-
### Proposal to improve performance
The current execution flow with prefix caching is as follows:
1. Scheduler takes the next prefill sequence:
a. Calculate how many blocks it needs.
b. …
-
-
-
|GameID|GameState File|
|-|-|
|9627da66a27c7f839ac45babab66b0c4|ClientGameState--2036574453-[2333-2333].txt|
-
-
-
## 🐛 Bug
When using `torch.nn.functional.nll_loss` with 16-bit CUDA tensors the default reduction='mean' produces NaNs. However doing the reduction manually gives accurate results at (arguably) min…
-
Thanks for making Nx!
I tried to use `value_and_grad` on a function that takes two inputs: a vectorized tensor and a non-vectorized tensor.
``` elixir
defmodule Foo do
import Nx.Defn
defn…
-
-
**Build Scans:**
- [elasticsearch-periodic #4021 / 7.5.2_bwc](https://gradle-enterprise.elastic.co/s/dxajyxtquipuq)
- [elasticsearch-periodic #3988 / 7.3.2_bwc](https://gradle-enterprise.elastic.co/s/…