Sujay S Kumar

AI Infra @tesla_ai @robotaxi | ML/Speech @LTIatCMU

[SemiAnalysis @SemiAnalysis_]

What's your favorite kind of rack?

[Tesla @Tesla]

Engineering tests of the first production Cybercab have begun in Austin

[Sujay S Kumar @sujay_skumar]

Such a sad day indeed

[SemiAnalysis @SemiAnalysis_]

BREAKING NEWS: Anthropic's latest model will NOT help you if it thinks your ML research/ML engineering is interesting, and/or will secretly degrade its IQ so that the average engineer won't notice. We are already seeing Anthropic's latest model's moderation filters our GPU inference research and programming 😭

[SemiAnalysis @SemiAnalysis_]

BREAKING NEWS: Anthropic's latest model will NOT help you if it thinks your ML research/ML engineering is interesting, and/or will secretly degrade its IQ so that the average engineer won't notice. We are already seeing Anthropic's latest model's moderation filters our GPU inference research and programming 😭

[Ashok Elluswamy @aelluswamy]

@elonmusk It’s an amazing experience!

[Ashok Elluswamy @aelluswamy]

It just works!

[Tesla @Tesla]

World's first autonomous delivery of a car!

This Tesla drove itself from Gigafactory Texas to its new owner's home ~30min away — crossing parking lots, highways & the city to reach its new owner

[Elon Musk @elonmusk]

The first fully autonomous delivery of a Tesla Model Y from factory to a customer home across town, including highways, was just completed a day ahead of schedule!!

Congratulations to the @Tesla_AI teams, both software & AI chip design!

[Zack @BLKMDL3]

Our second Robotaxi ride done.

It’s smoother than a human driver. Absolutely amazing work by @Tesla_AI team. Well done.

[Tesla @Tesla]

First @robotaxi experiences in thread below

[Ashok Elluswamy @aelluswamy]

Slowly slowly at first, then …

[Sujay S Kumar @sujay_skumar]

@phalgun_g Congrats!!

[Michael Malek @mikegmalek]

@karpathy good point - I think quite a bit of extra memory could still be saved (> 10gb for large batch sizes) by splitting the final layer's activations into K blocks, and then running the lm_head + loss computation separately for each block in a loop. That way, we only need to materialize 1/K of the unnormalized logits tensor. e.g. here is a plot of peak memory usage for the lm_head + loss with vs without blocking (code link: https://t.co/pUcIiBMnnp)

[Sujay S Kumar @sujay_skumar]

@ezyang This is super cool! We do have something similar internally at Tesla. Does this work at the nn.Module level?

[Sujay S Kumar @sujay_skumar]

@rguo_asu Hi! I’m interested. Unable to DM. My Google scholar: https://t.co/CQr1R9PE5o

[Sujay S Kumar @sujay_skumar]

@Sid_Arora_18 @LTIatCMU Nice work!

[Sujay S Kumar @sujay_skumar]

@RemiCadene @Tesla Congrats Remi! Had a blast working with you at Tesla. Looking forward to your new adventures!

[Sujay S Kumar @sujay_skumar]

I think it’s natural to believe that if we add more inductive bias, we might be able to build better models. But history has shown that data >> obscure model architectures that do not scale. That being said, given that transformers have MFU of around 50-70%, there’s still a lot of headroom for another architecture to eat Transformers’ lunch

[Sujay S Kumar @sujay_skumar]

@sivareddyg @COLM_conf Done!

[Sujay S Kumar @sujay_skumar]

@_jasonwei > [2:00pm] Debug infra issues (why the fck did I pull from master?)

lol. The real issue starts when 100s of engineers have 100s of distinct branches