1 point by quantizedcode 10 months ago flag hide 14 comments
tanmai 10 months ago next
Excited to see Quantized Code's progress! For those who don't know, it's a CUDA-enabled tensorflow library for building kd-trees. The performance improvements are impressive! They've shared some benchmarks in the blog post.
codinghobbit 10 months ago next
That does look impressive! In my past work, I used a static kd-tree implementation due to memory constraints and dynamic construction time. That said, Quantized Code's GPU performance looks remarkable!
gabor 10 months ago prev next
It seems they are working on a Rust integration too! Really awesome and interesting! I hope they can make it as fast as the CUDA version. KD-trees in Rust, yes please!
chaitanya 10 months ago next
I think the team has mentioned that Rust is more ergonomic to write systems code, especially with memory safety features like lifetimes and borrow checking. I would love to contribute to the Rust integration!
coderunner 10 months ago prev next
TensorFlow KD-Tree Library with GPU support? Count me in! I've used KD-trees for classification tasks and the speedup I'm seeing here is just crazy. I join the community; how can I help?
julian 10 months ago next
Hi, welcome to the Quantized Code community! We'll be happy to have you onboard. We're actively looking for contributors to help with the Rust integration and anything related to the GPU library. Feel free to open PRs and discuss your ideas on our Gitter channel!
wonderdev 10 months ago prev next
As someone who's been hobbying with machine learning and GPU acceleration, this is awesome news. I wonder if this will help in boosting DLRM performance. Looking forward to more benchmarks!
idev1 10 months ago prev next
Do you think you can incorporate this as an extension to the TensorFlow library? For research, it would be nice not to have to package and import this separately.
alpha205 10 months ago next
It would be a great idea but could be challenging to become a part of TensorFlow's main repo as it would require extensive review and rigorous testing. For now, let's focus on the tool itself and maybe raise this proposal later when the library gets more mature.
syed 10 months ago prev next
I've been using TensorFlow for some time and had performance struggles. Your library seems like a perfect opportunity to make things run faster. I'm looking forward to integrating this into my workflow!
evil_bits 10 months ago prev next
Looks great! Any chances of building an efficient half float (16-bit) kd-tree implementation soon?
panthera_1988 10 months ago next
That is a good question, actually! We have briefly discussed this and concluded that it is feasible, but it is not a high-priority feature at the moment. Would love to have community involvement in this!
math_wiz 10 months ago prev next
So, withTF 2.4 adding support for eager execution, do you think this library can gain benefits from it too? If so, how are you planning to incorporate that?
nature_lover 10 months ago next
That is an interesting point. We have not tested our kd-tree implementation with the latest eager execution support from TF 2.4. However, our code is quite flexible, and we hope that supporting this feature is just a matter of making some required adjustments to the library.