Best way to get into CUDA mode 😎 for free these days:
- sign up on
- automatically get 15 credits (they renew every month)
- fire up a Studio, switch to GPU
- hello CUDA!
Full process from 0 to 1 here:
To help developers get started with PyTorch, we’re making the 'Deep Learning with PyTorch' book, written by Luca Antiga and Eli Stevens, available for free to the community:
⚡️Proud to announce that starting this week I’m CTO at
@gridai_
⚡️
Really grateful to
@_willfalcon
for the opportunity, and privileged to be working with a team of amazingly talented, good hearted individuals
@PyTorchLightnin
+
@gridai_
❤️
BTW, we’re hiring! ⬅️
@MistralAI
Alright, so mixture of experts just dropped!
Adding support for it in lit-gpt will be pretty straightforward, we'll need to shard each model in the mixture on one or more GPUs.
Who's up for pairing on this?
Super cool to witness the
@PyTorch
2.0 announcement here in New Orleans, together with the
@LightningAI
team (like kids at Disneyland)
@soumithchintala
and the rest of the
@PyTorch
team are making our field leap forward substantially
and BTW thanks for the shoutout
This just to say that I'm having an unfair amount of fun 💃 hacking on with
@adrianwaelchli
@carmocca
@ThomasViehmann
@aniketmaurya
+ community
Just in the last hours: low-mem loading (cool!), Adapter + LoRA tuning on Dolly, RedPajama loading
Mixtral is a great LLM to use as a writing/coding companion.
Now there's
@LightningAI
Studio for it!
⚡️run Mixtral at 12.5 tok/s on a single A10G or T4 (on your free credits)
⚡️local execution, zero data transfer
⚡️bonus: chat app in 50 lines of Python
Today is a very special day, as
@ThomasViehmann
officially joins
@LightningAI
⚡️🎉🚀
Thomas and I go way back, a friendship born in open source:
@PyTorch
, the Manning book, TorchDrift, and more
Thomas will lead cutting edge work that will shape the future of AI, stay tuned! ⚡️
Big fan of TinyLlama ⬇️
⚡️it’s based on our beloved Lit-GPT
⚡️we liked it so much that we actually started a *full reproduction* a few weeks ago, we’re close to 2.5T tokens 🚀
⚡️we’ll publish a studio very soon that will allow anyone to fully reproduce
🔥🦙 The TinyLlama project can be a game-changer - its currently pretraining a 1.1B Llama model on 3 trillion tokens.
The team aim to achieve this within a span of "just" 90 days using 16 A100-40G GPUs 🚀🚀. The training has started on 2023-09-01. 🔥🦙
Now, overall, if a model
Idk I’m growing more uncomfortable with orgs gatekeeping AI in the name of safety, at the same time exposing humans to their products and generating societal effects.
Humanity has had a tendency to work things out when given a chance. It’s time to double down on open science.
So ChatGPT knows about my medical imaging vmtk library I developed in the 2000’s and produces (correct) instructions I didn’t write 😮
It *understands* code, there’s no other way.
Here’s my talk at
@aiusergroup
AI Use Conf in SF earlier this month:
I threw in a sneak peek of DSPy running CoT + retrieval on top of
@ollama
Mixtral, all running in
@LightningAI
Studios.
Studios are quite ideal for AI builders.
So Hangar:
- it's
#git
for numerical data
- it's FAST, space-efficient
- you can read from multiple branches at the same time, from multiple processes
- you can partially clone large data repos
- it comes with data loaders for
@PyTorch
and
@TensorFlow
Just merged a
@PyTorch
PR I've been working on for a while. You can now export the JIT trace of a computation in ONNX format and load it back / execute it either from Python or from pure C++ (libtorch). Good starting point for serving. BTW consider all this experimental for now.
So, this is something we at
@orobix
have been busy with for the past several months. It has been an incredible journey with
@MilestoneItaly
, stay tuned for more details in the next few weeks.
Spoiler:
@PyTorch
, 100% RL, no tricks! Ask
@dnlcrl
:-)
Passion, Neural AI, Excitement, Powered Up Multiplayer, Competition, Historical Challenges. This is MotoGP™19. Take your place on 06.06.19 on PS4, Xbox One, PC/Steam and later on Nintendo Switch.
#MotoGPvideogame
Amazing time and full house at the NeurIPS 2023 LLM Efficiency Challenge
Congrats to the winners and kudos to
@marksaroufim
Weiwei Yang
@LChoshen
et al for kicking this off
Lots of ideas for next year already!
Small is gonna win 💪
AI should be open source no strings attached, if we want to make collective progress.
I've seen a lot of derivative work from the original llama repo and the GPL license is tricky to manage. Here's an Apache 2.0 llama derived from nanoGPT.
Good take on pineapple pizza BTW 🍍🍕
So, I’m stepping down from the role of CEO at
@orobix
. I’ll be CTO, laser focused on consolidating our technologies and creating new ones, starting from our serving infrastructure. Really excited about the opportunity to focus. Best wishes to Pietro and Roberto, new CEO and COO!
Lightning 2.1 is here!
⚡Enhanced efficiency for training large models with FSDP
⚡Seamless scaling without substantial code changes
⚡Faster speeds using bitsandbytes advanced precision plugins
See all the updates ➡️
#GenAI
#MachineLearning
#LLMs
NYC has been treating us well these past 4 months. Overall, I find the city has a calming effect on me 🤷♂️
Lots of good challenges to tackle in the New Year, looking forward to doing that in the open.
Also, I need a barber 💈
2024 will definitely be the year of compilers.
Compilers will be pervasive, from accelerating computations to running on local devices, to expressing higher level logic on top of LLMs.
Very exciting space, and one where Python will shine if you embrace it (and not strangle it).
Also, in 2024 I’m expecting approaches like DSPy to finally land us on more solid grounds when it comes to building LLM-based systems.
We need higher levels of abstraction and compilers, which leads me to my next 2024 hunch.
Also, in 2024 I’m expecting approaches like DSPy to finally land us on more solid grounds when it comes to building LLM-based systems.
We need higher levels of abstraction and compilers, which leads me to my next 2024 hunch.
For 2024 I’m expecting the next quantum leap to be in model size, but downwards ⬇️
I’m extremely bullish on low rank methods (eg ) + implications on composition and pre-training.
We’re just scratching the surface, we’ll discover we can do a ton with less.
It’s surreal to be in NYC the day Roe vs Wade was overturned. Decades of conquers in civil rights being undone one sentence at a time. This is not the future we should set up our children for.
Teaching at
@WebValley
has been such a rewarding experience.
A group of brilliant high school students from around the world engaging in medical imaging and deep learning with
@PyTorch
for 3 weeks.
Fantastic initiative, thanks for the good vibes!
It feels good to be about to meet many from the
@PyTorch
community in person at the PyTorch Dev Conf today in SF.
Oh, and we have a poster from
@orobix
on how we get PyTorch to work in healthcare and manufacturing. Go team!
🚀 🚀 welcome
@tensorwerk
team to
@PyTorchLightnin
!
Under the leadership of
@lantiga
, the team has built world-class real-time model serving () and more!
@PyTorch
fans may also recognize Luca from Deep Learning with PyTorch!
This week at
@orobix
:
@dnlcrl
got our first visual attribution WGAN working,
@Giorgia_2988
and colleagues cracked running convnets on astrophysics data using hexagonal convolutions, racing cars have learned to pass opponents while drifting. All in
@PyTorch
. Very cool. Go team!
Judging from the latest announcements, it’s quite clear that the relevance of open source AI hinges on figuring out composition.
Monolithic weight-accessible models in isolation are not going to cut it.
It blows my mind that in just a few days I’ll have
@rasbt
as a co-worker, along with the amazing
@gridai_
team.
Send your application, I can ensure 2022 is going to be pretty interesting 🤓
I was convinced I had to wind down my public speaking. I’ve felt uncomfortable and unhappy with the results for most of last year.
This past week and today I just had fun doing it and I’m fully satisfied.
What changed? Sleep.
Between slides and sleep, choose the latter.
I had such a great time at the
@PyTorch
Dev Conf yesterday.
Great format, very high profile speakers, and above all, an exceptionally positive community. I'm going back with new friends and a lot of directions for the future.
@MistralAI
Alright, so mixture of experts just dropped!
Adding support for it in lit-gpt will be pretty straightforward, we'll need to shard each model in the mixture on one or more GPUs.
Who's up for pairing on this?
For 2024 I’m expecting the next quantum leap to be in model size, but downwards ⬇️
I’m extremely bullish on low rank methods (eg ) + implications on composition and pre-training.
We’re just scratching the surface, we’ll discover we can do a ton with less.
@HamelHusain
Take a look at , you can definitely lora + adapter fine tune on a single 3090 through Lightning Fabric
It’s llama (so non commercial) but we’re coming out with the gpt-neox family (stablelm, pythia) in a sister repo: (wip)
@DrJimFan
@karpathy
Great job
@cHHillee
, an additional bit of context
@DrJimFan
: this work evolved from lit-llama and and a lot of the additional feats have been / are being upstreamed to it
So it’s out! The first racing game where opponents are 100% governed by RL agents. Amazing to think that what started as an experiment between
@MilestoneItaly
and
@orobix
will now challenge gamers around the world. Kudos to
@MilestoneItaly
🤟💪 (technical blog post in the making)
When the red lights go out and that exciting sound starts to rise, you know the moment for great challenges has come.
#MotoGP19
is now available for PS4, Xbox One, PC/Steam. Nintendo Switch™ version will be available 27 June.
#TakeYourPlace
#MotoGPvideogame
The centerline algorithm is something I came up with during my PhD: it finds a weighted geodesic on the 3D Voronoi non-manifold.
I remember I was printing a paper on medial axes and another one on geodesics and 💡! It was great to spend time down these rabbit holes as a bioeng
Segmented tree branch detection, centerline extraction, and quantification are available in new
@3DSlicerApp
module. Useful for analyzing geometry of vessels, airways, etc. Powered by VMTK.
I feel the same way.
I’ll be covering some DSPy running on
@LightningAI
Studios during the workshop at the AI User Conf in SF on Wed.
Time to focus on building systems!
The jump from langchain => DSPy feels similar to the declarative => imperative shift that PyTorch/Chainer introduced to deep learning in the late 2010s
Suddenly there was much more flexibility and you could configure complex graphs without thinking about it too much
Bullish 🚀
Beyond stoked to be on this one 🙏
It’s hard to express how much contributing to
@PyTorch
in those early days has meant to me. I learned so much and got to meet such an amazing group of people, it has definitely shaped things.
The first full paper on
@pytorch
after 3 years of development.
It describes our goals, design principles, technical details uptil v0.4
Catch the poster at
#NeurIPS2019
Authored by
@apaszke
,
@colesbury
et. al.
Horrific news. I’m watching my kids get ready for school and I can’t stop thinking about what families are going through right now,just a few hours drive from here.
Landed in SF, getting a big cup of coffee to stat awake. Very excited about these coming days. I’ll be announcing a couple of things at RedisConf, plus new video coming out.
Pretty big deal if you ask me haha ⚡️
We made building AI at scale as seamless as coding on your laptop.
It's there right now: 1 free studio + 15 fresh credits every month, just sign up
Introducing Lightning AI Studios - A persistent GPU cloud environment. Setup once. Ready any time.
Code online. Code from your local IDE. Prototype. Train. Serve. Multi-node. All from the same place.
No credit card. 6 Free GPU hours/month.
We’ve been spearheading minimal model implementations for months with
You can do *a lot* when you pair those with amazing orchestration:
- finetune < $100
- serve at 0.00002 tok/s
My take: ➡️ make it hackable before you make it fast
Getting ready to talk about TorchDrift with Mr
@ThomasViehmann
in 40 minutes at the
@PyTorch
Ecosystem Day!
The best part is finding lots of friends on there, it's definitely a joyful experience 😊
I had such a great time at
@PyTorch
Conference in the last two days.
The event was a big success, I met lots of friends from the early days and was blown away by what
@PyTorch
has become.
Also I’m very honored to sit in the PyTorch Foundation Board representing
@LightningAI
.
Great post by
@soumithchintala
on the history of
@PyTorch
, a story of cross-pollination, openness, and attribution.
These were really strong foundations for
@PyTorch
, that
@soumithchintala
and
@apaszke
upheld without making compromises.
Stoked to see my name in there too :-)
PyTorch's design origins, its connection to Lua, its intertwined deep connection to JAX, its symbiotic connection to Chainer
The groundwork for PyTorch originally started in early 2016, online, among a band of Torch7's contributors.
Torch7 (~2010-2017)
These days, we also
LoRA remains my top choice among the various effective finetuning methods for LLMs. I've just developed and shared a "LoRA From Scratch" implementation here: .
This is a hands-on approach to building LoRA from the ground up, which is, in my opinion, a
I’ll be at
@pyconit
, Saturday and Sunday. Two talks:
1. Sat at noon, all about
@orobix
: what keeps us busy, (hard) lessons learned, cool stuff coming up
2. Sun at noon,
#RedisAI
: serving DL models from Redis. I’ll also introduce
@tensorwerk
and touch upon
#Hangar
Looking fwd!
As we are approaching general availability for
#RedisAI
1.0,
@itamarhaber
has baked this awesome update to the docs (with contributions from
@fcosta_oliveira
). Check it out at
The
@orobix
team, loved ones, a few kids and a pair of (very good) dogs taking it easy at Simone’s hideout.
I owe you lots, thanks for making it all happen.
A shout out to the whole team at
@orobix
.
We’re growing stronger every day and a lot ideas we had these past years are finally coming together. I’m lucky to be part of this and I can’t wait for our next steps.
And yes, we'll eventually get a website to match :-)
Big announcement: PyTorch Foundation!
PyTorch has large core investments from many companies. So, we're creating a neutral foundation for securing assets and interests.
Technical Governance is separate & secure in a Maintainer model.
Here's more context:
Made my day:
@satyanadella
featuring our (Eli,
@lantiga
, your's truly) book on his shelf for the Build keynote. The first part, targeted at deep learning beginners, arranges the topics in an excellent order for more in-depth teaching of deep learning, too. (Hint, hint.)
Check out for
- super-readable, single-file implementations
- ready to use integrations with environments
- powered by
@LightningAI
Fabric for perf and scale
- coming up: Dreamer, RLHF!
Very proud of the team at
@orobix
Looking forward to visiting
@Cornell
in Ithaca on Tue!
I’ll be chatting with Master’s students about supercharging engineering with AI, and do so live on
@LightningAI
Studios.
The seminar is at Thurston Hall at 10am, hit me up if you are around!
I was looking forward to this: deep learning-based data assimilation for Navier Stokes flow. Work like this is going to have a big impact in bridging 4D flow (or even X-Ray angiography and Doppler) and biofluids for clinical applications
Getting back from SF after a great week. Yeah, sure meeting friends, seeing Redis thrive and announcing RedisAI was all fun, but the real highlight was sharing my airbnb with these two very friendly fellas. Zero extra fee! I had such a great time it's really hard to leave...
10 years + 1: a new logo, a new website, new visions and a new way of understanding
#artificialintelligence
thanks to an extraordinary amount of experience gained “in the field” and thanks to the people we met on this journey. Read our story!
@simonsinek
Humbled and excited for the invitation!
In the spirit of OSS: my 2 cents will be provided “as is” and without any expressed or implied warranties, including, without limitation, the implied warranties of merchantability or fitness for a particular purpose.
Kudos to
@hhsecond
for
@tensorwerk
's
@PyTorch
2020 Hackathon submission!
Stockroom is the leanest end-to-end version control system for code, data, experiments and models on the planet🌏 fully integrated with
@PyTorch
. Me thinks you should check it out :-)
We are participating in the
@PyTorch
Summer Hackathon 2020 with our project Stockroom! 🚀 Thanks
@hhsecond
!
✨ Version model, data, parameters and artifacts along with git-versioned code! ✨
Introduction video 🎥:
Submission 📋:
It was a lot of fun to sit with
@bhutanisanyam1
and the rest of the gang.
We had a good time and even came up with advices for our listeners. Just make sure you don't follow any of them :-)
Giveaway + Release:
Here's my interview w 3 great contributors to
@PyTorch
:
All about their book: Deep Learning w PyTorch by
@ManningBooks
, Open Source and PyTorch.
Eli Stevens,
@lantiga
and
@ThomasViehmann
Audio:
Video:
Want to ease into
@PyTorchLightnin
⚡️ but not ready to fully commit yet? You can start supercharging your PyTorch code with LightningLite and get multi-device, multi-GPU, perf opt for free.
Hear all about it from the amazing
@adrianwaelchli
👈
2022: more family, more focus time, more creating new stuff
plus leveling up my sight-reading and surf-skating skills - not really hard given where I am on both right now 😄
then there is where we’re going as a species, great margins for improvement there as well