| theme | background | title | info | class | highlighter | drawings | transition | mdc | fonts | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
seriph |
Welcome to Slidev |
## Slidev Starter Template
Presentation slides for developers.
Learn more at [Sli.dev](https://sli.dev)
|
text-center |
shiki |
|
slide-left |
true |
|
Twitter: @thejaan Email: [email protected]
Press Space for next page
Lightning makes things easier and abstracts the details. In research (or in edge cases), you might need to go deeper. For example, torch.compile, custom metrics, etc.
- 📝 Incentives - compete with AnyScale, Ray, SkyPilot (multi-cloud, recommended)
- 🚀 Risk - Need to be careful relying on venture capital-backed code (ideal customer profile may be different, may need to pivot, short runway and high return expectations from LPs)
- 🔍 Research - usually need to use algorithms like HyperBand and use fully-sharded data parallel models for best performance. Supported in Ray and SkyPilot, work-in-progress for Lightning.
- 🔥 Time Horizon - behavior change is hard. Your future self will thank you but need to learn by experience when to use a shortcut / abstraction versus when to use the details. For me, I need to see the details almost always (not assembly/CUDA regularly though!)
"Writing a training loop is not that hard; why use a library?"
- Amazon Web Services, Google Cloud Platform, Microsoft Azure - all competing for mindshare
- If you reduce your cognitive load, you give up mindshare in exchange for "efficiency".
- It is up to you to decide when this trade-off is worth it.
- For me, it rarely is: libraries move fast, and I need to understand the details to be able to debug and optimize.
Hover on the bottom-left corner to see the navigation's controls panel, learn more
| right / space | next animation or slide |
| left / shiftspace | previous animation or slide |
| up | previous slide |
| down | next slide |
Here!