why the model size with 32-bit is equal to model with 4-bit

[xuzhiyuan1022]

No description provided.

[ChaofanTao]

Hi, during training, the parameters in both the trained full-precision model and quantized model are stored in float32 format, and there are limited possible choices of data in the quantized model.

The model size compresses when the quantized model converts the data format to INT during deployment. Also, the parameters of the proposed transform can be removed during deployment.

[xuzhiyuan1022]

thank you for your reply.

grad_alpha = (grad_output * (sign * i + (input_q - input) * (1 - i))).sum()

when update alpha,why use (input_q-input)

[ChaofanTao]

The (input_q-input) here considers the difference between input_q and input when updating alpha. That is an optional setting.