[Neural Speed] Fix a blocker on Windows platforms

scripts/inference.py

@@ -18,15 +18,12 @@
 from typing import List, Optional
 import subprocess
 from transformers import AutoTokenizer
+import neural_speed
 
 model_maps = {"gpt_neox": "gptneox", "llama2": "llama", "gpt_bigcode": "starcoder"}
 build_path = Path(Path(__file__).parent.absolute(), "../build/")
 
 
-def is_win():
-    return sys.platform.startswith('win')
-
-
 def main(args_in: Optional[List[str]] = None) -> None:
     parser = argparse.ArgumentParser(description="main program llm running")
     parser.add_argument("--model_name", type=str, help="Model name: String", required=True)
@@ -130,19 +127,8 @@ def main(args_in: Optional[List[str]] = None) -> None:
     args = parser.parse_args(args_in)
     print(args)
     model_name = model_maps.get(args.model_name, args.model_name)
-    if is_win():
-        path = Path(args.build_dir, "./Bin/Release/run_{}.exe".format(model_name))
-    else:
-        if args.one_click_run == "True":
-            import neural_speed
-            package_path = os.path.dirname(neural_speed.__file__)
-            path = Path(package_path, "./run_{}".format(model_name))
-        else:
-            path = Path(args.build_dir, "./bin/run_{}".format(model_name))
-
-    if not path.exists():
-        print("Please build graph first or select the correct model name.")
-        sys.exit(1)
+    package_path = os.path.dirname(neural_speed.__file__)
+    path = Path(package_path, "./run_{}".format(model_name))
 
     cmd = [path]
     cmd.extend(["--model", args.model])

scripts/quantize.py

@@ -17,15 +17,12 @@
 import argparse
 from typing import List, Optional
 import subprocess
+import neural_speed
 
 model_maps = {"gpt_neox": "gptneox", "llama2": "llama", "gpt_bigcode": "starcoder"}
 build_path = Path(Path(__file__).parent.absolute(), "../build/")
 
 
-def is_win():
-    return sys.platform.startswith('win')
-
-
 def str2bool(v):
     if isinstance(v, bool):
         return v
@@ -100,19 +97,8 @@ def main(args_in: Optional[List[str]] = None) -> None:
     args = parser.parse_args(args_in)
 
     model_name = model_maps.get(args.model_name, args.model_name)
-    if is_win():
-        path = Path(args.build_dir, "./Bin/Release/quant_{}.exe".format(model_name))
-    else:
-        if args.one_click_run == "True":
-            import neural_speed
-            package_path = os.path.dirname(neural_speed.__file__)
-            path = Path(package_path, "./quant_{}".format(model_name))
-        else:
-            path = Path(args.build_dir, "./bin/quant_{}".format(model_name))
-    if not path.exists():
-        print(path)
-        print("Please build graph first or select the correct model name.")
-        sys.exit(1)
+    package_path = os.path.dirname(neural_speed.__file__)
+    path = Path(package_path, "./quant_{}".format(model_name))
 
     cmd = [path]
     cmd.extend(["--model_file", args.model_file])