Precision control (#126)

* macos12 full build (x86)

* add support for setting precision via --dtype

* setup default dtype as float32

* setup default dtype as float32

Author: mikekgfb

export.py

@@ -167,7 +167,7 @@ def cli():
     parser.add_argument(
         "-d",
         "--dtype",
-        default=None,
+        default="float32",
         help="Override the dtype of the model (default is the checkpoint dtype). Options: bf16, fp16, fp32",
     )
     parser.add_argument("-v", "--verbose", action="store_true")

generate.py

@@ -603,7 +603,7 @@ def cli():
     parser.add_argument(
         "-d",
         "--dtype",
-        default=None,
+        default="float32",
         help="Override the dtype of the model (default is the checkpoint dtype). Options: bf16, fp16, fp32",
     )
 

quantize.py

@@ -37,7 +37,7 @@ def name_to_dtype(name):
     if name in name_to_dtype_dict:
         return name_to_dtype_dict[name]
     else:
-        raise RuntimeError("unsupported dtype specified")
+        raise RuntimeError(f"unsupported dtype name {name} specified")
 
 name_to_dtype_dict = {
     "fp32" : torch.float,