Replies: 2 comments
-
之前在我的环境里 pytorch 2 在 inference 的时候有机率 coredumps(和是否使用 torch.compile 无关),就先回滚回了 1。 |
Beta Was this translation helpful? Give feedback.
0 replies
-
v4 会包含一个配置项来开启 compile。 |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
最近尝试了pytorch 2.X的compile功能,使用后训练速度大概提升了10%。一个小的遗憾是test_play由于目前输入输出的batch size不固定暂时无法优化。
Pytorch版本:2.1.0.dev20230315 py3.10_cuda11.8_cudnn8.7.0_0
mode:
mortal = torch.compile(mortal, mode="reduce-overhead")
current_dqn = torch.compile(current_dqn, mode="reduce-overhead")
next_rank_pred = torch.compile(next_rank_pred, mode="reduce-overhead")
效果:~6.8batch/s → ~7.5 batch/s
Beta Was this translation helpful? Give feedback.
All reactions