Fix g_idx init for GPTQ (#2222)

Signed-off-by: Kaihui-intel <[email protected]>
Co-authored-by: Wang, Chang <[email protected]>

Author: Kaihui-intel

neural_compressor/transformers/models/modeling_auto.py

@@ -60,6 +60,7 @@
 
 def build_woq_model(model, quantization_config):
     bits = quantization_config.bits
+    g_idx = getattr(quantization_config, "desc_act", False)
     for n, m in model.named_modules():
         if n in quantization_config.modules_to_not_convert:
             continue
@@ -70,7 +71,6 @@ def build_woq_model(model, quantization_config):
                 not getattr(quantization_config, "sym", False),
             )
             use_optimum_format = True
-            g_idx = hasattr(m, "g_idx") and m.g_idx is not None
 
             with init_empty_weights():
                 new_module = INCWeightOnlyLinear(