Pytorch - Lazy initialization of models (#11471)

* lazy_init_weights

* remove ipdb

* save int

* add necessary code

* remove unnecessary utils

* Update src/transformers/models/t5/modeling_t5.py

* clean

* add tests

* correct

* finish tests

* finish tests

* fix some more tests

* fix xlnet & transfo-xl

* fix more tests

* make sure tests are independent

* fix tests more

* finist tests

* final touches

* Update src/transformers/modeling_utils.py

* Apply suggestions from code review

* Update src/transformers/modeling_utils.py

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* Update src/transformers/modeling_utils.py

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* clean tests

* give arg positive name

* add more mock weights to xlnet

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

tests/test_modeling_xlnet.py

@@ -594,6 +594,18 @@ class XLNetModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase)
         # xlnet cannot keep gradients in attentions or hidden states
         return
 
+    # overwrite from test_modeling_common
+    def _mock_init_weights(self, module):
+        if hasattr(module, "weight") and module.weight is not None:
+            module.weight.data.fill_(3)
+        if hasattr(module, "bias") and module.bias is not None:
+            module.bias.data.fill_(3)
+
+        for param in ["q", "k", "v", "o", "r", "r_r_bias", "r_s_bias", "r_w_bias", "seg_embed", "mask_emb"]:
+            if hasattr(module, param) and getattr(module, param) is not None:
+                weight = getattr(module, param)
+                weight.data.fill_(3)
+
     def _check_hidden_states_for_generate(
         self, batch_size, hidden_states, min_length, max_length, config, use_cache=False, num_beam_groups=1
     ):