🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

🚀 We fixed a major LLM post-training bottleneck! Our new method (TBA) combines trajectory balance with asynchronous training to speed up LLM RL 5-50x while improving results+scalability. For example, using VinePPO's GSM8K setup, we obtain +1.2% accuracy and 50x faster RL.

We worked a bit on non-tensor data within tensordict, it's looking pretty good! The reshaping feature is really joyful to play with

The true coding assistant will be able to understand when I say "Oh you know what I mean c'mon" Now I still feel it's the assistant telling me "Oh you know what I mean"

torchrl 🤝 gymnasium happy ever after With the help of the Farama Foundation team, we managed to make TorchRL compatible with gymnasium v1.1 onward!

Occasional reminder that you can get a 6x speedup in torchrl by using cudagraphs + compile for training

Then, you go async and get a 12x speedup compared to eager mode...

Claude writes a PR, Claude reads the PR on gh and congratulates itself of the quality, then approves it.

It’s mid 2025, we have autonomous cars, AI etc, and almost all the online forms I fill still mistakenly put a cap letter at the beginning of my email address