德感,小将军直率且在意武艺,二皇子肤浅更看美貌,而反派则要突出其自私、虚伪的特质。这些小传直接决定 NPC 的说话风格,尤其是面对玩家辱骂时的反应,比如反派可能会嘴硬反驳,庶妹可能会委屈卖惨,渣男未婚夫可能会恼羞成怒。
。新收录的资料对此有专业解读
(I) = Intel Ice Lake/Cooper Lake
Save StorySave this story