Global parameters and variables¶
GPU memory manager¶
Certain ptens operations, specifically linmaps
and gather
, involve creating large
temporary objects.
However, memory allocation and deallocation on GPUs can have a large latency.
To remedy this problem, the ptens backend can use a fast custom memory manager.
The amount of GPU memory allocated to this memory manager (in MB)
can be set with the vram_manager.reset function:
>> vram_manager.reset(1000)
This variable should be set with care to find the right balance between the fraction of memory available to PyTorch objects vs. the ptens backend.
ptens status¶
The ptens_base.status_str
function returns information about the current ptens session in a
string:
>> print(ptens_base.status_str())
---------------------------------------
Ptens 0.0
CUDA support: OFF
AtomsPack cat cache: 0
Graph cache: 0
Graph elist cache: 0
Subgraph cache: 0
---------------------------------------