1.. _torch_cuda_memory: 2 3Understanding CUDA Memory Usage 4=============================== 5To debug CUDA memory use, PyTorch provides a way to generate memory snapshots that record the state of allocated CUDA memory 6at any point in time, and optionally record the history of allocation events that led up to that snapshot. 7 8The generated snapshots can then be drag and dropped onto the interactiver viewer hosted at `pytorch.org/memory_viz <https://pytorch.org/memory_viz>`_ which 9can be used to explore the snapshot. 10 11Generating a Snapshot 12===================== 13The common pattern for recording a snapshot is to enable memory history, run the code to be observed, and then save a file with a pickled snapshot: 14 15.. code-block:: python 16 17 # enable memory history, which will 18 # add tracebacks and event history to snapshots 19 torch.cuda.memory._record_memory_history() 20 21 run_your_code() 22 torch.cuda.memory._dump_snapshot("my_snapshot.pickle") 23 24Using the visualizer 25==================== 26 27Open `pytorch.org/memory_viz <https://pytorch.org/memory_viz>`_ and drag/drop the pickled snapshot file into the visualizer. 28The visualizer is a javascript application that runs locally on your computer. It does not upload any snapshot data. 29 30 31Active Memory Timeline 32---------------------- 33 34The Active Memory Timeline shows all the live tensors over time in the snapshot on a particular GPU. Pan/Zoom over the plot to look at smaller allocations. 35Mouse over allocated blocks to see a stack trace for when that block was allocated, and details like its address. The detail slider can be adjusted to 36render fewer allocations and improve performance when there is a lot of data. 37 38.. image:: _static/img/torch_cuda_memory/active_memory_timeline.png 39 40 41Allocator State History 42----------------------- 43 44The Allocator State History shows individual allocator events in a timeline on the left. Select an event in the timeline to see a visual summary of the 45allocator state at that event. This summary shows each individual segment returned from cudaMalloc and how it is split up into blocks of individual allocations 46or free space. Mouse over segments and blocks to see the stack trace when the memory was allocated. Mouse over events to see the stack trace when the event occurred, 47such as when a tensor was freed. Out of memory errors are reported as OOM events. Looking at the state of memory during an OOM may provide insight into why 48an allocation failed even though reserved memory still exists. 49 50.. image:: _static/img/torch_cuda_memory/allocator_state_history.png 51 52The stack trace information also reports the address at which an allocation occurred. 53The address b7f064c000000_0 refers to the (b)lock at address 7f064c000000 which is the "_0"th time this address was allocated. 54This unique string can be looked up in the Active Memory Timeline and searched 55in the Active State History to examine the memory state when a tensor was allocated or freed. 56 57Snapshot API Reference 58====================== 59 60.. currentmodule:: torch.cuda.memory 61.. autofunction:: _record_memory_history 62.. autofunction:: _snapshot 63.. autofunction:: _dump_snapshot 64
