Borislav Stanimirov

Ok. so, the main problem is views. Views are pointed when the compute graph is built. Obviously if there are views to external tensors, which are then redirected this will...

@slaren so this is the current version. The limitations are as mentioned above. Example usage is added. I talked with @ggerganov and there is no metal support yet. It's not...

It's trivial now, but it won't be trivial with any of these features: * Support for resetting external data of existing tensors (and thus no requirement to rebuild the cgraph)....

BTW, @slaren how do you intend to implement external data to tensors?

If split tensor ops are eliminated, cuda extras can easily be eliminated as well. Otherwise it doesn't seem possible. An alternative approach for views would be to reinitialize them when...

Maybe be C++23 proof and implement this: https://en.cppreference.com/w/cpp/header/stacktrace

@blozano-tt you can configure with `-DCMP_SOURCE_CACHE="some/path"`. This however contrary to the PR takes precedence over an env var. If it's defined, the env var is ignored. As a side note...

Ah. I think I get it now. The goal of this PR is not to make CPM_SOURCE_CACHE_DEFAULT something that one would configure with, but to reduce snippets like yours and...

No plans yet to reconcile such clashes (no ideas too, AFAIK). The workaround is to add boost *before* those packages with the union of libraries that's actually needed. There is...

This is a rather old version of CPM that you're using. Changes to the use of `FetchContent` were implemented since then. Does the problem persist if you switch to a...