Rafi Ayub

@official_j3rck "slightly less"

Fun times, fun people, slightly less existential dread about AI Come join

So, I did a thing 🙂 This was really just a fun little side project - I wanted to spend some time working on agents, and Pokemon was the most fun way I could come up with. And then it kinda took off! 3.7 Sonnet is so fun to watch play!

Thanks for the contribution!! Now you can do SFT, KD, QAT, RLHF, *and* RL completely within torchtune … let’s post-train some crazy models 🫡🫡🫡

btw torchtune is officially reinforcement learning - the GRPO implementation is officially merged, the entire codebase is really clean and modifiable, so go out there, reinforcement learn your LLMs, and any contributions welcome!

@official_j3rck No no, queue up Severance Season 2

Twin Peaks Season 3 queued up now No one release anything cool in the meantime

@official_j3rck We are your family

Starting a fine tuning run now so I have an excuse to leave “family holiday time” every 30 mins to check loss curves

If you have a lot of experience training and fine-tuning ML models and want to help bring that expertise to the community, we’re looking to hire a new member for the torchtune team! Full description in link below

New release new release new release!!! 🎁 Kaggle collab notebooks 🎁 Early exit training recipe 🎁 QAT + LoRA github.com/pytorch/torcht…

starting a VC fund that invests in founders who sleep 8 hours a night, have significant others, consistently work out and bathe regularly

ai ain’t taking everyone’s jobs, let’s be real. what it’s actually doing is making the human touch worth its weight in gold. when a single prompt can spit out a track, live acts and real bands are gonna hit harder than ever, trust me. as prompt-to-video tech blows up, live theater is about to flex its status exclusive, top-tier, luxury vibes only. jazz clubs, underground poetry readings, and the rise of live painting as full blown performance art shows on a mainstream level. i can already picture schools bussing kids in just to catch a glimpse of real human creativity, raw and unfiltered instead of homeworks like watching a movie on netflix or a documentary human made art’s about to go premium, mega luxurious, almost untouchable. handwritten books, custom-made crafts, anything touched by human hands will be the new holy grail of pop culture. give it a decade even less, and that’ll be the game. just some thoughts running through my head lately. feels like today is the perfect time to create art, start today.

Knowledge distillation is how the Llama 3.2 1B and 3B models still pack a punch and why they’re our most used models. It will remain key in creating lightweight, performant, task-specific LLMs to automate production workflows. Read more for how this is done in torchtune 👇

Activation offloading is one of those few free lunches - 20% reduction in memory with only a 1% slowdown by leveraging high throughput memory bandwidth. Just enable with a single parameter in your torchtune config 💪

@marksaroufim @t0kenl1mit @kakemeister MHA and transformer components have been rewritten so many times for different models because of small but significant differences. Only recently has the field converged to more standard decoder components. Still, it’s been hard to maintain a unified MHA just within torchtune

@t0kenl1mit @kakemeister pytorch.org/docs/stable/ge… should be deprecated soon - we'll have an updated doc on this soon. (the lack of a notice proves your point)

If you could change one thing about PyTorch what would it be?

Super exciting collab to make it easier to fine tune w your favorite Kaggle models 🙏

Wake up and do not go directly to work. Make yourself a coffee. Go for a walk. Seize a little time for yourself before the day seizes you.

@ashwinb This must’ve been written by a Southern Hemispheran

Some intense smelly BS.

@official_j3rck git back to git

bored bored bored

PyTorch 2.5 is here 🔥 We are excited to announce the release of #PyTorch 2.5, featuring a new CuDNN backend for SDPA, regional compilation of torch.compile, & TorchInductor CPP backend performance speedup Read more in our blog: hubs.la/Q02TRs9p0

One interesting idea in Anthropic’s essay was that AI could accelerate progress in medicine not by analyzing data but by playing PI, designing and conducting experiments (with robots) based on literature it’s trained on Hard to realize that with current LLMs, which cannot reason