Sign In

Roshan Sumbaly

Director of AI at Meta - Llama

Roshan Sumbaly, a prominent individual in the tech industry, currently serves as the leader of the Llama Applied group in Meta AI. With a solid educational background, he holds degrees in Computer Science from prestigious institutions such as Stanford University and Birla Institute of Technology and Science, Pilani. Roshan has amassed a wealth of experience through his various leadership roles in major organizations like Meta, Coursera, and LinkedIn.

Highlights

Jul 23 · linkedin

Llama 3.1 is here - this time with 405B! 🚀

We at Meta are doubling down on our commitment to open and accessible AI - innovating not just at the foundation model layer, but also moving up the stack to help the ecosystem.

Starting with the new Llama in the room - 405B! After our April launch, we completed training our largest capacity model on 15T+ tokens taking over 30M H100-hours. We pushed the context length to 128K, added multilingual support, native tooling capability and pushed the limits on math and reasoning. This was no simple feat as it required careful coordination to merge all these capability experts back into a single model.

The 405B model has been a game changer internally acting as a powerful synthetic data generator and reward model for rejection sampling, for our smaller capacity 8B and 70B models. Both of which have also been upgraded with all the same capabilities and are now SOTA across several benchmarks. We've also made it our mission to improve the price/performance ratio of the Llama family - so we're open-sourcing an FP8 version of 405B, enabling standard single-node-8-GPU inference vs having to deal with distributed inference.

We also pushed into multimodality with the addition of visual reasoning and speech capabilities via adapters. The team had a lot of learnings here as we tried to parallelize everything and hot-swap checkpoints - definitely read the paper for more details on post-training, safety (data mixing strategy to balance vs helpfulness), factuality and red-teaming.

We're also moving up the stack - Llama was always built to be part of an overall system that can orchestrate several components, including calling external tools or running system-level safety. Today we’re excited to open-source a full reference system that parallels our internal production system and should act as a good starting point to build “agentic applications”. Further, we realized while building how fragmentation of frameworks externally for ML developers building Llama variants was slowing them down - so we're starting an RFC to define canonical interfaces that we hope the community adopts. We're calling it Llama Stack!

I'll end up saying Llama has thrived because it is a community effort. We’re partnering with 25+ partners to integrate our models, and eventually parts of Llama Stack, for everyone to play with on day 0. If you are in this space and want to talk, ping me anytime.

We’re just getting started - Llama Herd of models and system has more exciting updates coming, so stay tuned. And congrats to the team for making this happen.

Back to building!

Paper: https://lnkd.in/gDceRfiQ Download: https://lnkd.in/geZE8Yrf

Jan 19 · linkedin

When we started the Gen AI group in Meta AI beginning of 2023, one of the first meetings I got pulled into was to help with capacity planning. After having spent 2 years micro-optimizing every FLOP for training on-device models for RayBan series of glasses, it took me some time to internalize the massive compute footprint ask we were putting forward to train and deploy our Llama and Emu series of foundation models. Excited to have Mark share where we landed with the world!

We're bringing online 350K H100s, just this year, to enable our grand vision towards general intelligence. Refreshing internal dashboards daily to see these land has been an exhilarating and humbling experience. This compute is essential for us to continue innovating on the Llama family of frontier models - a key building block towards our north star.

Lots to do in 2024. I'm hiring experienced language researchers (and research leads/managers) to enable democratization of these Llama models both internally and externally via open source.

Reach out via LinkedIn message if you'd like to come aboard!

Roshan Sumbaly
Add to my network

Location

San Francisco Bay Area