Agent skill · nvidia
nvshmem-troubleshoot-and-report-bugs
Diagnose NVSHMEM runtime failures and prepare bug reports for launch, crashes, hangs, correctness, transport, or topology issues.
What it needs
About 9k tokens when loaded.
What this skill does
NVSHMEM Troubleshooting and Bug Reporting Purpose Diagnose NVSHMEM runtime failures from high-signal evidence. Match documented problems conservatively, verify version-sensitive claims with live official documentation, separate NVSHMEM defects from application or environment failures, and prepare a portable report when escalation is useful. Prerequisites Require no credentials, API keys, or privileged access. Work from the user's exact launch command, complete output, and generated logs. Use direct read-only inspection only when the user supplies access and asks for it. Use $nvshmem-docs when it is registered to verify version-specific behavior, and $nvshmem-install when it is registered for installation or build handoffs. Otherwise use the live official documentation linked from the bundled knowledge base. Keep troubleshooting-knowledge-base.md available for diagnosis and bug-report-template.md available for escalation. Inputs Required: Obtain the exact command or complete batch-script launch stanza and complete stdout and stderr from every PE, including lines before the first error and any per-PE logs. Conditional: Obtain output that confirms INFO-level NVSHMEM logging before diagnosing an in-process NVSHMEM failure. Apply the documented pre-start exception only when NVSHMEM demonstrably cannot emit logs. Optional: Obtain the full NVSHMEM version, CUDA and driver versions, launcher and bootstrap, transport settings, PE/node/GPU counts, rank-to-GPU mapping, topology, smallest failing scale, and last known-good configuration only when they distinguish plausible causes. Source precedence: Prefer raw artifacts from the current run, then explicit values in the current user prompt or invocation, then non-conflicting agent context, then focused follow-up answers. Report conflicts and prefer the current raw artifact unless the user confirms a correction. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills nvidia/nvshmem-troubleshoot-and-report-bugs