Use CPU if CUDA is not available (#18)

As it's reasonable, isn't it?

Author: malfet

.github/workflows/compile.yml

@@ -38,8 +38,8 @@ jobs:
         run: |          
           export MODEL_PATH=checkpoints/stories15M/stories15M.pt
           export MODEL_REPO=stories15M
-          python generate.py --checkpoint_path ${MODEL_PATH} --temperature 0 --device cpu | tee output_eager
-          python generate.py --compile --checkpoint_path ${MODEL_PATH} --temperature 0 --device cpu | tee output_compiled
+          python generate.py --checkpoint_path ${MODEL_PATH} --temperature 0 | tee output_eager
+          python generate.py --compile --checkpoint_path ${MODEL_PATH} --temperature 0 | tee output_compiled
           python aoti_export.py --checkpoint_path ${MODEL_PATH} --output_path ./${MODEL_REPO}.so
           python generate.py --checkpoint_path ${MODEL_PATH} --temperature 0 --dso ./${MODEL_REPO}.so | tee output_aoti
           # echo "******************************************"

generate.py

@@ -558,7 +558,7 @@ def cli():
         default=None,
         help="Draft checkpoint path.",
     )
-    parser.add_argument("--device", type=str, default="cuda", help="Device to use")
+    parser.add_argument("--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu", help="Device to use")
     parser.add_argument(
         "--dso_path",
         type=Path,