Deepak Singh

exploring | prev. mercor, photonics, drug discovery, harvard cs + neuro

[Deepak Singh @deepaksinghml]

Many thanks to @modal for generously supporting this project with GPU credits!

[Deepak Singh @deepaksinghml]

(1/4) @flappyairplanes raised $1B to develop megakernels for unusual, data-efficient learning.

Since I don't have $1B, I built Latch: a compiler that generates training megakernels from PyTorch. It's open source!

[Deepak Singh @deepaksinghml]

(4/4) Latch is a research preview for H100s and currently generates megakernels for MoE, state-space, diffusion, and energy-based models.

Try it out: https://t.co/j7CdketC9o

[Deepak Singh @deepaksinghml]

@ayushsatyam146 @flappyairplanes This is cool! Particularly interested in folding small ops into their consumers to cut synchronization (seen that overhead in Latch too).

Have you tried any of this on backward graphs?

[Deepak Singh @deepaksinghml]

@aakash_thatte @flappyairplanes The eventual goal is to combine execution and data efficiency gains (e.g. this recipe using finer experts and recurrence is more data efficient but runs slower)

Finer experts *should* benefit from fusion so I'm hopeful this is a good direction!