[Longformer, Bert, Roberta, ...] Fix multi gpu training (#7272)

* fix multi-gpu

* fix longformer

* force to delete unnecessary layers

* fix notifications

* fix warning

* fix roberta

* fix tests

* remove hasattr

* fix tests

* fix roberta

* merge and clean authorized keys
This commit is contained in:
Patrick von Platen
2020-09-25 20:33:21 +02:00
committed by GitHub
parent 2c8ecdf8a8
commit e50a931c11
16 changed files with 179 additions and 49 deletions

View File

@@ -183,14 +183,14 @@ class AutoModelTest(unittest.TestCase):
def test_from_pretrained_identifier(self):
model = AutoModelWithLMHead.from_pretrained(SMALL_MODEL_IDENTIFIER)
self.assertIsInstance(model, BertForMaskedLM)
self.assertEqual(model.num_parameters(), 14830)
self.assertEqual(model.num_parameters(only_trainable=True), 14830)
self.assertEqual(model.num_parameters(), 14410)
self.assertEqual(model.num_parameters(only_trainable=True), 14410)
def test_from_identifier_from_model_type(self):
model = AutoModelWithLMHead.from_pretrained(DUMMY_UNKWOWN_IDENTIFIER)
self.assertIsInstance(model, RobertaForMaskedLM)
self.assertEqual(model.num_parameters(), 14830)
self.assertEqual(model.num_parameters(only_trainable=True), 14830)
self.assertEqual(model.num_parameters(), 14410)
self.assertEqual(model.num_parameters(only_trainable=True), 14410)
def test_parents_and_children_in_mappings(self):
# Test that the children are placed before the parents in the mappings, as the `instanceof` will be triggered