fix conflicts and enable bias_shifting for input_maxes before tuning

Signed-off-by: Lu, Yintong <yintong.lu@intel.com>

neural_compressor/adaptor/pytorch.py

@@ -1880,9 +1880,17 @@ def qdq_quantize(self, model, tune_cfg):
                 input_power = torch.pow(abs_input_max, alpha)
                 weight_power = torch.pow(weight_max, 1 - alpha)
                 scale = torch.clip(input_power / weight_power, min=1e-5)
+                if 'os_bias' in info: #lyt_os_debug_0822
+                    bias_alpha = info['os_bias']
+                else:
+                    bias_alpha = None
                 for op_name in absorbed_layer:
                     module = fetch_module(q_model, op_name)
-                    new_module = SQLinearWrapper(module, 1.0 / scale, input_minmax, alpha)
+                    new_module = SQLinearWrapper(module, 1.0/scale, input_minmax, alpha, bias_alpha=bias_alpha, layer_name=_) #lyt_os_debug_0822
+                    if bias_alpha is not None:
+                        logger.info(f"lyt_debug qdq_quantize bias_wrapper {_}, {bias_alpha.size()}, layer_name: {_}, absorbed_layer: {absorbed_layer}")
+                    else:
+                        logger.info(f"lyt_debug qdq_quantize bias_wrapper {_}, None, layer_name: {_}, absorbed_layer: {absorbed_layer}")
                     set_module(q_model, op_name, new_module)
                     logger.debug(f"Current SmoothQuant alpha of {op_name} is {alpha}")
 

neural_compressor/adaptor/torch_utils/model_wrapper.py

@@ -106,11 +106,13 @@ def _wrap_lwq_layer(model, lwq_layers, op_cfgs):
 
 
 class SQLinearWrapper(torch.nn.Module):
-    def __init__(self, module, input_scale, input_minmax, alpha=0.5, dtype=torch.quint8):
+    def __init__(self, module, input_scale, input_minmax, alpha=0.5, dtype=torch.quint8, bias_alpha=None, layer_name=None): #lyt_os_debug
         super().__init__()
         self.register_buffer("input_scale", input_scale)
         self.alpha = alpha
         self.dtype = dtype
+        self.bias_alpha=bias_alpha #lyt_os_debug
+        self.layer_name = layer_name
         # calculate and only save scale, zero_point to avoid memory usage
         self.scale, self.zero_point = self._calculate_qparams(input_scale, input_minmax, dtype)
         self.add_module("sq_linear", module)
@@ -122,6 +124,8 @@ def weight(self):
         return self.sq_linear.weight
 
     def forward(self, X):
+        if self.bias_alpha != None:  #lyt_os_debug_0822
+            X = X - self.bias_alpha
         if self.ipex:
             X = self.sq_linear(X)
         else:
@@ -157,7 +161,20 @@ def _update_sq_linear(self):
         # remove mul and reset sq_linear for ipex inference
         scale = self.input_scale.view(1, self.input_scale.shape[0])
         with torch.no_grad():
+            import copy #lyt_os_debug
+            layer_weight = copy.deepcopy(self.sq_linear.weight)  #lyt_os_debug
             self.sq_linear.weight /= scale
+            if self.bias_alpha != None: #lyt_os_debug_0822:
+                res = torch.matmul(self.bias_alpha, layer_weight.transpose(0, 1))
+                layer_bias = copy.deepcopy(self.sq_linear.bias)
+                if self.sq_linear.bias != None:
+                    bias_tmp = self.sq_linear.bias + res
+                    self.sq_linear.bias.data.copy_(bias_tmp)
+                    logger.info(f"lyt_debug bias shifted AP: {self.layer_name} {self.sq_linear.bias.data.size()}, {torch.all(layer_bias.data == self.sq_linear.bias.data)}")
+                else:
+                    self.sq_linear.bias = torch.nn.Parameter(res)
+                    logger.info(f"lyt_debug bias created AP: {self.layer_name} {self.sq_linear.bias.data.size()}, {layer_bias==None}")
+
 
     def _recover_sq_linear(self):
         # remove mul and reset sq_linear for ipex inference