Updates in Nsight Python 1.1.1#

API Changes#

  • Raise an error when no kernels were profiled (#83): @nsight.analyze.kernel now raises a RuntimeError when no kernel launch could be attributed to any nsight.annotate() region, instead of returning an empty DataFrame. Previously, with the CUPTI backend this failed silently when CUDA was used before the first profiled call.

  • Warn on ignored annotation failures (#83): When an annotation created with ignore_failures=True raises, a RuntimeWarning naming the annotation and the ignored exception is now emitted. The measured metric for that run is still set to NaN.

Fixes#

  • Custom info is no longer read back from the output directory (#86): Values gathered by info_collectors are now passed from collection to extraction in memory. Previously they were saved to collected_info.pkl under output_prefix and loaded from that path, so a file already at that path could be loaded in place of the collected values. collected_info.pkl is no longer written, and collector values that cannot be pickled are no longer replaced with None.

  • Improved kernel-count and missing-report errors (#83): The errors raised when an annotation launches more than one kernel, or a different number of kernels in different runs, now report the kernel and run counts and suggest warming up code that autotunes on its first call. The error raised when no NVIDIA Nsight Compute report is found now includes the report path and the most likely cause.

  • Plain-text output when stdout is not a terminal (#83): ANSI color and cursor-movement codes are only emitted when stdout is a terminal. When output is redirected to a file, a pipe, or a captured log, only the final progress bar is printed.