Rebase

Author: Chun-I Tsai

backends/qualcomm/quantizer/quantizer.py

@@ -153,11 +153,6 @@ def __post_init__(self):
             if self.act_observer
             else per_channel_quant_config_func()
         )
-        self.per_block_quant_config = (
-            per_block_quant_config_func(act_observer=act_observer)
-            if self.act_observer
-            else per_block_quant_config_func()
-        )
         self.use_per_channel_weight_quant_ops = set()
         if self.is_conv_per_channel:
             self.use_per_channel_weight_quant_ops.update(
@@ -173,6 +168,12 @@ def __post_init__(self):
                     torch.ops.aten.linear.default,
                 }
             )
+        if per_block_quant_config_func:
+            self.per_block_quant_config = (
+                per_block_quant_config_func(act_observer=self.act_observer)
+                if self.act_observer
+                else per_block_quant_config_func()
+            )
 
 
 class QnnQuantizer(Quantizer):

backends/qualcomm/tests/test_qnn_delegate.py

@@ -1238,7 +1238,6 @@ def test_qnn_backend_conv2d_block(self):
                 module = self.get_qdq_module(
                     module,
                     sample_input,
-                    is_conv_per_block=True,
                     quant_dtype=QuantDtype.use_16a4w_block,
                     block_size_map={"conv2d": (1, 128, 1, 1)},
                 )