-
-
Notifications
You must be signed in to change notification settings - Fork 24
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
add cuda support #72
add cuda support #72
Conversation
Hi! This is the friendly automated conda-forge-linting service. I just wanted to let you know that I linted all conda-recipes in your PR ( |
hmmm doesn't seem to be that simple! |
Ok, for my own notes: bazel cache is in |
OK, I think this might work now :) |
Success!
|
You can try this one: https://anaconda.org/wolfv/jaxlib/files |
Which one should be the default variant? Cuda or no cuda? Should we make a jaxlib-gpu meta package? |
@wolfv thank you so much for your work here! I think no-cuda should be the default. Most of us don't carry around an NVIDIA-equipped laptop. 😄. jaxlib-gpu could be a good idea. It'd have to be paired with a "jax-gpu" meta-package, would that be right? The reason I ask is because Trying to work backwords, wondering what your thoughts are on establishing a 'canonical' way of installing JAX via conda/mamba? Here are some ways I can think of: conda install -c conda-forge jax # installs cpu-only? gpu-only?
conda install -c conda-forge jax jaxlib-gpu # this way to install GPU?
conda install -c conda-forge jax-gpu # this way to install GPU? |
Hi @ericmjl the GPU packages will also work with a non-gpu laptop (but the download is pretty big). We could do both, a |
These are great thoughts. Thanks a ton, @wolfv 😄. I personally think that having to worry about the cpu/gpu divide is a bit troublesome. I've never tried installing a GPU-compiled jaxlib on a CPU-only machine; have you tried that? Does the following code block work correctly? import jax.numpy as np
a = np.arange(3) If it does, then I can see a path to making the GPU package the default thing installed from conda-forge, under a command But if not, then I think having a jax/jaxlib vs. jax-gpu/jaxlib-gpu divide is a sensible thing to do. We can pick and choose based on what we need. |
As far as I understood the discussion on the tensorflow-feedstock, we would only install the |
@xhochy I am afraid that isn't true. unfortunately the However, the package itself works on CPU as well (because it comes with code for CPU and GPU at the same time). I strongly believe this is true for both tensorflow and jaxlib. Now, since a "cpu-only" package is available, we could introduce a proper run dependency on |
I think that sounds like a good idea, @wolfv! |
Just want to say thank you to all. I spent yesterday running into all sort of problems and conflicts trying to get jax, tensorflow, tensorflow-probability, distrax etc running on my 3090. Using this package is the only that work. Curious what block it from being merged? |
I can give this another push today :) |
Your wish is my command!
Let's see if this works. Trying to build it "locally" on a server right now.