Does tvm support quantization


#1

All the models currently are executed for float32 datatype. Are any quantized models(say INT8) also supported, if yes please brief me about them.


#2

You mean suppor existing INT8 quantized model (for example TFLite) or support quantize FP32 model?


#3

If you could answer both the questions then it would be great. About the first one I found out that work is in progress for existing INT8 quantized model am I right?
Please comment on the second one also.


#4

Yes. The fisrt one we are working in progress.
The second one is also in progress. https://github.com/dmlc/tvm/pull/2116 is PR for second one you want to know.


#5

Thankyou very much @FrozenGene !!