How to default to EGL on GPU and OSMESA on CPU?

Abierto

#419 abierto el 17/8/2023 0 comentarios 4 reacciones 0 asignados Ver en GitHub

Nadie ha tomado este issue todavía.

Evaluación

Dificultad: 4/5
Tiempo estimado: 3-5 días
Aptitud para principiantes: 35/100

Tipo de issue: Nueva funcionalidad
Claridad: Bastante claro
Estado de actividad: Estancado
Stack tecnológico: python
Área: computer-graphics, machine-learning

Línea de trabajo

Comienza rastreando cómo se selecciona la preferencia de backend entre GLFW, EGL y OSMESA, y cómo se gestiona MUJOCO_GL. Compara opciones para establecer un orden de backends configurable o valores predeterminados conscientes de la GPU; después, verifica que las máquinas con GPU elijan EGL headless, mientras que las máquinas que solo tienen CPU recurran a OSMESA sin requerir ninguna configuración del entorno del usuario.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

Right now, the preference order is glfw > egl > osmesa. In the common scenario of training an RL agent on a GPU machine, this means windows will pop up, which is typically not desired. I could also pin os.environ['MUJOCO_GL'] = 'egl' in my code to avoid this behavior. However, then the code will crash when running on a CPU-only machine.

Is there any way to set up my RL training code so that it will do headless rendering on GPU but osmesa on CPU, skipping the glfw backend? The goal here is for my RL training code to work out of the box when other people run it, without them having to set up their own env variable.

The potential workarounds I can think of all have their own downsides:

Detecting whether the machine has a GPU by asking the deep learning framework, to only set the backend to EGL if a GPU is available: This would require importing JAX inside the env processes and hog GPU memory.
Detecting whether the machine has a GPU by checking for common executable like nvcc or nvidia-smi: This doesn't work in many server environments (including internally at Google), where those binaries aren't available at runtime.

If DMC would just allow specifying a preference order for the backends or change the default to favor EGL over GLFW, that would simplify things a lot.

Lenguaje dominante: Python
Estrellas: 4.7k
Forks: 765
Métricas de merge de PR: Sin PR fusionados en 30 d