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

[tests] remove flax-pt equivalence and cross tests (#36283)

Browse Source
This commit is contained in:
Joao Gante
2025-02-19 15:13:27 +00:00
committed by GitHub
parent fa8cdccd91
commit 99adc74462
39 changed files with 33 additions and 3103 deletions
Download Patch File Download Diff File Expand all files Collapse all files

112
tests/models/gpt2/test_modeling_flax_gpt2.py
View File

@@ -13,14 +13,12 @@
# limitations under the License.
import tempfile
import unittest
import numpy as np
import transformers
from transformers import GPT2Config, GPT2Tokenizer, is_flax_available, is_torch_available
from transformers.testing_utils import is_pt_flax_cross_test, require_flax, slow
from transformers import GPT2Config, GPT2Tokenizer, is_flax_available
from transformers.testing_utils import require_flax, slow
from ...test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
@@ -29,15 +27,8 @@ if is_flax_available():
import jax
import jax.numpy as jnp
from transformers.modeling_flax_pytorch_utils import (
convert_pytorch_state_dict_to_flax,
load_flax_weights_in_pytorch_model,
)
from transformers.models.gpt2.modeling_flax_gpt2 import FlaxGPT2LMHeadModel, FlaxGPT2Model
if is_torch_available():
import torch
class FlaxGPT2ModelTester:
def __init__(
@@ -255,105 +246,6 @@ class FlaxGPT2ModelTest(FlaxModelTesterMixin, unittest.TestCase):
self.assertListEqual(output_string, expected_string)
# overwrite from common since `attention_mask` in combination
# with `causal_mask` behaves slighly differently
@is_pt_flax_cross_test
def test_equivalence_pt_to_flax(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
# prepare inputs
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
pt_inputs = {k: torch.tensor(v.tolist()) for k, v in prepared_inputs_dict.items()}
# load corresponding PyTorch class
pt_model_class_name = model_class.__name__[4:] # Skip the "Flax" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
batch_size, seq_length = pt_inputs["input_ids"].shape
rnd_start_indices = np.random.randint(0, seq_length - 1, size=(batch_size,))
for batch_idx, start_index in enumerate(rnd_start_indices):
pt_inputs["attention_mask"][batch_idx, :start_index] = 0
pt_inputs["attention_mask"][batch_idx, start_index:] = 1
prepared_inputs_dict["attention_mask"][batch_idx, :start_index] = 0
prepared_inputs_dict["attention_mask"][batch_idx, start_index:] = 1
pt_model = pt_model_class(config).eval()
fx_model = model_class(config, dtype=jnp.float32)
fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
fx_model.params = fx_state
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs, pt_outputs):
self.assert_almost_equals(fx_output[:, -1], pt_output[:, -1].numpy(), 4e-2)
with tempfile.TemporaryDirectory() as tmpdirname:
pt_model.save_pretrained(tmpdirname)
fx_model_loaded = model_class.from_pretrained(tmpdirname, from_pt=True)
fx_outputs_loaded = fx_model_loaded(**prepared_inputs_dict).to_tuple()
self.assertEqual(
len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch"
)
for fx_output_loaded, pt_output in zip(fx_outputs_loaded, pt_outputs):
self.assert_almost_equals(fx_output_loaded[:, -1], pt_output[:, -1].numpy(), 4e-2)
# overwrite from common since `attention_mask` in combination
# with `causal_mask` behaves slighly differently
@is_pt_flax_cross_test
def test_equivalence_flax_to_pt(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
with self.subTest(model_class.__name__):
# prepare inputs
prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
pt_inputs = {k: torch.tensor(v.tolist()) for k, v in prepared_inputs_dict.items()}
# load corresponding PyTorch class
pt_model_class_name = model_class.__name__[4:] # Skip the "Flax" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
pt_model = pt_model_class(config).eval()
fx_model = model_class(config, dtype=jnp.float32)
pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)
batch_size, seq_length = pt_inputs["input_ids"].shape
rnd_start_indices = np.random.randint(0, seq_length - 1, size=(batch_size,))
for batch_idx, start_index in enumerate(rnd_start_indices):
pt_inputs["attention_mask"][batch_idx, :start_index] = 0
pt_inputs["attention_mask"][batch_idx, start_index:] = 1
prepared_inputs_dict["attention_mask"][batch_idx, :start_index] = 0
prepared_inputs_dict["attention_mask"][batch_idx, start_index:] = 1
# make sure weights are tied in PyTorch
pt_model.tie_weights()
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs).to_tuple()
fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
for fx_output, pt_output in zip(fx_outputs, pt_outputs):
self.assert_almost_equals(fx_output[:, -1], pt_output[:, -1].numpy(), 4e-2)
with tempfile.TemporaryDirectory() as tmpdirname:
fx_model.save_pretrained(tmpdirname)
pt_model_loaded = pt_model_class.from_pretrained(tmpdirname, from_flax=True)
with torch.no_grad():
pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
self.assertEqual(
len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch"
)
for fx_output, pt_output in zip(fx_outputs, pt_outputs_loaded):
self.assert_almost_equals(fx_output[:, -1], pt_output[:, -1].numpy(), 4e-2)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes: