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

remove ipex_optimize_model usage (#38632)

Browse Source 
* remove ipex_optimize_model usage

Signed-off-by: YAO Matrix <matrix.yao@intel.com>

* update Dockerfile

Signed-off-by: root <root@a4bf01945cfe.jf.intel.com>

---------

Signed-off-by: YAO Matrix <matrix.yao@intel.com>
Signed-off-by: root <root@a4bf01945cfe.jf.intel.com>
Co-authored-by: root <root@a4bf01945cfe.jf.intel.com>
This commit is contained in:
Yao Matrix
2025-06-07 02:04:44 +08:00
committed by GitHub
parent 5009252a05
commit dc76eff12b
7 changed files with 10 additions and 231 deletions
Download Patch File Download Diff File Expand all files Collapse all files

146
tests/trainer/test_trainer.py
View File

@@ -79,7 +79,6 @@ from transformers.testing_utils import (
require_deepspeed,
require_galore_torch,
require_grokadamw,
require_intel_extension_for_pytorch,
require_liger_kernel,
require_lomo,
require_non_hpu,
@@ -1325,37 +1324,6 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
train_output = trainer.train()
self.assertEqual(train_output.global_step, 10)
@require_torch_bf16
@require_intel_extension_for_pytorch
def test_number_of_steps_in_training_with_ipex(self):
for mix_bf16 in [True, False]:
tmp_dir = self.get_auto_remove_tmp_dir()
# Regular training has n_epochs * len(train_dl) steps
trainer = get_regression_trainer(
learning_rate=0.1, use_ipex=True, bf16=mix_bf16, use_cpu=True, output_dir=tmp_dir
)
train_output = trainer.train()
self.assertEqual(train_output.global_step, self.n_epochs * 64 / trainer.args.train_batch_size)
# Check passing num_train_epochs works (and a float version too):
trainer = get_regression_trainer(
learning_rate=0.1,
num_train_epochs=1.5,
use_ipex=True,
bf16=mix_bf16,
use_cpu=True,
output_dir=tmp_dir,
)
train_output = trainer.train()
self.assertEqual(train_output.global_step, int(1.5 * 64 / trainer.args.train_batch_size))
# If we pass a max_steps, num_train_epochs is ignored
trainer = get_regression_trainer(
learning_rate=0.1, max_steps=10, use_ipex=True, bf16=mix_bf16, use_cpu=True, output_dir=tmp_dir
)
train_output = trainer.train()
self.assertEqual(train_output.global_step, 10)
def test_torch_compile_loss_func_compatibility(self):
config = LlamaConfig(vocab_size=100, hidden_size=32, num_hidden_layers=3, num_attention_heads=4)
tiny_llama = LlamaForCausalLM(config)
@@ -2628,69 +2596,6 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
expected_acc = AlmostAccuracy()((pred + 1, y))["accuracy"]
self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
@require_torch_bf16
@require_intel_extension_for_pytorch
def test_evaluate_with_ipex(self):
for mix_bf16 in [True, False]:
with tempfile.TemporaryDirectory() as tmp_dir:
trainer = get_regression_trainer(
a=1.5,
b=2.5,
use_ipex=True,
compute_metrics=AlmostAccuracy(),
bf16=mix_bf16,
use_cpu=True,
output_dir=tmp_dir,
)
results = trainer.evaluate()
x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
pred = 1.5 * x + 2.5
expected_loss = ((pred - y) ** 2).mean()
self.assertAlmostEqual(results["eval_loss"], expected_loss)
expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
# With a number of elements not a round multiple of the batch size
trainer = get_regression_trainer(
a=1.5,
b=2.5,
use_ipex=True,
eval_len=66,
compute_metrics=AlmostAccuracy(),
bf16=mix_bf16,
use_cpu=True,
output_dir=tmp_dir,
)
results = trainer.evaluate()
x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
pred = 1.5 * x + 2.5
expected_loss = ((pred - y) ** 2).mean()
self.assertAlmostEqual(results["eval_loss"], expected_loss)
expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
# With logits preprocess
trainer = get_regression_trainer(
a=1.5,
b=2.5,
use_ipex=True,
compute_metrics=AlmostAccuracy(),
preprocess_logits_for_metrics=lambda logits, labels: logits + 1,
bf16=mix_bf16,
use_cpu=True,
output_dir=tmp_dir,
)
results = trainer.evaluate()
x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
pred = 1.5 * x + 2.5
expected_loss = ((pred - y) ** 2).mean()
self.assertAlmostEqual(results["eval_loss"], expected_loss)
expected_acc = AlmostAccuracy()((pred + 1, y))["accuracy"]
self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
def test_predict(self):
with tempfile.TemporaryDirectory() as tmp_dir:
trainer = get_regression_trainer(a=1.5, b=2.5, output_dir=tmp_dir)
@@ -2830,57 +2735,6 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
self.assertTrue(np.array_equal(labels[0], trainer.eval_dataset.ys[0]))
self.assertTrue(np.array_equal(labels[1], trainer.eval_dataset.ys[1]))
@require_torch_bf16
@require_intel_extension_for_pytorch
def test_predict_with_ipex(self):
for mix_bf16 in [True, False]:
with tempfile.TemporaryDirectory() as tmp_dir:
trainer = get_regression_trainer(
a=1.5, b=2.5, use_ipex=True, bf16=mix_bf16, use_cpu=True, output_dir=tmp_dir
)
preds = trainer.predict(trainer.eval_dataset).predictions
x = trainer.eval_dataset.x
self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))
# With a number of elements not a round multiple of the batch size
trainer = get_regression_trainer(
a=1.5, b=2.5, eval_len=66, use_ipex=True, bf16=mix_bf16, use_cpu=True, output_dir=tmp_dir
)
preds = trainer.predict(trainer.eval_dataset).predictions
x = trainer.eval_dataset.x
self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))
# With more than one output of the model
trainer = get_regression_trainer(
a=1.5, b=2.5, double_output=True, use_ipex=True, bf16=mix_bf16, use_cpu=True, output_dir=tmp_dir
)
preds = trainer.predict(trainer.eval_dataset).predictions
x = trainer.eval_dataset.x
self.assertEqual(len(preds), 2)
self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))
# With more than one output/label of the model
trainer = get_regression_trainer(
a=1.5,
b=2.5,
double_output=True,
label_names=["labels", "labels_2"],
use_ipex=True,
bf16=mix_bf16,
use_cpu=True,
output_dir=tmp_dir,
)
outputs = trainer.predict(trainer.eval_dataset)
preds = outputs.predictions
labels = outputs.label_ids
x = trainer.eval_dataset.x
self.assertEqual(len(preds), 2)
self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))
self.assertTrue(np.array_equal(labels[0], trainer.eval_dataset.ys[0]))
self.assertTrue(np.array_equal(labels[1], trainer.eval_dataset.ys[1]))
def test_dynamic_shapes(self):
eval_dataset = DynamicShapesDataset(batch_size=self.batch_size)
model = RegressionModel(a=2, b=1)