[Solved] Platform:WindowlessEglApplication::tryCreateContext(): no EGL devices found / WindowlessContext: Unable to create windowless context.

[Billxgy]

This is solved by me, but I would like to upload the debug report to help others.

🐛 Bug
Running python examples/example.py --scene data/scene_datasets/habitat-test-scenes/skokloster-castle.glb inside the supplied SLURM container fails with
Platform::WindowlessEglApplication::tryCreateContext(): no EGL devices found / WindowlessContext: Unable to create windowless context.
EGL enumerates zero devices even though nvidia-smi reports a healthy GPU.

Steps to Reproduce

• Use Container since no sudo permission in a GPU Cluster.
• Run python examples/example.py --scene data/scene_datasets/habitat-test-scenes/skokloster-castle.glb.
• Logs show repeated warnings: EGL fails to find a CUDA device; eglQueryDevicesEXT returns 0.
• LD_DEBUG=libs reveals mixed NVIDIA EGL versions (570.195.03) being loaded alongside the host driver 570.158.01.

Expected Behavior
Habitat-Sim should create a windowless EGL context and render frames without manual library surgery.

Additional Context

• Containers bundle newer NVIDIA EGL libraries (570.195.03) while the host driver is 570.158.01.
• Conda packages include Mesa’s libEGL.so / libGLdispatch.so, which preload ahead of the system’s NVIDIA stack.
• Resulting symbol lookups (ErrorF, __malloc_hook, __free_hook, etc.) fail, causing zero EGL devices.
• Workaround: remove the mismatched 570.195.03 binaries, disable Mesa EGL in the conda env, and set

export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu
export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libc_malloc_debug.so.0:/usr/lib/x86_64-linux-gnu/libEGL_nvidia.so.0
export __EGL_VENDOR_LIBRARY_FILENAMES=/workspace/egl_config/10_nvidia.json

After restoring consistent driver/library versions, eglQueryDevicesEXT reports >0 devices and examples run normally.

System Info

• OS: Linux (cluster node, driver-managed; containerized environment)
• PyTorch install: prebundled with conda environment (conda)
• Build: using precompiled habitat-sim binaries from repo (no custom build)
• Python: 3.9
• GPU: NVIDIA H800 80 GB (single GPU exposed as /dev/nvidia1)
• CUDA: 12.8 (per nvidia-smi)
• Key dependencies: NumPy 1.26.4, SciPy <1.11 (warns about NumPy incompatibility); Mesa EGL libraries present in env by default
• Additional note: no sudo access; container image mixes host and container NVIDIA components.

[manainszu]

I could not work by you method