TFlarge

Summary

TFlarge is a hack to allow using large tensorflow models on GPUs with limited VRAM.

How to use

Go to the TFlarge directory, then:

$ ./build

$ export LD_PRELOAD=/path/to/tflarge.so

[use tensor flow normally]

Limitations

• TFlarge works by replacing allocations in VRAM with managed allocations. This means you are not limited by the amount of VRAM you have
• Because all allocations are managed, there is a performance hit, but it's still faster than running on the CPU. For example, it is faster to train large models on my GTX 1660 Super than on my 5950X CPU.