torch.cuda.nvtx.range_start#
- torch.cuda.nvtx.range_start(msg)[source]#
Mark the start of a range with string message. It returns an unique handle for this range to pass to the corresponding call to rangeEnd().
A key difference between this and range_push/range_pop is that the range_start/range_end version supports range across threads (start on one thread and end on another thread).
Returns: A range handle (uint64_t) that can be passed to range_end().