Why is the preserved model different from the pyTorch model after the quant aware training

[yangyyt]

quant aware model:
pythorch:
layer_0.scale
layer_0.zero_point
layer_0._packed_params.dtype
layer_0._packed_params._packed_params
layer_1.scale
layer_1.zero_point
layer_1._packed_params.dtype
layer_1._packed_params._packed_params
layer_end.scale
layer_end.zero_point
layer_end._packed_params.dtype
layer_end._packed_params._packed_params
quant.scale
quant.zero_point

but pytorch lightning:
layer_0.weight
layer_0.bias
layer_0.activation_post_process.fake_quant_enabled
layer_0.activation_post_process.observer_enabled
layer_0.activation_post_process.scale
layer_0.activation_post_process.zero_point
layer_0.activation_post_process.activation_post_process.min_val
layer_0.activation_post_process.activation_post_process.max_val
layer_0.weight_fake_quant.fake_quant_enabled
layer_0.weight_fake_quant.observer_enabled
layer_0.weight_fake_quant.scale
layer_0.weight_fake_quant.zero_point
layer_0.weight_fake_quant.activation_post_process.min_vals
layer_0.weight_fake_quant.activation_post_process.max_vals
layer_0a.activation_post_process.fake_quant_enabled
layer_0a.activation_post_process.observer_enabled
layer_0a.activation_post_process.scale
layer_0a.activation_post_process.zero_point
layer_0a.activation_post_process.activation_post_process.min_val
layer_0a.activation_post_process.activation_post_process.max_val
layer_1.weight
layer_1.bias
layer_1.activation_post_process.fake_quant_enabled
layer_1.activation_post_process.observer_enabled
layer_1.activation_post_process.scale
layer_1.activation_post_process.zero_point
layer_1.activation_post_process.activation_post_process.min_val
layer_1.activation_post_process.activation_post_process.max_val
layer_1.weight_fake_quant.fake_quant_enabled
layer_1.weight_fake_quant.observer_enabled
layer_1.weight_fake_quant.scale
layer_1.weight_fake_quant.zero_point
layer_1.weight_fake_quant.activation_post_process.min_vals
layer_1.weight_fake_quant.activation_post_process.max_vals
layer_1a.activation_post_process.fake_quant_enabled
layer_1a.activation_post_process.observer_enabled
layer_1a.activation_post_process.scale
layer_1a.activation_post_process.zero_point
layer_1a.activation_post_process.activation_post_process.min_val
layer_1a.activation_post_process.activation_post_process.max_val
layer_end.weight
layer_end.bias
layer_end.activation_post_process.fake_quant_enabled
layer_end.activation_post_process.observer_enabled
layer_end.activation_post_process.scale
layer_end.activation_post_process.zero_point
layer_end.activation_post_process.activation_post_process.min_val
layer_end.activation_post_process.activation_post_process.max_val
layer_end.weight_fake_quant.fake_quant_enabled
layer_end.weight_fake_quant.observer_enabled
layer_end.weight_fake_quant.scale
layer_end.weight_fake_quant.zero_point
layer_end.weight_fake_quant.activation_post_process.min_vals
layer_end.weight_fake_quant.activation_post_process.max_vals
quant.activation_post_process.fake_quant_enabled
quant.activation_post_process.observer_enabled
quant.activation_post_process.scale
quant.activation_post_process.zero_point
quant.activation_post_process.activation_post_process.min_val
quant.activation_post_process.activation_post_process.max_val

[yangyyt]

How can I save a qat model like pyTorch?

[tchaton]

@Borda