feat(examples): don't use mps for float8 dtype.

huggingface · Sep 17, 2024 · ec1f85e · ec1f85e
1 parent ffae30c
commit ec1f85e
Showing 1 changed file with 2 additions and 5 deletions.
diff --git a/examples/vision/image-classification/pets/quantize_vit_model.py b/examples/vision/image-classification/pets/quantize_vit_model.py
@@ -37,7 +37,6 @@ def test(model, device, test_loader):
             data, target = batch["pixel_values"], batch["labels"]
             data, target = data.to(device), target.to(device)
             output = model(data).logits
-            # print("*****I am after output", output)
             if isinstance(output, QTensor):
                 output = output.dequantize()
             test_loss += F.nll_loss(output, target, reduction="sum").item()  # sum up batch loss
@@ -66,17 +65,15 @@ def main():
     parser.add_argument("--activations", type=str, default="int8", choices=["none", "int8", "float8"])
     args = parser.parse_args()
 
-    # torch.manual_seed(args.seed)
-
     dataset_kwargs = {}
 
     if args.device is None:
         if torch.cuda.is_available():
             device = torch.device("cuda")
             cuda_kwargs = {"num_workers": 1, "pin_memory": True, "shuffle": True}
             dataset_kwargs.update(cuda_kwargs)
-        elif torch.backends.mps.is_available():
-            device = torch.device("cpu")
+        elif all([torch.backends.mps.is_available(), args.weights != "float8", args.activations != "float8"]):
+            device = torch.device("mps")
         else:
             device = torch.device("cpu")
     else: