jtdavies· @jtdavies · X·· 13 小时前AI 评分
某模型获大幅升级:1m 上下文与更快推理实测
作者称某模型迎来重大升级,速度更快、kv-pool 显著增大、上下文最高达 1m,并提到 MiaAI_lab 的 TensorFold recipe 快 20%。作者实测 400k 上下文下达到 56 tok/s(YaRN x2),并计划当晚测试 1m 上下文。
This is stupidly good, a MAJOR upgrade!
Very fast, significantly larger kv-pool, and up to 1m context. (@MiaAI_lab's TensorFold recipe is 20% faster).
Cognisant ability is looking excellent.
I've tested 400k context with 56 tok/s (YaRN x2) and go up to 1m tonight.
来源:jtdavies · x.com