Specific int8 operators for conv and fully connected?

[mikhilg10]

In basic gguf format are all the weights and activations are being processed as fp16 or there are separate operators for int8 and other formats?. Operators like convolutional , fully connected layer and ffn??
What I understood from ggml that quantization is storage purposes only. While inference it'll convert qX to fp16 or 32. Is my understanding wrong??
And
If I want to add the int8 specific operators how to do that. Can anyone help me???

[FSSRepo]

Quantization is mostly used to reduce the size of the weights that need to be multiplied (matrix multiplication ggml_mul_mat); the rest of the operations are performed in fp16 or fp32, depending on the case.

[mikhilg10]

so the activations are used as it is in fp32 and fp16??