From 683b3e304af15660401928b0456464ce52673c28 Mon Sep 17 00:00:00 2001
From: Megvii Engine Team <megengine@megvii.com>
Date: Thu, 8 Dec 2022 19:11:05 +0800
Subject: [PATCH] fix(mge/quantization): disable creation of qint8 dtype with
 zeropoint

GitOrigin-RevId: d3b87c303f91ae333e8ba42c23180b9626836d39
---
 imperative/python/megengine/core/tensor/dtype.py  | 7 +++++++
 imperative/python/megengine/quantization/utils.py | 3 +++
 2 files changed, 10 insertions(+)

diff --git a/imperative/python/megengine/core/tensor/dtype.py b/imperative/python/megengine/core/tensor/dtype.py
index ea9eb3585..ba3d56959 100644
--- a/imperative/python/megengine/core/tensor/dtype.py
+++ b/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)}},
diff --git a/imperative/python/megengine/quantization/utils.py b/imperative/python/megengine/quantization/utils.py
index 00699d3fb..216c2172a 100644
--- a/imperative/python/megengine/quantization/utils.py
+++ b/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
-- 
GitLab