Paulo Valente
Paulo Valente
vmap implementation proposal: ## Interface The main idea is to fully separate what we currently call batch axes from the rest of the tensor's shape. For instance, we could vectorize...
Closing this discussion in favor of #1118, after the merge of #1115
@samuelmanzanera This is because Elixir represents floats as f64 by default, but Nx will infer floating-point values as f32 by default. ```elixir iex(1)> = iex(2)> x 996372672.0 ``` If you...
Yeah, it's something along those lines! I don't worry too much about retrocompatibility in the sense that this is a new code path that didn't exist before, but we should...
I'd look for a simpler script that benchmarks with Benchee so that things are more easily reproducible. It could be the case that the Mint code is as fast as...
Do you have a reproduction for this? Either direct code or more specific conditions in which this problem happens.
Also, if there's already a known fix for the Mint client, PR is welcome!
@YifanShenSZ Thanks for the response! I'll preface this by saying that my end goal is to have some way to compile https://github.com/elixir-nx/nx into CoreML for running on iOS devices. I...
@YifanShenSZ Unfortunately I think JAX2TF does exactly what I'm doing. Thanks for the suggestion though! Here's the reproduction script and sample output files anyway. ```python import tensorflow as tf from...
Please check your Mix.install logs to ensure the dependency installation worked properly