fix(mge/quantization): disable creation of qint8 dtype with zeropoint

GitOrigin-RevId: d3b87c303f91ae333e8ba42c23180b9626836d39

imperative/python/megengine/core/tensor/dtype.py

@@ -146,6 +146,13 @@ def create_quantized_dtype(
     else:
         # Don't trick to combine with is_unsigned. Metadata should not contain
         # invalid field to keep consistent with c dtype.
+        if zp is not None:
+            raise ValueError(
+                "'{}' only supports symmetric quantization, "
+                "zero_point should be None, but actually got '{}'".format(
+                    dtype_meta.name, zp
+                )
+            )
         return np.dtype(
             dtype_meta.np_dtype_str,
             metadata={"mgb_dtype": {"name": dtype_meta.cname, "scale": float(scale)}},

imperative/python/megengine/quantization/utils.py

@@ -210,6 +210,9 @@ def fake_quant_bias(bias: Tensor, inp: Tensor, w_qat: Tensor) -> Tensor:
         if inp_params.scale is None or w_params.scale is None:
             return b_qat
 
+        if inp_params.mode != QuantMode.SYMMERTIC:
+            return b_qat
+
         # TODO: support different mode
         if inp_params.mode != w_params.mode:
             return b_qat