David (drbh) Holtz
@davidh.bsky.social
Positive > Negative
github.com/drbh
ML engineer at HuggingFace 🤗
github.com/drbh
ML engineer at HuggingFace 🤗
It’s entertaining/exciting to see GPUs discover async pipelining like more traditional CPUs
The Hopper TMA unit is a good example of this, it introduces async GEMM at the hardware level pytorch.org/blog/hopper-...
Along with newer hardware, are new fun kernel algorithms pytorch.org/blog/cutlass...
The Hopper TMA unit is a good example of this, it introduces async GEMM at the hardware level pytorch.org/blog/hopper-...
Along with newer hardware, are new fun kernel algorithms pytorch.org/blog/cutlass...
November 20, 2024 at 6:42 PM
It’s entertaining/exciting to see GPUs discover async pipelining like more traditional CPUs
The Hopper TMA unit is a good example of this, it introduces async GEMM at the hardware level pytorch.org/blog/hopper-...
Along with newer hardware, are new fun kernel algorithms pytorch.org/blog/cutlass...
The Hopper TMA unit is a good example of this, it introduces async GEMM at the hardware level pytorch.org/blog/hopper-...
Along with newer hardware, are new fun kernel algorithms pytorch.org/blog/cutlass...
Fantastic post by @FL33TW00D that helps demystify positional encoding. A must read for all ML engineers fleetwood.dev/posts/you-co...
November 18, 2024 at 4:03 PM
Fantastic post by @FL33TW00D that helps demystify positional encoding. A must read for all ML engineers fleetwood.dev/posts/you-co...