Diagnose CUDA crashes by analyzing FlashInfer API logs in PyTorch workloads.
npx skills add https://github.com/sunxxuns/flashinfer --skill debug-cuda-crash
Please help me install this Agent Skill. Skill: debug-cuda-crash Source: https://github.com/sunxxuns/flashinfer/tree/main/.claude/skills/debug-cuda-crash Command: npx skills add https://github.com/sunxxuns/flashinfer --skill debug-cuda-crash
This Skill provides a guided approach to diagnosing CUDA crashes by enabling and analyzing FlashInfer's API logs, turning cryptic failures into actionable insights.
High-intent search queries and answers about installing and using this skill.
Debug a CUDA crash by enabling FlashInfer API logs to capture and inspect API calls and tensor metadata, revealing the root cause of memory errors or illegal accesses.
Diagnose illegal memory accesses by setting FLASHINFER_LOGLEVEL to 1 and FLASHINFER_LOGDEST to stdout before running your PyTorch CUDA workload to generate actionable logs.
Trace NaN and Inf issues across multi-GPU runs by enabling FlashInfer API logging to capture tensor metadata and identify offending inputs and shapes at the failing step.
This approach works with batch_decode_with_padded_kv_cache by logging API call traces and tensor metadata to quickly identify out-of-bounds errors and offending shapes.
Reproduce a CUDA crash by configuring the FlashInfer logging environment variables, then executing a CUDA-enabled PyTorch script that exercises the failing API to generate the trace.
Your CUDA workload fails cryptically because API logging is disabled; enabling FlashInfer logs turns these failures into actionable insights by recording the API call trace and tensor metadata.
jwu1980-testCapture CUDA API call traces and inputs to debug crashes.
scottgl9Record CUDA kernel calls and tensor states to diagnose GPU errors.
flashinfer-aiDiagnose CUDA crashes by capturing pre-crash API inputs with FlashInfer logging.
ariusewyDebug CUDA crashes in FlashInfer applications using API logging and compute-sanitizer.
