SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Fix CPU offload + disk offload tests (#27204)

Browse Source 
Fix disk offload tests + weight sharing issues
This commit is contained in:
Lysandre Debut
2023-11-01 19:25:23 +01:00
committed by GitHub
parent c9e72f55b2
commit 95020f208e
14 changed files with 155 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/models/vitdet/test_modeling_vitdet.py
View File

@@ -182,7 +182,11 @@ class VitDetModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
# TODO: Fix me (once this model gets more usage)
@unittest.skip("Does not work on the tiny model as we keep hitting edge cases.")
def test_disk_offload(self):
def test_disk_offload_bin(self):
super().test_disk_offload()
@unittest.skip("Does not work on the tiny model as we keep hitting edge cases.")
def test_disk_offload_safetensors(self):
super().test_disk_offload()
# TODO: Fix me (once this model gets more usage)

6
tests/models/whisper/test_modeling_whisper.py
View File

@@ -1788,7 +1788,11 @@ class WhisperEncoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.
pass
@unittest.skip(reason="Some undefined behavior encountered with tiny versions of this model. Skip for now.")
def test_disk_offload(self):
def test_disk_offload_bin(self):
pass
@unittest.skip(reason="Some undefined behavior encountered with tiny versions of this model. Skip for now.")
def test_disk_offload_safetensors(self):
pass
@unittest.skip(reason="Some undefined behavior encountered with tiny versions of this model. Skip for now.")

52
tests/test_modeling_common.py
View File

@@ -2578,7 +2578,45 @@ class ModelTesterMixin:
@require_accelerate
@mark.accelerate_tests
@require_torch_gpu
def test_disk_offload(self):
def test_disk_offload_bin(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
if model_class._no_split_modules is None:
continue
inputs_dict_class = self._prepare_for_class(inputs_dict, model_class)
model = model_class(config).eval()
model = model.to(torch_device)
torch.manual_seed(0)
base_output = model(**inputs_dict_class)
model_size = compute_module_sizes(model)[""]
with tempfile.TemporaryDirectory() as tmp_dir:
model.cpu().save_pretrained(tmp_dir, safe_serialization=False)
with self.assertRaises(ValueError):
max_size = int(self.model_split_percents[0] * model_size)
max_memory = {0: max_size, "cpu": max_size}
# This errors out cause it's missing an offload folder
new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
max_size = int(self.model_split_percents[1] * model_size)
max_memory = {0: max_size, "cpu": max_size}
new_model = model_class.from_pretrained(
tmp_dir, device_map="auto", max_memory=max_memory, offload_folder=tmp_dir
)
self.check_device_map_is_respected(new_model, new_model.hf_device_map)
torch.manual_seed(0)
new_output = new_model(**inputs_dict_class)
self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
@require_accelerate
@mark.accelerate_tests
@require_torch_gpu
def test_disk_offload_safetensors(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
@@ -2595,17 +2633,11 @@ class ModelTesterMixin:
with tempfile.TemporaryDirectory() as tmp_dir:
model.cpu().save_pretrained(tmp_dir)
with self.assertRaises(ValueError):
max_size = int(self.model_split_percents[0] * model_size)
max_memory = {0: max_size, "cpu": max_size}
# This errors out cause it's missing an offload folder
new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
max_size = int(self.model_split_percents[1] * model_size)
max_memory = {0: max_size, "cpu": max_size}
new_model = model_class.from_pretrained(
tmp_dir, device_map="auto", max_memory=max_memory, offload_folder=tmp_dir
)
# This doesn't error out as it's in safetensors and doesn't need an offload folder
new_model = model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
self.check_device_map_is_respected(new_model, new_model.hf_device_map)
torch.manual_seed(0)