Fix npu group size setting of optimize_model=False (#12256)

python/llm/src/ipex_llm/transformers/npu_model.py

@@ -198,7 +198,8 @@ class _BaseAutoModelClass:
             from ipex_llm.transformers.npu_models.convert import optimize_llm
             optimize_llm(model)
             with torch.no_grad():
-                cls.load_convert(qtype, model, "cpu", modules_to_not_convert, *args, **kwargs)
+                cls.load_convert(qtype, model, "cpu", modules_to_not_convert,
+                                 quantization_group_size, *args, **kwargs)
                 if hasattr(model, "llm"):
                     create_npu_kernels(model.llm)
                 else:

python/llm/src/ipex_llm/transformers/npu_models/linear.py

@@ -130,7 +130,7 @@ class QuantizedLinear(torch.nn.Module):
         weight: torch.Tensor,
         scale: torch.Tensor,
         bias: Optional[torch.Tensor] = None,
-        group_size: int = False,
+        group_size: int = 0,
     ):
         """Initialize the QuantizedLinear class.