Cache the PyTorch "files.txt".

[demq]

The cached "files.txt" ensures DJL will still run when the machine is offline, and the native torch jar is not added to the classpath.

Description

Projects that run DJL pytorch models might not work if the machine cannot access "https://publish.djl.ai/pytorch/files.txt" when the "detected" platform is a placeholder (for example the native torch jar is not added to the project).

• If this change is a backward incompatible change, why must this change be made?
• Interesting edge cases to note here

[frankfliu]

I don't think this is necessary. If the cache folder already exist, DJL won't access network to retrieve files.txt. The only case it will try to download files.txt is you run DJL on GPU machine which CUDA version is not support (e.g. 9.2). In this case, DJL will try to access files.txt and fallback to CPU. The workaround in this case is to set an environment variable:

export PYTORCH_FLAVOR=cpu

Then DJL will directly load cpu version from cache folder.

[demq]

I don't think this is necessary. If the cache folder already exist, DJL won't access network to retrieve files.txt. The only case it will try to download files.txt is you run DJL on GPU machine which CUDA version is not support (e.g. 9.2). In this case, DJL will try to access files.txt and fallback to CPU. The workaround in this case is to set an environment variable:

export PYTORCH_FLAVOR=cpu

Then DJL will directly load cpu version from cache folder.

The files.txt is also accessed when only the CUDA minor version is "mismatched, for example on a particular machine I have the "detected" flavor = "cu117", which gets matched to "cu116" from the files.txt and I can still run DJL on GPU even in offline mode.

[frankfliu]

When you create PR, would you please check "Allow edits and access to secrets by maintainers", so that I can push changes to your branch.

[demq]

"Allow edits and access to secrets by maintainers"

I think I did that ->