Y
HN Search
Hacker News Search
new
|
comments
|
top
|
jobs
automatistist
searching PlanetScale…
1.
▲
2.
▲
3.
▲
4.
▲
5.
▲
6.
▲
4 ms
·
1.
▲
by
automatistist
3y ago
> The standard practice for achieving fast inference is to rewrite the entire model inference loop in C++, as in FasterTransformer, and call out to special fused kernels in CUDA. But this means that any changes to the model require painf