debug-cuda-crash

To debug CUDA crashes from illegal memory access, enable API logging with the @flashinfer_api decorator to capture pre-crash tensor inputs and metadata. Set the required FLASHINFER_LOGLEVEL and FLASHINFER_LOGDEST environment variables to record function calls before failure.

The best way to diagnose kernel failures and NaN/Inf errors is to use API logging to inspect input tensors prior to the crash. Applying the @flashinfer_api decorator captures function calls and tensor metadata, allowing you to reproduce crashes in controlled environments.

You can capture PyTorch API inputs before a CUDA kernel crash by enabling the @flashinfer_api logging decorator. Configure FLASHINFER_LOGLEVEL and FLASHINFER_LOGDEST to output detailed logs, then run your script to reproduce the crash and inspect the recorded inputs.

Yes, you need to set the FLASHINFER_LOGLEVEL and FLASHINFER_LOGDEST environment variables to enable API logging for debugging. These settings control the verbosity and destination of your logs, allowing you to capture function calls and tensor metadata before a crash.

Your CUDA application may crash due to illegal memory access, NaN/Inf values, or kernel failures. You can reproduce the crash for debugging by running your script in a controlled environment with API logging enabled to observe function calls and extract actionable input data before failure.