In Pyodide micropip only allows to install pure python wheels (i.e. that don't have compiled extensions). The filename for those wheels ends with none-any.whl (see PEP 427).
If you look at Pytorch wheels currently available on PyPi, their filenames ends with e.g. x86_64.whl so it means that they would only work on the x86_64 architecture and not in the WebAssembly VM.
The general solution to this is to add a package to the Pyodide build system. However in the case of pytorch, there is a blocker that cffi is currently not supported in pyodide (GH-pyodide#761), while it's required at runtime by pytorch (see an example of build setup from conda-forge). So it is unlikely that pytorch would be availble in pyodide in the near future.
In Pyodide micropip only allows to install pure python wheels (i.e. that don't have compiled extensions). The filename for those wheels ends with
none-any.whl
(see PEP 427).If you look at Pytorch wheels currently available on PyPi, their filenames ends with e.g.
x86_64.whl
so it means that they would only work on the x86_64 architecture and not in the WebAssembly VM.The general solution to this is to add a package to the Pyodide build system. However in the case of pytorch, there is a blocker that cffi is currently not supported in pyodide (GH-pyodide#761), while it's required at runtime by pytorch (see an example of build setup from conda-forge). So it is unlikely that pytorch would be availble in pyodide in the near future.