also afaik you cant install the full cuda toolchain via condo either, you dont get nvcc.
@soumithchintala
-
PyTorch CUDA Dependencies Available via pip packages now
By
–
@_seemethere can answer a few of your questions, especially around analysis of usage.
pytorch pip packages don't need the cudatoolkit, and we've been working with NVIDIA (thanks to @ptrblck_de
) to have cublas, cudnn etc. available via pip as well. -
In-house Expertise Requirements for Robotics and AI Systems
By
–
most small shops and academic shops don't have the in-house all-rounded expertise to rewrite all parts of the ecosystem they need. 1x or matic are different, you have world class well rounded experts who can write navigation, IK, slam, etc. in-house.
-
ROS ecosystem value exceeds middleware architecture
By
–
ROS is just a pub/sub in the same way Python is just a programming language.
The biggest benefit of both is their ecosystem, not their architecture.
Rewriting the middleware? then you gotta build an entirely new ecosystem.
that's why most small robotics projects and shops still -
AMD Peer-to-Peer Bandwidth Performance Impact on Distributed Inference
By
–
If you're evaluating which accelerator to choose for your future workloads beware that AMD's peer-to-peer bandwidth is 7x slower than all-to-all bandwidth. As long as you use full nodes or single gpus, you have nothing to worry about. But if you have to deploy TP=2, TP=4, or
-
Building Large-Scale Clusters to Train and Release Llama
By
–
its super fun to build very very large clusters, train llama on them, and release it for y'all to enjoy — and talk in great detail about how we did it!
It's also really fun to partner with @Ahmad_Al_Dahle in creating this disruptive chaos Join us, there's lots of work to do! -
Accelerator Support: System Dependencies and Binary Storage Challenges
By
–
there are two problems, you've identified both:
1. accelerator support involves system package and driver version resolution that we didn't have to meaningfully confront with CPUs (CPUs just had to deal with arch)
2. accelerator binaries are large — so we had to move to s3 -
PyTorch modernizes accelerator support with improved uv integration
By
–
we're willing to fix it to work better with uv (which is fantastic) cc: @_seemethere I think PyTorch was forced to solve the accelerator support problem before Python packaging was ready for it, so we solved it as best as we can — but we're willing to move to a better world if
-
GPT Action Simulation: Frame Testing and Prompt Optimization
By
–
not actually play play — i had a bunch of screen-caps and forward-simulated what GPT would take as an action.
I tried sending one frame, 10 frames, 24 frames, etc. and asked it to take one of several actions — and played with the prompting to adjust the state space and action -
GPT Boosts Robot Task Accuracy by 10 Percent Plus
By
–
Definitely wouldn't try it on real-world robots like this.
But, we did use GPT pretty successfully on RUMs ( https://
robotutilitymodels.com ) as an inference-time augmentation — bumped accuracy of tasks by a huge margin (10%+)