Pytorch nightly vs Stable

[yoad.tewel]

Hey

So I saw this post: Rtx 3070 slower with mixed precision autocast in pytorch 1.7, where @ptrblck said:

@ptrblck

cudnn8.0.5 ships with the updated heuristics for the 3090 and cudnn8.1.x will cover the complete 30xx series. You could try out the nightly PyTorch build, which already uses cudnn8.0.5 and check, if the performance is improved.

So I have some questions about nightly build:

1. Which version of pytorch is the nightly? (If stable is 1.7.0, what is the version number of nightly)
2. How stable is nightly? I guess its less stable than the stable version, but should I expect problems working with it? (I work with CNNs and GANS, nothing speical)
3. When I install the nightly build with pip/conda, all the cuda and cudnn stuff are inside the binaries? And if so, the only thing I need is nvidia drivers?
4. Are there any specifications on the nightly build, like which cuda and cudnn it uses, when was the latest version released etc?

Thank you!

[ptrblck]

1. The nightly release is tagged with the date e.g. as: 1.8.0.dev20201109
2. You might see more issues, as the nightly will be rebuild every day from the current master branch. That being said, I don’t have the feeling that the nightly is terribly unstable.
3. Yes
4. You can use the install command from the website (same as for the stable release) and specify which cudatoolkit you would like to use.

[yoad.tewel]

@ptrblck Thanks!

So another quesiton, if the current nightly is 1.8.0, are there binaries of earlier nightly versions which are maybe more stable, like 1.7.1?

[ptrblck]

The nightly is not 1.8.0, which would be a future stable release, but 1.8.0.devDATE.
The stable releases are not created by tagging a specific nightly release as stable, but a versioned branch is created (e.g. at the moment for 1.7.1), which gets specific commits/fixes etc., while the master branch and the nightlies keep moving.