Check out the work we’ve been doing on LLM Inference in PyTorch! As models become more capable, extensibility will be as critical as performance - a point of emphasis across all of our libraries. Looking forward to the feedback and sharing more of what we have cooking :)

We recently enabled quantization-aware training in torchtune using torchao. QAT helps recover any quality degradation resulting from quantizing the model - especially useful when you’re resource constrained at inference time. 👇🏽awesome work by Andrew Or! pytorch.org/blog/quantizat…

I'll be at the PyTorch Conference on Sept 18 to talk about Triton kernels, CUDA, finetuning, hand deriving derivatives, backprop, Unsloth AI & more! Thanks to Kartikay Khandelwal for inviting me! If anyone has burning questions, lmk! Or come and say hi! Event: events.linuxfoundation.org/pytorch-confer…

Really excited to be speaking at the PyTorch Conference! I'll be talking about LLMs and fine-tuning. If you're around, come say hi! :)

I will be at PyTorch Conf all tomorrow! get in touch if you wanna say hi, or swing by the panel I’m on @ 4:45 !

torchtune v0.3.0 is officially released! We’ve added a lot of improvements in memory and performance and RLHF. The highlights [1/2]: - Full support for FSDP2 in full and (Q)LoRA finetune, upto +30%-150% speedup - RLHF with PPO, SimPO, thanks to a community contributor - DoRA

WaveForms AI will be at #NeurIPS in Vancouver Join us at our event with a16z on Wed (6pm-8pm): lu.ma/9wu5944v (limited room) Justine Moore Marco Mascorro Kartikay Khandelwal Also come talk to us at our booth in the conf! #swag #stickers

Come join us! We’re building something really cool and having a ton of fun doing it :)