For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Python function
begin_launch_trace
begin_launch_trace()
max.driver.begin_launch_trace() → None
Starts a process-global recording of enqueued device operations.
Records kernel launches, memory copies, and memsets across all
streams into one enqueue-ordered list, clearing any previous trace.
No stream or device handle is needed, so work enqueued on streams the
caller does not hold (e.g. a compiled graph’s internal stream) is still
captured. Only CUDA and HIP devices record entries; on other devices the
trace is always empty. Intended for tests and debugging: pair with
take_launch_trace to assert which device work a code path enqueues
and on which stream.