Skander Moalla

icon indicating a website link https://skandermoalla.com/

icon company @CLAIRE-Labo EPFL icon location Lausanne icon location PhD in Machine Learning at EPFL. Deep Reinforcement Learning.

Results 27 comments of Skander Moalla

[BUG] Problems with BatchedEnv on accelerated device with single envs on cpu

Actually there is another issue with ParallelEnv. The native truncation key is not faithful. ```python from tensordict.nn import TensorDictModule from torch import nn from torchrl.envs import ExplorationType, ParallelEnv, StepCounter, TransformedEnv...

[BUG] Problems with BatchedEnv on accelerated device with single envs on cpu

This is now solved, right?

[Feature][Rootless] Remove any need to use root privileges for development

This only applies to rootful containers. For containers with mounted namespaces (Pyxis+Enroot, apptainer) no privileges are needed.

[UX] Avoid rebuilding the project

#27 adds the build dependencies in the images and builds without isolation. This makes the build much faster in the cases where it's only Python code to link.

[BUG] Plans to update the build frontend?

> Small note though: I'm more open to change things on the RL side than tensordict, the reason being that parts of this lib is at risk of being moved...

[Feature] A minimalistic version of the template

Seems like just copying the `installation` directory gets the job done. The documentation should probably be rewritten to be agnostic to the project structure, but the installation code is usable...

torch.compile() throws exception when LigerKernel is used

I have the same issue even without `torch.compile`. I can maybe also open a separate issue? ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer device = "cuda" ckpt = "meta-llama/Llama-3.2-1B-Instruct"...