feat(mge/quantization): add cambricon-quantization-example

GitOrigin-RevId: ed578ca92b

python_module/megengine/module/qat/module.py

@@ -52,7 +52,7 @@ class QATModule(Module):
             self.weight_fake_quant = safe_call(qconfig.weight_fake_quant)
 
     def _enable_exec(self, with_module, func, enable):
-        if not with_module:
+        if not with_module or not func:
             return
         if enable:
             func.enable()

python_module/megengine/module/quantized/linear.py

@@ -32,11 +32,13 @@ class Linear(QuantizedModule):
         inp_scale = mgb.dtype.get_scale(inp.dtype)
         w_scale = mgb.dtype.get_scale(self.weight.dtype)
         bias_dtype = mgb.dtype.qint32(inp_scale * w_scale)
-        return F.linear(
+        ret = F.linear(
             inp,
             self.weight,
             None if self.bias is None else self.bias.astype(bias_dtype),
-        ).astype(self.output_dtype)
+        )
+        ret = ret if self.output_dtype is None else ret.astype(self.output_dtype)
+        return ret
 
     @classmethod
     def from_qat_module(cls, qat_module: QAT.Linear):