Fix passing torch_dtype and device_map via model_init_kwargs

Signed-off-by: Thomas Parnell <[email protected]>

Author: tdoublep

auto_fp8/modeling.py

@@ -71,10 +71,10 @@ def skip(*args, **kwargs):
         torch.cuda.empty_cache()
 
         # Important defaults
-        if not hasattr(model_init_kwargs, "torch_dtype"):
+        if not "torch_dtype" in model_init_kwargs:
             model_init_kwargs["torch_dtype"] = "auto"
 
-        if not hasattr(model_init_kwargs, "device_map"):
+        if not "device_map" in model_init_kwargs:
             model_init_kwargs["device_map"] = "auto"
 
         merged_kwargs = {**model_init_kwargs, **cached_file_kwargs}