Log FlashInfer API inputs to diagnose CUDA crashes.
npx skills add https://github.com/fw-ai/flashinfer --skill debug-cuda-crash-fw-ai
Please help me install this Agent Skill. Skill: debug-cuda-crash Source: https://github.com/fw-ai/flashinfer/tree/main/.claude/skills/debug-cuda-crash Command: npx skills add https://github.com/fw-ai/flashinfer --skill debug-cuda-crash-fw-ai
This Skill helps you diagnose and fix CUDA crashes in your FlashInfer applications by providing detailed logging of API inputs before errors occur.
Enable detailed API logging to a file named debug.log by running export FLASHINFER_LOGLEVEL=3 and export FLASHINFER_LOGDEST=debug.log before executing your Python script.
High-intent search queries and answers about installing and using this skill.
You can debug illegal memory access in FlashInfer by enabling API logging to capture tensor shapes, dtypes, and values before the crash. Set FLASHINFER_LOGLEVEL=3 and FLASHINFER_LOGDEST=debug.log, reproduce the crash, and examine the logged inputs to pinpoint the problematic tensor configuration.
FlashInfer model training crashes with NaN or Inf values due to invalid tensor inputs or incorrect data types. Enabling API logging captures tensor statistics before the error, allowing you to identify the exact data or configuration that triggers the invalid numerical values during execution.
The best way to capture tensor inputs before a CUDA out-of-memory error is enabling API logging in FlashInfer. By setting FLASHINFER_LOGLEVEL=3, the system records tensor shapes and values continuously, ensuring pre-crash data is available for diagnosis when the OOM error halts execution.
Yes, FlashInfer supports advanced debugging using compute-sanitizer and cuda-gdb alongside API logging. These tools enable deeper kernel-level inspection through printf statements and multi-process logging to diagnose complex CUDA crashes beyond standard tensor input analysis.
Yes, FlashInfer API logging supports multi-process environments for diagnosing CUDA crashes. Advanced debugging techniques include multi-process logging, allowing you to track tensor inputs and identify root causes like incorrect data types or illegal memory access across distributed training workflows.
API logging for debugging CUDA crashes is limited to capturing tensor shapes, dtypes, and values before the error occurs. While it identifies root causes like illegal memory access or OOM, complex kernel-level issues may require additional tools like compute-sanitizer or cuda-gdb for full diagnosis.
flashinfer-aiDiagnose CUDA crashes by capturing pre-crash API inputs with FlashInfer logging.
FutureMLS-LabDiagnose SGLang CUDA crashes by recording kernel API calls and tensor metadata.
ariusewyDebug CUDA crashes in FlashInfer applications using API logging and compute-sanitizer.
Nabilhassan12345Diagnose CUDA crashes in SGLang with kernel API logging.
