debug-cuda-crash

Debug a CUDA crash in SGLang by enabling kernel API logging to capture function boundaries, inputs, and outputs. Set SGLANG_KERNEL_API_LOGLEVEL and SGLANG_KERNEL_API_LOGDEST, run your code path to reproduce the crash, and inspect the logged tensor metadata.

An illegal memory access during CUDA kernel execution in SGLang can be diagnosed by configuring kernel API logging. This captures verbose tensor metadata and dumps at the function boundary before the device-side failure occurs.

Yes, you can configure level-based kernel API logging to capture specific tensor inputs and outputs. By adjusting the log level, you filter APIs and control the verbosity of tensor metadata dumps recorded around the failure.

The best way to reproduce device-side asserts across multi-GPU setups is to use kernel boundary tracing. Configure log destinations and apply selective logging to generate reproducible crash repros with detailed tensor metadata.

Enabling kernel API logging in SGLang captures inputs and outputs at function boundaries, which introduces overhead. Use level-based logging and API filtering to selectively control dumps and minimize performance impact during crash reproduction.

You need kernel API logging for crash dumps when encountering device-side asserts or illegal memory accesses during CUDA kernel execution. It provides actionable tensor input and output logging to trace failures at the kernel boundary.