As I understand we do not have currently accumulation data type attribute for ops, right? We just seems need to introduce such one similar to dtype. The default value should match to dtype if it is defined and to type if dtype is not defined. The change of accumulator type is a matter of some additional logic this will define, for example, that in this certain model the third convolution should accumulate into fp32, for example to get good accuracy. If there is no such additional logic, the default behaviour (matching of accum to dtype and type) is good enough
--- [Visit Topic](https://discuss.tvm.apache.org/t/rfc-relay-fp32-fp16-model-support/9994/4) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.apache.org/email/unsubscribe/18a82f309ab9fe88d9b1536ae1919987612f9e0d80099fe29d2ddb9492eaaace).
