Ivy Zheng
ivyzhengx.bsky.social
Ivy Zheng
@ivyzhengx.bsky.social
JAX & Flax @ GoogleDeepmind
Gemini, neural net frameworks & performance
video games, and other random things
she/her, wife of a wife 🌈
in the eyes of a Chinese this question looks like "which one of you is Democrat and which one Republican" :D
February 9, 2025 at 6:13 AM
It's not, and it's hard to get real apple to apple comparison nowadays. Users likely will use whatever wholelistically better for them. I wonder how often is it that someone other than the model publisher writes kernels/other low level improvements for a model...
November 21, 2024 at 10:01 AM
Hi Akarshan, we are aware of this and actively working on it. The slowdown mostly comes from model traversals between nnx.jit and jax.jit, so it'll be less significant on large-flops models. A new release today (0.10.2) should improve it a bit and more fixes are coming (some from jax level too)
November 20, 2024 at 5:59 AM
Hi Catherine, thanks for this awesome list! Could I be added too? I work in Google for ML infra and frameworks
November 11, 2024 at 12:15 AM
yeah I guess for a product with such life span users already got used to it - which is great! And no worries you are all good!
November 2, 2024 at 6:36 PM
Thanks for replying! The historical technical restraint was true indeed (same for weibo in its early years) but in 2024 this feels more like a product decision. As a user I'd love to at least have a long text option - like if so you don't have to split your reply too! But one can only hope...
November 2, 2024 at 8:49 AM