Fix g_idx init in transformers-like API (#2204)

Signed-off-by: Kaihui-intel <[email protected]>

Author: Kaihui-intel

neural_compressor/transformers/models/modeling_auto.py

@@ -70,6 +70,7 @@ 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(
@@ -80,7 +81,7 @@ def build_woq_model(model, quantization_config):
                     group_size=quantization_config.group_size,
                     zp=zp,
                     bias=m.bias is not None,
-                    g_idx=True,
+                    g_idx=g_idx,
                     use_optimum_format=use_optimum_format,
                 )
             set_module(model, n, new_module)

neural_compressor/transformers/quantization/utils.py

@@ -206,14 +206,8 @@ def _replace_linear(
                         device=device,
                         use_optimum_format=getattr(module, "use_optimum_format", True),
                     )
-                    if quantization_config.quant_method.value == "gptq":
-                        g_idx = getattr(
-                            module,
-                            "g_idx",
-                            torch.zeros(in_features, dtype=torch.int32).to(device),
-                        )
-                    else:
-                        g_idx = None
+                    # g_idx is only present when using GPTQ quantization method
+                    g_idx = module.g_idx if hasattr(module, "g_idx") else None
                     model._modules[name].set_scales_zps_gidx(
                         (
                             module.scales