Slow training or OOM Error on single gpu about director HOT 4 CLOSED

Okay almost all time is spent during the train step on the GPU, so there isn't really anything to speed it up other than getting a faster GPU.

from director.

@danijar Just a quick follow-up on this issue. Whenever you get a chance, any updates or suggestions would be awesome. Appreciate your work on this project!

from director.

What's your GPU?

Could you post a screenshot of the"timer.*avg" and "timer.*frac" scalar metrics from TensorBoard/WandB?

You can use --jax.prealloc False if you don't want to allocate all memory upfront, at a small performance cost.

from director.

Hey Danijar,

Thanks a lot for getting back to me!
I'm using a GPU cluster at our university, so my options are Tesla T4 (16GB VRAM, I hear it's around 3-4x slower than V100), an unknown GPU with 24GB VRAM (I'll update this comment once I find its name), and NVIDIA Titan Xp (12GB VRAM).

Here are the TensorBoard results for running dmc_vision / dmc_walker_walk:

I also tried using --jax.prealloc False, but it seems not to work, probably because Director doesn't use Jax and that config is only available for the Dreamer v3 codebase.

Looking forward to your thoughts on this. Thanks again!

from director.