Hi all, I’m interested in deploying pre-quantized networks, specially with 2-bit and 4-bit quantization with VTA but I know the usage for the accelerator is still restricted to 8bit and only certain models are supported. Nevertheless, the tech report and the hardware design state that it’s possible to have lower integer precision data types. My approach was to generate a custom bitstream with the desired data widths, considering always the buffer sizes but I still haven’t been able to come up with a configuration that works. As this issue was mentioned in the v0.6 roadmap but wasn’t addressed I thought I could get some further guidelines in order to bring this enhancement. Can anybody help me out?
Thanks in advance