File size: 17,942 Bytes
dfefe0b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 | diff --git a/docker/transformers-all-latest-gpu/Dockerfile b/docker/transformers-all-latest-gpu/Dockerfile
index f9e9aa17897f..8b3024fe92cc 100644
--- a/docker/transformers-all-latest-gpu/Dockerfile
+++ b/docker/transformers-all-latest-gpu/Dockerfile
@@ -10,8 +10,6 @@ SHELL ["sh", "-lc"]
# to be used as arguments for docker build (so far).
ARG PYTORCH='2.6.0'
-# (not always a valid torch version)
-ARG INTEL_TORCH_EXT='2.3.0'
# Example: `cu102`, `cu113`, etc.
ARG CUDA='cu121'
# Disable kernel mapping for now until all tests pass
@@ -32,8 +30,6 @@ RUN python3 -m pip install --no-cache-dir -U tensorflow==2.13 protobuf==3.20.3 "
RUN python3 -m pip uninstall -y flax jax
-RUN python3 -m pip install --no-cache-dir intel_extension_for_pytorch==$INTEL_TORCH_EXT -f https://developer.intel.com/ipex-whl-stable-cpu
-
RUN python3 -m pip install --no-cache-dir git+https://github.com/facebookresearch/detectron2.git pytesseract
RUN python3 -m pip install -U "itsdangerous<2.1.0"
diff --git a/docs/source/en/perf_infer_cpu.md b/docs/source/en/perf_infer_cpu.md
index 7522a013d0d9..af93d3cabd69 100644
--- a/docs/source/en/perf_infer_cpu.md
+++ b/docs/source/en/perf_infer_cpu.md
@@ -78,26 +78,3 @@ python examples/pytorch/question-answering/run_qa.py \
--no_cuda \
--jit_mode_eval
```
-
-## IPEX
-
-[Intel Extension for PyTorch](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/getting_started.html) (IPEX) offers additional optimizations for PyTorch on Intel CPUs. IPEX further optimizes TorchScript with [graph optimization](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/features/graph_optimization.html) which fuses operations like Multi-head attention, Concat Linear, Linear + Add, Linear + Gelu, Add + LayerNorm, and more, into single kernels for faster execution.
-
-Make sure IPEX is installed, and set the `--use_opex` and `--jit_mode_eval` flags in [`Trainer`] to enable IPEX graph optimization and TorchScript.
-
-```bash
-!pip install intel_extension_for_pytorch
-```
-
-```bash
-python examples/pytorch/question-answering/run_qa.py \
---model_name_or_path csarron/bert-base-uncased-squad-v1 \
---dataset_name squad \
---do_eval \
---max_seq_length 384 \
---doc_stride 128 \
---output_dir /tmp/ \
---no_cuda \
---use_ipex \
---jit_mode_eval
-```
diff --git a/docs/source/en/perf_train_cpu.md b/docs/source/en/perf_train_cpu.md
index 1eab6afbde23..29641a99843f 100644
--- a/docs/source/en/perf_train_cpu.md
+++ b/docs/source/en/perf_train_cpu.md
@@ -17,30 +17,9 @@ rendered properly in your Markdown viewer.
A modern CPU is capable of efficiently training large models by leveraging the underlying optimizations built into the hardware and training on fp16 or bf16 data types.
-This guide focuses on how to train large models on an Intel CPU using mixed precision and the [Intel Extension for PyTorch (IPEX)](https://intel.github.io/intel-extension-for-pytorch/index.html) library.
+This guide focuses on how to train large models on an Intel CPU using mixed precision. AMP is enabled for CPU backends training with PyTorch.
-You can Find your PyTorch version by running the command below.
-
-```bash
-pip list | grep torch
-```
-
-Install IPEX with the PyTorch version from above.
-
-```bash
-pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
-```
-
-> [!TIP]
-> Refer to the IPEX [installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation) guide for more details.
-
-IPEX provides additional performance optimizations for Intel CPUs. These include additional CPU instruction level architecture (ISA) support such as [Intel AVX512-VNNI](https://en.wikichip.org/wiki/x86/avx512_vnni) and [Intel AMX](https://www.intel.com/content/www/us/en/products/docs/accelerator-engines/what-is-intel-amx.html). Both of these features are designed to accelerate matrix multiplication. Older AMD and Intel CPUs with only Intel AVX2, however, aren't guaranteed better performance with IPEX.
-
-IPEX also supports [Auto Mixed Precision (AMP)](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/features/amp.html) training with the fp16 and bf16 data types. Reducing precision speeds up training and reduces memory usage because it requires less computation. The loss in accuracy from using full-precision is minimal. 3rd, 4th, and 5th generation Intel Xeon Scalable processors natively support bf16, and the 6th generation processor also natively supports fp16 in addition to bf16.
-
-AMP is enabled for CPU backends training with PyTorch.
-
-[`Trainer`] supports AMP training with a CPU by adding the `--use_cpu`, `--use_ipex`, and `--bf16` parameters. The example below demonstrates the [run_qa.py](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering) script.
+[`Trainer`] supports AMP training with CPU by adding the `--use_cpu`, and `--bf16` parameters. The example below demonstrates the [run_qa.py](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering) script.
```bash
python run_qa.py \
@@ -54,7 +33,6 @@ python run_qa.py \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir /tmp/debug_squad/ \
- --use_ipex \
--bf16 \
--use_cpu
```
@@ -65,7 +43,6 @@ These parameters can also be added to [`TrainingArguments`] as shown below.
training_args = TrainingArguments(
output_dir="./outputs",
bf16=True,
- use_ipex=True,
use_cpu=True,
)
```
diff --git a/docs/source/en/perf_train_cpu_many.md b/docs/source/en/perf_train_cpu_many.md
index bd332d15e1ba..0de9f895a78a 100644
--- a/docs/source/en/perf_train_cpu_many.md
+++ b/docs/source/en/perf_train_cpu_many.md
@@ -75,8 +75,7 @@ python3 run_qa.py \
--doc_stride 128 \
--output_dir /tmp/debug_squad/ \
--no_cuda \
- --ddp_backend ccl \
- --use_ipex
+ --ddp_backend ccl
```
</hfoption>
@@ -115,7 +114,6 @@ python3 run_qa.py \
--output_dir /tmp/debug_squad/ \
--no_cuda \
--ddp_backend ccl \
- --use_ipex \
--bf16
```
@@ -201,8 +199,7 @@ spec:
--output_dir /tmp/pvc-mount/output_$(date +%Y%m%d_%H%M%S) \
--no_cuda \
--ddp_backend ccl \
- --bf16 \
- --use_ipex;
+ --bf16;
env:
- name: LD_PRELOAD
value: "/usr/lib/x86_64-linux-gnu/libtcmalloc.so.4.5.9:/usr/local/lib/libiomp5.so"
diff --git a/src/transformers/trainer.py b/src/transformers/trainer.py
index ca03db994715..e0d6555706bb 100755
--- a/src/transformers/trainer.py
+++ b/src/transformers/trainer.py
@@ -157,7 +157,6 @@
is_galore_torch_available,
is_grokadamw_available,
is_in_notebook,
- is_ipex_available,
is_liger_kernel_available,
is_lomo_available,
is_peft_available,
@@ -1915,29 +1914,6 @@ def torch_jit_model_eval(self, model, dataloader, training=False):
return model
- def ipex_optimize_model(self, model, training=False, dtype=torch.float32):
- if not is_ipex_available():
- raise ImportError(
- "Using IPEX but IPEX is not installed or IPEX's version does not match current PyTorch, please refer"
- " to https://github.com/intel/intel-extension-for-pytorch."
- )
-
- import intel_extension_for_pytorch as ipex
-
- if not training:
- model.eval()
- dtype = torch.bfloat16 if not self.is_in_train and self.args.bf16_full_eval else dtype
- # conv_bn_folding is disabled as it fails in symbolic tracing, resulting in ipex warnings
- model = ipex.optimize(model, dtype=dtype, level="O1", conv_bn_folding=False, inplace=not self.is_in_train)
- else:
- if not model.training:
- model.train()
- model, self.optimizer = ipex.optimize(
- model, dtype=dtype, optimizer=self.optimizer, inplace=True, level="O1"
- )
-
- return model
-
def compare_trainer_and_checkpoint_args(self, training_args, trainer_state):
attributes_map = {
"logging_steps": "logging_steps",
@@ -1967,10 +1943,6 @@ def compare_trainer_and_checkpoint_args(self, training_args, trainer_state):
logger.warning_once(warning_str)
def _wrap_model(self, model, training=True, dataloader=None):
- if self.args.use_ipex:
- dtype = torch.bfloat16 if self.use_cpu_amp else torch.float32
- model = self.ipex_optimize_model(model, training, dtype=dtype)
-
if is_sagemaker_mp_enabled():
# Wrapping the base model twice in a DistributedModel will raise an error.
if isinstance(self.model_wrapped, smp.model.DistributedModel):
diff --git a/src/transformers/training_args.py b/src/transformers/training_args.py
index 2d58e26af319..bbb1f22350ea 100644
--- a/src/transformers/training_args.py
+++ b/src/transformers/training_args.py
@@ -1581,6 +1581,12 @@ def __post_init__(self):
FutureWarning,
)
self.use_cpu = self.no_cuda
+ if self.use_ipex:
+ warnings.warn(
+ "using `use_ipex` is deprecated and will be removed in version 4.54 of 🤗 Transformers. "
+ "You only need PyTorch for the needed optimizations on Intel CPU and XPU.",
+ FutureWarning,
+ )
self.eval_strategy = IntervalStrategy(self.eval_strategy)
self.logging_strategy = IntervalStrategy(self.logging_strategy)
diff --git a/tests/trainer/test_trainer.py b/tests/trainer/test_trainer.py
index 41bc61e8a7f3..65a16066c979 100644
--- a/tests/trainer/test_trainer.py
+++ b/tests/trainer/test_trainer.py
@@ -79,7 +79,6 @@
require_deepspeed,
require_galore_torch,
require_grokadamw,
- require_intel_extension_for_pytorch,
require_liger_kernel,
require_lomo,
require_non_hpu,
@@ -1325,37 +1324,6 @@ def test_number_of_steps_in_training(self):
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 @@ def test_evaluate_with_jit(self):
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 @@ def test_predict_with_jit(self):
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)
|