Edan Toledo

thanks for making this script. When i find the time i will try to properly nail down the issue and solve it. Whilst your sampling idea is not bad, ultimately...

https://github.com/instadeepai/flashbax/pull/58 There was another bug in PER that rounded new transitions priority value which this thread helped me find. I also just made it configurable. Essentially, the sum tree implementation...

Let me think about this - when i have time i'll play around with this idea. But also some input from @SimonDuToit who is now the maintainer would be useful

Hello, so we haven't specifically asked the JAX maintainers about this issue. However, important to note for reverb and stable baselines that the memory is not stored on the GPU...

Yes i believe so, i did the benchmarks a while ago but I'm sure i would have created the data on device.

Hey. This is in my roadmap at some point but it's quite an involved algorithm. I have written it before but some thought will need to go into how to...

my code is quite messy right now - it was for a paper I submitted a while ago and it was for a multi-agent use case using dreamer and graph...

> Hey, I finally have some time to work on this. It seems that PGX [flips the board](https://github.com/sotetsuk/pgx/blob/25d5a50272cc181c225a52e54ffd4ce10101b15d/pgx/chess.py#L240) at each step, therefore the observation should be consistent between co-players: >...

Hmm, let me look into this. I unfortunately dont have access to a GPU machine currently so itll be hard for me to test this however regardless this reminds me...

@thomashirtz Did you ever figure out the issue?