Makes sense to me to release nightlies to s3 only. Can we reduce size by cutting down on the SMs we release? Was the main complaint around cuda release sizes?
On Dec 1, 2019 9:43 PM, "Lausen, Leonard" <lau...@amazon.com.INVALID> wrote: Hi MXNet Community, since more than 2 months our binary Python nightly releases published on Pypi are broken. The problem is that our binaries exceed Pypi's size limit. Decreasing the binary size by adding compression breaks third-party libraries loading libmxnet.so https://github.com/apache/incubator-mxnet/issues/16193 Sheng requested Pypi to increase their size limit: https://github.com/pypa/pypi-support/issues/50 Currently "the biggest cost for PyPI from [the many MXNet binaries with nightly release to Pypi] is the bandwidth consumed when several hundred mirrors attempt to mirror each release immediately after it's published". So Pypi is not inclined to allow us to upload even larger binaries on a nightly schedule. Their compromise is to allow it on a weekly cadence. However, I would like the community to revisit the necessity of releasing pre- release binaries to Pypi on a nightly (or weekly) cadence. Instead, we can release nightly binaries ONLY to a public S3 bucket and instruct users to install from there. On our side, we only need to prepare a html document that contains links to all released nightly binaries. Finally users will install the nightly releases via pip install --pre mxnet-cu101 -f http://mxnet.s3.amazonaws.com/mxnet-cu101/ nightly.html Instead of pip install --pre mxnet-cu101 Of course proper releases and release candidates should still be made available via Pypi. Thus releases would be installed via pip install mxnet-cu101 And release candidates via pip install --pre mxnet-cu101 This will substantially reduce the costs of the Pypi project and in fact matches the installation experience provided by PyTorch. I don't think the benefit of not including "-f http://mxnet.s3.amazonaws.com/mxnet-cu101/nightly.html"; matches the costs we currently externalize to the Pypi team. This suggestion seems uncontroversial to me. Thus I would like to start lazy consensus. If there are no objections, I will assume lazy consensus on stopping nightly releases to Pypi in 72hrs. Best regards Leonard
