Justus Schock

Staff Research Engineer @LightningAI ⚡️

[Justus Schock @justusschock]

It’s finally out. Come and see what we’ve been working on for years! You can try for free ⚡️

[Lightning AI ⚡️ @LightningAI]

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.

https://t.co/fa3S0wOkKu

[Lightning AI ⚡️ @LightningAI]

A robust, thoroughly tested collection of 100+ metrics with distributed synchronization for large-scale training and extensive plotting support: TorchMetrics 1.0.0 is live. 🚀
Learn more 👉 https://t.co/XwiTeD5KcD

#PyTorch #MachineLearning #DataScience

[Lightning AI ⚡️ @LightningAI]

🤔 Can you guess the secret location of our meetup with @stabilityAI in NYC? 🤫

Register now to find out! 👉👉https://t.co/387bMxJ1ud

📅May 19th | 6 PM
📍NYC | Secret Location

🚀Show off your AI demos, be inspired and join the movement to democratize AI!

#KeepAIOpenSource #NYCevents

[Lightning AI ⚡️ @LightningAI]

Calling the AI community of NYC 🗽

Come celebrate the power of open-source AI in our meetup with @StabilityAI!

📅May 19th | 6 PM
📍NYC | Secret Location

🚀Show off your AI demos, be inspired and join the movement to democratize AI!

RSVP here 👉https://t.co/387bMxJ1ud

#AI #OpenSource #NYC

[Justus Schock @justusschock]

@rasbt would you consider @geoffreyhinton to be still part of AI? During his NeurIPS2022 keynote he said, all his experiments are MATLAB

[Lightning AI ⚡️ @LightningAI]

With the help of our community, we are excited to announce PyTorch Lightning 2.0🎉🎉

Install now 👉https://t.co/Y4rxJpRkDG

Highlights Include:
⚡️ Commitment to backward compatibility in the 2.0 series
⚡️Simplified abstraction layers, removed legacy functionality, integrations out of the main repo. This improves the project's readability and debugging experience.
⚡️Introducing Fabric! Scale any PyTorch model with just a few lines of code.

#PyTorch #PyTorchLightning

[Justus Schock @justusschock]

@Luk8YrselfIshan @LightningAI Full training time depends on number and type of nodes as well as chosen model and dataset. We are currently working on cutting these times though!

[Justus Schock @justusschock]

@muchomuchacho @LightningAI Since we're just using deepspeed for this example, you can consolidate with the python file created with the checkpoint (https://t.co/0W41DK50rn)

[Justus Schock @justusschock]

@muchomuchacho @LightningAI We don't do so on purpose. If you need to do inference in a sharded way, you need to load the checkpoint when the model is sharded already or you might run into OOM issues (same with fine-tuning). (1/2)

[Justus Schock @justusschock]

@Bewinxed @LightningAI Currently in the works, will be up next (shouldn't take long)... So stay tuned :)

[Justus Schock @justusschock]

@NotMiles_ @LightningAI Not yet, but if you want a vision transformer, you should be able to just use the same techniques and plugin your vision model.
Vision examples are noted down for the future!

[Justus Schock @justusschock]

@NotMiles_ @LightningAI Also checkout https://t.co/2puBFSPcEt our minimal wrapper around minGPT and nanoGPT (using their Code without any changes!). You have support for sharding models with offloading via DeepSpeed or torch native Fully Sharded Data Parallel (FSDP)

[Justus Schock @justusschock]

@NotMiles_ @LightningAI Check out our examples https://t.co/0J0W1Tb6JE and https://t.co/LqpsugIWOV

[Eden Afek @M_a_d_a_m_E]

⚡ Lightning allows you to finetune LLMs with billions of parameters!

Train your own text classifier with 3B parameters on the cloud with just a few clicks - no need to worry about infrastructure management.

Try it out: https://t.co/BodTeiyxyI

[Justus Schock @justusschock]

@JetBrains_Fleet Will there be a possibility to user docker containers as python remote interpreters like in pycharm?

[Justus Schock @justusschock]

@pythiccoder @aniketmaurya try the same prompt multiple times. Currently the filter may be a bit oversensitive, but better to be on the safe side!

[Justus Schock @justusschock]

@pythiccoder @aniketmaurya most of the time it does not. It actually depends on the samples noise in the beginning.

[Lightning AI ⚡️ @LightningAI]

With just a few days of work, you too can build an intelligent app that is:

📈 Scalable
💪 Performant & distributed
🎨 Fully customizable

Learn how 👉👉👉 https://t.co/nxr1E8Z8IQ

#Diffusion #AI #ML #BuildWithLightning

[Justus Schock @justusschock]

@DorGoldenberg @rasbt @yaroslavvb you would have to sample it on rank0 and then broadcast it manually. This would probably be the boilerplate @chaton_thomas meant and lightning has primitives for that.

[William Falcon ⚡️ @williamfalcon]

Excited to share the official torch metrics paper was published in JOSS, nice job @PyTorchLightnin team!
Torch metrics is a package for reproducible metrics optimized for scale even with the most complex distributed @PyTorch models