Hello, thank you for the excellent IsaacGym product!
I’ve encountered an issue with setting up graphics_device_id, with camera sensor, which results in a Segmentation fault (core dumped).
Specifically, I’m operating IsaacGym on an 8 GPU server (all RTX 3090, Driver version 545.29.06), and attempting to specify GPUs using the CUDA_VISIBLE_DEVICES environment variable. It seems to work only with certain numbers for graphics_device_id; using others triggers a segmentation fault.
For example, on one server with 8 GPUs, I need to set the graphics_device_id to 0 (using CUDA_VISIBLE_DEVICES=0,2), while on another server, it has to be set to 1 (using CUDA_VISIBLE_DEVICES=4,0). In both cases, the compute_device_id is consistently set to 0. Could this issue be related to how CUDA_VISIBLE_DEVICES maps to a specific graphics_device_id ? Or might there be another underlying reason for this problem?
Also, I came across a forum post that might be related (Not sure what to set for "graphics_device_id"), but I’m unsure how to accurately determine the correct GRAHPICS_DEVICE_ID. Could you provide more guidance on this? Thank you!