Fix all offload and MP tests (#17533)

src/transformers/modeling_utils.py

@@ -574,7 +574,6 @@ def _load_state_dict_into_meta_model(
     for param_name, param in state_dict.items():
         # First part of the test is always true as load_state_dict_keys always contains state_dict keys.
         if param_name not in loaded_state_dict_keys or param_name not in expected_keys:
-            print(param_name)
             continue
 
         if param_name.startswith(start_prefix):
@@ -2124,6 +2123,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             if model._no_split_modules is None:
                 raise ValueError(f"{model.__class__.__name__} does not support `device_map='auto'` yet.")
             no_split_modules = model._no_split_modules
+            # Make sure tied weights are tied before creating the device map.
+            model.tie_weights()
             device_map = infer_auto_device_map(
                 model, no_split_module_classes=no_split_modules, dtype=torch_dtype, max_memory=max_memory
             )

tests/models/opt/test_modeling_opt.py

@@ -63,7 +63,7 @@ class OPTModelTester:
         use_labels=False,
         vocab_size=99,
         hidden_size=16,
-        num_hidden_layers=2,
+        num_hidden_layers=5,
         num_attention_heads=4,
         intermediate_size=4,
         hidden_act="gelu",

tests/models/t5/test_modeling_t5.py

@@ -515,6 +515,8 @@ class T5ModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
     test_resize_embeddings = True
     test_model_parallel = True
     is_encoder_decoder = True
+    # The small T5 model needs higher percentages for CPU/MP tests
+    model_split_percents = [0.8, 0.9]
 
     def setUp(self):
         self.model_tester = T5ModelTester(self)

tests/test_modeling_common.py

@@ -153,6 +153,7 @@ class ModelTesterMixin:
     test_model_parallel = False
     is_encoder_decoder = False
     has_attentions = True
+    model_split_percents = [0.5, 0.7, 0.9]
 
     def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
         inputs_dict = copy.deepcopy(inputs_dict)
@@ -2217,12 +2218,7 @@ class ModelTesterMixin:
     @require_accelerate
     @require_torch_gpu
     def test_disk_offload(self):
-        if all([model_class._no_split_modules is None for model_class in self.all_model_classes]):
-            return
-
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        if isinstance(getattr(config, "num_hidden_layers", None), int) and config.num_hidden_layers < 4:
-            config.num_hidden_layers = 4
 
         for model_class in self.all_model_classes:
             if model_class._no_split_modules is None:
@@ -2234,8 +2230,7 @@ class ModelTesterMixin:
             base_output = model(**inputs_dict)
 
             model_size = compute_module_sizes(model)[""]
-            # We test several splits of sizes to make sure it works.
-            max_size = int(0.4 * model_size)
+            max_size = int(self.model_split_percents[0] * model_size)
             with tempfile.TemporaryDirectory() as tmp_dir:
                 model.cpu().save_pretrained(tmp_dir)
 
@@ -2256,12 +2251,7 @@ class ModelTesterMixin:
     @require_accelerate
     @require_torch_gpu
     def test_cpu_offload(self):
-        if all([model_class._no_split_modules is None for model_class in self.all_model_classes]):
-            return
-
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        if isinstance(getattr(config, "num_hidden_layers", None), int) and config.num_hidden_layers < 4:
-            config.num_hidden_layers = 4
 
         for model_class in self.all_model_classes:
             if model_class._no_split_modules is None:
@@ -2274,7 +2264,7 @@ class ModelTesterMixin:
 
             model_size = compute_module_sizes(model)[""]
             # We test several splits of sizes to make sure it works.
-            max_gpu_sizes = [int(p * model_size) for p in [0.5, 0.7, 0.9]]
+            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents]
             with tempfile.TemporaryDirectory() as tmp_dir:
                 model.cpu().save_pretrained(tmp_dir)
 
@@ -2292,12 +2282,7 @@ class ModelTesterMixin:
     @require_accelerate
     @require_torch_multi_gpu
     def test_model_parallelism(self):
-        if all([model_class._no_split_modules is None for model_class in self.all_model_classes]):
-            return
-
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        if isinstance(getattr(config, "num_hidden_layers", None), int) and config.num_hidden_layers < 4:
-            config.num_hidden_layers = 4
 
         for model_class in self.all_model_classes:
             if model_class._no_split_modules is None:
@@ -2310,7 +2295,7 @@ class ModelTesterMixin:
 
             model_size = compute_module_sizes(model)[""]
             # We test several splits of sizes to make sure it works.
-            max_gpu_sizes = [int(p * model_size) for p in [0.5, 0.7, 0.9]]
+            max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents]
             with tempfile.TemporaryDirectory() as tmp_dir:
                 model.cpu().save_pretrained(tmp_dir)