diff --git a/language/llama2-70b/ci_file/qllama2_70b_forward_test.py b/language/llama2-70b/ci_file/qllama2_70b_forward_test.py
index 15d368c444..8581976cf1 100644
--- a/language/llama2-70b/ci_file/qllama2_70b_forward_test.py
+++ b/language/llama2-70b/ci_file/qllama2_70b_forward_test.py
@@ -7,6 +7,7 @@
 import furiosa_llm_models
 import joblib
 import model_compressor
+import model_compressor_impl
 import torch
 import yaml
 from torch.nn.functional import pad
@@ -114,7 +115,7 @@ def obtain_traced_model_dict(model):
             prefill_model,
             prefill_input_names,
             prefill_concrete_args,
-        ) = model_compressor.helper.llama_custom_symbolic_trace(
+        ) = model_compressor_impl.helper.llama_custom_symbolic_trace(
             model,
             input_names=["input_ids", "attention_mask", "position_ids"],
             disable_check=True,
@@ -123,7 +124,7 @@ def obtain_traced_model_dict(model):
             decode_model,
             decode_input_names,
             decode_concrete_args,
-        ) = model_compressor.helper.llama_custom_symbolic_trace(
+        ) = model_compressor_impl.helper.llama_custom_symbolic_trace(
             model,
             input_names=[
                 "input_ids",
@@ -228,7 +229,7 @@ def get_generator_for_golden_model(
         "decode_model": quant_golden_models["decode"],
     }
 
-    return model_compressor.helper.QuantCausalLM(
+    return model_compressor_impl.helper.QuantCausalLM(
         quant_golden_models, golden_model_type, golden_input_names, golden_concrete_args
     )
 
@@ -315,7 +316,7 @@ def perform_generation(
     generation_output_dictionary = dict()
     with torch.no_grad():
         for idx, test_data in enumerate(test_data_list):
-            if type(generator) == model_compressor.helper.QuantCausalLM:
+            if type(generator) == model_compressor_impl.helper.QuantCausalLM:
                 output = generator.generate(**test_data, **gen_kwargs)
             elif type(generator) == MLPerfSubmissionGreedySearch:
                 # mlperf submission