←新闻详情

UCSD Hao AI Lab更新于 10月13日 17:37

FastVideo FastH3 解码:VAE 批处理与 CUDA 图

(8/13) Decode: the 28 overlapping tiles of each temporal unit go through the VAE decoder as one batch, with units spread across GPUs. The fused decoder is repla

AI 导读

(8/13)解码:每个时间单元的 28 个重叠 tile 作为一个批次通过 VAE 解码器,各单元分散在不同 GPU 上。融合后的解码器以 CUDA 图方式重放:14 次图启动即可解码一段 10 秒视频。

正文

当前显示原文,中文全文译文尚未就绪。

(8/13) Decode: the 28 overlapping tiles of each temporal unit go through the VAE decoder as one batch, with units spread across GPUs. The fused decoder is replayed as a CUDA graph: 14 graph launches decode a 10-second clip. https://t.co/1OSrAKx39u

来源

UCSD Hao AI Lab