Add CONFIG command for configuring/loading backends about redisai HOT 9 CLOSED

@lantiga @K-Jo What do you think the best way of distributing Docker images in this case. We can definitely support docker images for each backend separately but we should support a single image with all the backends as well (we can let the user load it dynamically but we probably need to install all the required CUDA versions beforehand).

from redisai.

@hhsecond Good point. I think having different container images for each backend is a bit overkill and in any case it could limit what users can do (the PyTorch backend is needed for SCRIPT, etc).
I'd rather build at most 3 images:

• all CPU
• TF GPU, TORCH GPU, ONNX CPU
• all GPU (with CUDA 9.0 + 9.1)

from redisai.

/cc @hhsecond @gkorland @K-Jo I've got dynamic loading implemented. I just have an issue with jemalloc that I'll likely need help with.

from redisai.

@mnunberg @MeirShpilraien can you please see if you can help?

from redisai.

FYI, @MeirShpilraien helped me and the issue is now solved.

from redisai.

@hhsecond Good point. I think having different container images for each backend is a bit overkill and in any case it could limit what users can do (the PyTorch backend is needed for SCRIPT, etc).
I'd rather build at most 3 images:

• all CPU
• TF GPU, TORCH GPU, ONNX CPU
• all GPU (with CUDA 9.0 + 9.1)

When we add TPU and AMD GPU would that go in #2 and #3?

from redisai.

@andresrinivasan these would be specific enough to warrant separate containers IMHO

from redisai.

Closed by #159

from redisai.

Let's keep the comments in mind for packaging

from redisai.