How to read
Colored bands are allocation lifetimes; color follows the “Color by” setting.
Orange dashed is total reserved allocator memory, including inactive cached blocks. Enable “Reserved” to show it.
Light dotted is reserved memory for each CUDA Graph or MemPool private pool.
White dashed is the high-water mark of active allocations.
Yellow vertical marks a call to
_snapshot(); red marks an OOM. Multiple yellow lines mean memory was sampled multiple times.0→NThe x-axis is allocator event order, not elapsed time.
max-graph
python-buckets-4-shared
Click an allocation to inspect its stack trace