Revert "Support Fake GroupWise Quant (PaddlePaddle#61900)"

This reverts commit 2175de0.

python/paddle/nn/quant/format.py

@@ -46,14 +46,7 @@ def from_quanter(quanter):
 
 
 class LinearQuanter(Layer):
-    def __init__(
-        self,
-        scales,
-        zero_point=None,
-        quant_axis=None,
-        bit_length=8,
-        group_size=128,
-    ):
+    def __init__(self, scales, zero_point=None, quant_axis=None, bit_length=8):
         super().__init__()
         scales = paddle.to_tensor(scales, dtype="float32")
         scale_attr = paddle.framework.ParamAttr(
@@ -72,21 +65,9 @@ def __init__(
         )
         self._quant_axis = -1 if quant_axis is None else quant_axis
         self._bit_length = bit_length
-        self._group_size = group_size
 
     def forward(self, input):
         if in_dynamic_mode():
-            if len(self._scales.shape) > 1:
-                bnt = (1 << (self._bit_length - 1)) - 1
-                new_s = paddle.repeat_interleave(
-                    self._scales, self._group_size, 0
-                )
-                quant_weight = paddle.clip(
-                    paddle.round(input.cast('float32') / new_s * bnt),
-                    -bnt - 1,
-                    bnt,
-                )
-                return quant_weight.cast(input.dtype)
             return _C_ops.quantize_linear(
                 input.cast('float32'),
                 self._scales,
@@ -124,14 +105,7 @@ def from_quanter(quanter):
 
 
 class LinearDequanter(Layer):
-    def __init__(
-        self,
-        scales,
-        zero_point=None,
-        quant_axis=None,
-        bit_length=8,
-        group_size=128,
-    ):
+    def __init__(self, scales, zero_point=None, quant_axis=None, bit_length=8):
         super().__init__()
         scales = paddle.to_tensor(scales, dtype="float32")
         scale_attr = paddle.framework.ParamAttr(
@@ -150,18 +124,9 @@ def __init__(
         )
         self._quant_axis = -1 if quant_axis is None else quant_axis
         self._bit_length = bit_length
-        self._group_size = group_size
 
     def forward(self, input):
         if in_dynamic_mode():
-            if len(self._scales.shape) > 1:
-                bnt = (1 << (self._bit_length - 1)) - 1
-                new_s = paddle.repeat_interleave(
-                    self._scales, self._group_size, 0
-                )
-                quant_dequant_weight = input.cast('float32') / bnt * new_s
-                return quant_dequant_weight.cast(input.dtype)
-
             return _C_ops.dequantize_linear(
                 input.cast('float32'),
                 self._scales,

python/paddle/quantization/observers/__init__.py

@@ -14,6 +14,5 @@
 # limitations under the License.
 
 from .abs_max import AbsmaxObserver
-from .groupwise import GroupWiseWeightObserver
 
-__all__ = ["AbsmaxObserver", "GroupWiseWeightObserver"]
+__all__ = ["AbsmaxObserver"]