Fix python code scan warnings

Set CUDA on by default in CMakeLists.txt

CMakeLists.txt

@@ -19,7 +19,7 @@ check_language(CUDA)
 if (CMAKE_CUDA_COMPILER)
     set (CUDA_FOUND TRUE)
 endif()
-
+set(USE_CUDA 1)
 
 if("${CMAKE_C_COMPILER_ID}" STREQUAL "GNU" AND CMAKE_C_COMPILER_VERSION VERSION_LESS 9)
   message(FATAL_ERROR  "GCC version must be greater than or equal to 9")

download_llama.py

@@ -8,7 +8,7 @@
 def check_ort_import():
     try:
         import onnxruntime as ort
-    except:
+    except Exception as ex:
         return False
 
     return True
@@ -22,13 +22,13 @@ def get_models(args: argparse.Namespace):
         "-m",
         "onnxruntime.transformers.models.llama.convert_to_onnx",
         "-m",
-        "meta-llama/Llama-2-7b-hf",
+        "meta-llama/Llama-2-7b-chat-hf",
         "--output",
-        "llama2-7b-fp32-cpu",
+        "llama2-7b-chat-int4-gpu",
         "--precision",
-        "fp32",
+        "int4",
         "--execution_provider",
-        "cpu"
+        "cuda"
     ])
     # subprocess.run(["mv", "gpt2_parity_results.csv", args.output_folder])
 

download_test_model.py

@@ -7,7 +7,7 @@
 def check_ort_import():
     try:
         import onnxruntime as ort
-    except:
+    except Exception as ex:
         return False
 
     return True

src/pybind/llama_hl.py

@@ -1,5 +1,4 @@
 import onnxruntime_genai as og
-import numpy as np
 import time
 from transformers import LlamaTokenizer
 

src/pybind/run_llama_opt_ort.py

@@ -1,5 +1,4 @@
 import datetime
-import numpy as np
 from transformers import LlamaTokenizer
 import onnxruntime_genai as og
 

src/pybind/test_beam.py

@@ -1,5 +1,4 @@
 import onnxruntime_genai as og
-import numpy as np
 from transformers import GPT2Tokenizer
 
 device_type = og.DeviceType.CPU

src/pybind/test_greedy.py

@@ -1,5 +1,4 @@
 import onnxruntime_genai as og
-import numpy as np
 from transformers import GPT2Tokenizer
 
 device_type = og.DeviceType.CPU