Such a privilege to have contributed to this paper! I'm over the moon Thanks a mil to Albert Bou, Matteo Bettini, Sebastian Dittert, Shagun Sodhani, Xiaomeng Yang, Vikash Kumar and gianni de fabritiis

3 papers (with 1 spotlight) accepted at ICLR 2025 : 1. TorchRL torchrl : decision-making library for PyTorch openreview.net/forum?id=QxIto… 2. Motif: Intrinsic Motivation from AI Feedback openreview.net/forum?id=tmBKI… 3. Decision transformers for #offline #RL openreview.net/forum?id=vpV7f…

Kudos to vmoens for enabling pointwise arithmetic operations (using foreach) in tensordicts (data structure for torchrl )! github.com/pytorch/tensor… This should be helpful for usecases like implementing optimizers!

Great opportunity to work on an interesting problem and learn from an amazing team!

While I am not attending ICLR 2025 , my amazing co-authors are presenting several interesting works. Check them out below:

This is a very interesting debugging story! Also a great example of how "knowing the domain" can pay off.

I did not know that Claude.AI (from Anthropic ) is not available in Canada! support.anthropic.com/en/articles/84…

This is fixed now :) anthropic.com/news/introduci…

Attending my first Collision Conf next week and the vibes are very different from the typical science conferences that I am used to. Any tips from pro attendees on how to make the most of the #CollisionConf will be helpful! And hit me up if you want to meet during the event!

Giving a talk on "torch.func: Functional Transforms in PyTorch" next week at the Toronto Machine Learning Society (TMLS) (Toronto Machine Learning Summit) If you are attending and want to connect, send me a dm :)

Just came across nnsight.net for #interpreting the internals of PyTorch models. Skimmed the docs and it looks pretty awesome. Would love to hear some thoughts from people who have already tried it

While RL is coming back elsewhere.

distillation.... mmm 🍻 state-of-the-art Mamba models with 1% of the compute, by leveraging pretrained Transformers! key insight: project the (quadratic) attention matrices onto (structured) SSM matrix mixers before end-to-end training led by students Aviv Bick Kevin Li

📢 𝐈 𝐚𝐦 𝐨𝐧 𝐭𝐡𝐞 𝐣𝐨𝐛 𝐦𝐚𝐫𝐤𝐞𝐭 for both industry and academic positions. Please reach out if I'd be a good fit for your industry research group or academic department. Research: reinforcement learning, efficient exploration, RLHF & more! hmishfaq.github.io

I have some 1-month gift subscriptions for The Pragmatic Engineer (The Pragmatic Engineer) to give away (courtesy of Gergely Orosz ). I’ve been subscribed for a while and find it super helpful for staying updated on tech and engineering. If you’d like to try it out, just reply 🙂

Today we're opensourcing a LeanRL, a simple RL library that provides recipes for fast RL training using torch.compile and cudagraphs. Using these, we get >6x speed-ups compared to the original CleanRL implementations. github.com/pytorch-labs/l… A thread ⬇️

My team at AI at Meta is hiring research #interns to work on newer architectures like #SSMs and #distillation (esp. cross-architecture distillation). If this interests you, please apply at metacareers.com/jobs/119904986…. You can dm your resume/queries here or at linkedin.com/in/shagunsodha…

I learnt a lot from CUDA mode community just by following along the discussions in the discord. Excited that the community is expanding its scope to GPU mode . Looking forward to learning even more from the community!

This is a great opportunity, esp for MS students who are often not eligible for research internships. Gabriel Synnaeve is an amazing mentor and FAIR Paris is an amazing team!

Excited to host a tutorial on training large deep learning models with PyTorch FSDP at #ODSCWest this Tuesday, where I'll share strategies for scaling models efficiently. 🔥 Open Data Science #AI #ML #PyTorch #scaling