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

enable misc cases on XPU & use device agnostic APIs for cases in tests (#38192)

Browse Source 
* use device agnostic APIs in tests

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

* more

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

* fix style

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

* add reset_peak_memory_stats API

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

* update

---------

Signed-off-by: Matrix Yao <matrix.yao@intel.com>
Signed-off-by: YAO Matrix <matrix.yao@intel.com>
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
This commit is contained in:
Yao Matrix
2025-05-20 16:09:01 +08:00
committed by GitHub
parent dbc4b91db4
commit 3bd1c20149
13 changed files with 52 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
tests/models/aria/test_modeling_aria.py
View File

@@ -30,6 +30,7 @@ from transformers import (
)
from transformers.models.idefics3 import Idefics3VisionConfig
from transformers.testing_utils import (
backend_empty_cache,
require_bitsandbytes,
require_torch,
require_torch_large_accelerator,
@@ -302,7 +303,7 @@ class AriaForConditionalGenerationIntegrationTest(unittest.TestCase):
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
@slow
@require_torch_large_accelerator

8
tests/models/cohere/test_tokenization_cohere.py
View File

@@ -17,7 +17,11 @@ import unittest
from functools import lru_cache
from transformers import CohereTokenizerFast
from transformers.testing_utils import require_jinja, require_tokenizers, require_torch_multi_gpu
from transformers.testing_utils import (
require_jinja,
require_tokenizers,
require_torch_multi_accelerator,
)
from ...test_tokenization_common import TokenizerTesterMixin, use_cache_if_possible
@@ -55,7 +59,7 @@ class CohereTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
return CohereTokenizerFast.from_pretrained(pretrained_name, **kwargs)
# This gives CPU OOM on a single-gpu runner (~60G RAM). On multi-gpu runner, it has ~180G RAM which is enough.
@require_torch_multi_gpu
@require_torch_multi_accelerator
def test_torch_encode_plus_sent_to_model(self):
super().test_torch_encode_plus_sent_to_model()

3
tests/models/colpali/test_modeling_colpali.py
View File

@@ -29,6 +29,7 @@ from transformers.models.colpali.configuration_colpali import ColPaliConfig
from transformers.models.colpali.modeling_colpali import ColPaliForRetrieval, ColPaliForRetrievalOutput
from transformers.models.colpali.processing_colpali import ColPaliProcessor
from transformers.testing_utils import (
backend_empty_cache,
require_torch,
require_vision,
slow,
@@ -303,7 +304,7 @@ class ColPaliModelIntegrationTest(unittest.TestCase):
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
@slow
def test_model_integration_test(self):

4
tests/models/idefics2/test_modeling_idefics2.py
View File

@@ -35,7 +35,7 @@ from transformers.testing_utils import (
require_flash_attn,
require_torch,
require_torch_gpu,
require_torch_multi_gpu,
require_torch_multi_accelerator,
require_torch_sdpa,
slow,
torch_device,
@@ -583,7 +583,7 @@ class Idefics2ForConditionalGenerationIntegrationTest(unittest.TestCase):
cleanup(torch_device, gc_collect=True)
@slow
@require_torch_multi_gpu
@require_torch_multi_accelerator
def test_integration_test(self):
model = Idefics2ForConditionalGeneration.from_pretrained(
"HuggingFaceM4/idefics2-8b-base",

4
tests/models/phi4_multimodal/test_modeling_phi4_multimodal.py
View File

@@ -31,7 +31,7 @@ from transformers import (
is_torch_available,
is_vision_available,
)
from transformers.testing_utils import require_soundfile, require_torch, slow, torch_device
from transformers.testing_utils import backend_empty_cache, require_soundfile, require_torch, slow, torch_device
from transformers.utils import is_soundfile_available
from ...generation.test_utils import GenerationTesterMixin
@@ -296,7 +296,7 @@ class Phi4MultimodalIntegrationTest(unittest.TestCase):
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
def test_text_only_generation(self):
model = AutoModelForCausalLM.from_pretrained(

3
tests/models/qwen2_5_vl/test_modeling_qwen2_5_vl.py
View File

@@ -29,6 +29,7 @@ from transformers import (
is_vision_available,
)
from transformers.testing_utils import (
backend_empty_cache,
is_flaky,
require_cv2,
require_flash_attn,
@@ -421,7 +422,7 @@ class Qwen2_5_VLIntegrationTest(unittest.TestCase):
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
@slow
def test_small_model_integration_test(self):

3
tests/models/qwen2_vl/test_modeling_qwen2_vl.py
View File

@@ -28,6 +28,7 @@ from transformers import (
is_vision_available,
)
from transformers.testing_utils import (
backend_empty_cache,
require_flash_attn,
require_torch,
require_torch_gpu,
@@ -367,7 +368,7 @@ class Qwen2VLIntegrationTest(unittest.TestCase):
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
@slow
def test_small_model_integration_test(self):

8
tests/models/whisper/test_modeling_whisper.py
View File

@@ -32,7 +32,6 @@ from transformers import WhisperConfig
from transformers.testing_utils import (
is_flaky,
require_flash_attn,
require_non_xpu,
require_torch,
require_torch_accelerator,
require_torch_fp16,
@@ -42,7 +41,7 @@ from transformers.testing_utils import (
slow,
torch_device,
)
from transformers.utils import cached_property, is_torch_available, is_torchaudio_available
from transformers.utils import cached_property, is_torch_available, is_torch_xpu_available, is_torchaudio_available
from transformers.utils.import_utils import is_datasets_available
from ...generation.test_utils import GenerationTesterMixin
@@ -2431,11 +2430,10 @@ class WhisperModelIntegrationTests(unittest.TestCase):
" How many different species are there in the chilli? How many different species are there in the chilli?",
)
@require_non_xpu
@slow
@require_torch_gpu
@require_torch_accelerator
def test_speculative_decoding_distil(self):
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
torch_dtype = torch.float16 if (torch.cuda.is_available() or is_torch_xpu_available()) else torch.float32
model_id = "openai/whisper-large-v2"
model = WhisperForConditionalGeneration.from_pretrained(
model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True

4
tests/tensor_parallel/test_tensor_parallel.py
View File

@@ -21,9 +21,11 @@ from transformers import is_torch_available
from transformers.integrations.tensor_parallel import get_packed_weights, repack_weights
from transformers.testing_utils import (
TestCasePlus,
backend_device_count,
get_torch_dist_unique_port,
require_huggingface_hub_greater_or_equal,
require_torch_multi_gpu,
torch_device,
)
@@ -168,4 +170,4 @@ class TestTensorParallel(TestCasePlus):
@require_torch_multi_gpu
class TestTensorParallelCuda(TestTensorParallel):
nproc_per_node = torch.cuda.device_count()
nproc_per_node = backend_device_count(torch_device)

7
tests/test_modeling_common.py
View File

@@ -73,6 +73,7 @@ from transformers.models.auto.modeling_auto import (
)
from transformers.testing_utils import (
CaptureLogger,
backend_empty_cache,
get_device_properties,
hub_retry,
is_flaky,
@@ -2652,7 +2653,7 @@ class ModelTesterMixin:
config = self.model_tester.get_large_model_config()
for model_class in self.all_parallelizable_model_classes:
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
# 1. single gpu memory load + unload + memory measurements
# Retrieve initial memory usage (can easily be ~0.6-1.5GB if cuda-kernels have been preloaded by previous tests)
@@ -2668,7 +2669,7 @@ class ModelTesterMixin:
del model
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
# 2. MP test
# it's essential to re-calibrate the usage before the next stage
@@ -2692,7 +2693,7 @@ class ModelTesterMixin:
del model
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
@require_torch_gpu
@require_torch_multi_gpu

19
tests/trainer/test_trainer.py
View File

@@ -66,6 +66,7 @@ from transformers.testing_utils import (
backend_max_memory_allocated,
backend_memory_allocated,
backend_reset_max_memory_allocated,
backend_reset_peak_memory_stats,
evaluate_side_effect_factory,
execute_subprocess_async,
get_gpu_count,
@@ -1654,7 +1655,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
self.assertFalse(is_any_loss_nan_or_inf(log_history_filter))
def test_train_and_eval_dataloaders(self):
if torch_device == "cuda":
if torch_device in ["cuda", "xpu"]:
n_gpu = max(1, backend_device_count(torch_device))
else:
n_gpu = 1
@@ -4106,7 +4107,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
mod = MyModule()
# 1. without TorchDynamo (eager baseline)
a = torch.ones(1024, 1024, device="cuda", requires_grad=True)
a = torch.ones(1024, 1024, device=torch_device, requires_grad=True)
a.grad = None
trainer = CustomTrainer(model=mod)
# warmup
@@ -4115,17 +4116,17 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
# resets
gc.collect()
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
backend_empty_cache(torch_device)
backend_reset_peak_memory_stats(torch_device)
orig_loss = trainer.training_step(mod, {"x": a})
orig_peak_mem = torch.cuda.max_memory_allocated()
orig_peak_mem = backend_max_memory_allocated(torch_device)
torchdynamo.reset()
del trainer
# 2. TorchDynamo nvfuser
with tempfile.TemporaryDirectory() as tmp_dir:
a = torch.ones(1024, 1024, device="cuda", requires_grad=True)
a = torch.ones(1024, 1024, device=torch_device, requires_grad=True)
a.grad = None
args = TrainingArguments(output_dir=tmp_dir, torch_compile_backend="nvfuser")
trainer = CustomTrainer(model=mod, args=args)
@@ -4135,11 +4136,11 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
# resets
gc.collect()
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
backend_empty_cache(torch_device)
backend_reset_peak_memory_stats(torch_device)
loss = trainer.training_step(mod, {"x": a})
peak_mem = torch.cuda.max_memory_allocated()
peak_mem = backend_max_memory_allocated(torch_device)
torchdynamo.reset()
del trainer

5
tests/utils/test_cache_utils.py
View File

@@ -21,6 +21,7 @@ from transformers import set_seed
from transformers.generation.configuration_utils import ALL_CACHE_IMPLEMENTATIONS
from transformers.testing_utils import (
CaptureStderr,
backend_device_count,
cleanup,
get_gpu_count,
is_torch_available,
@@ -210,8 +211,8 @@ class CacheIntegrationTest(unittest.TestCase):
if not has_accelerator:
self.skipTest("Offloaded caches require an accelerator")
if cache_implementation in ["offloaded_static", "offloaded_hybrid_chunked"]:
if torch.cuda.device_count() != 1:
self.skipTest("Offloaded static caches require exactly 1 GPU")
if backend_device_count(torch_device) != 1:
self.skipTest("Offloaded static caches require exactly 1 accelerator")
@parameterized.expand(TEST_CACHE_IMPLEMENTATIONS)
def test_cache_batched(self, cache_implementation):