[tests] skip tests for xpu (#33553)

* enable

* fix

* add xpu skip

* add marker

* skip for xpu

* add more

* add one more

src/transformers/testing_utils.py

@@ -849,6 +849,13 @@ def require_torch_xpu(test_case):
     return unittest.skipUnless(is_torch_xpu_available(), "test requires XPU device")(test_case)
 
 
+def require_non_xpu(test_case):
+    """
+    Decorator marking a test that should be skipped for XPU.
+    """
+    return unittest.skipUnless(torch_device != "xpu", "test requires a non-XPU")(test_case)
+
+
 def require_torch_multi_xpu(test_case):
     """
     Decorator marking a test that requires a multi-XPU setup (in PyTorch). These tests are skipped on a machine without

tests/extended/test_trainer_ext.py

@@ -31,6 +31,7 @@ from transformers.testing_utils import (
     get_torch_dist_unique_port,
     require_apex,
     require_bitsandbytes,
+    require_non_xpu,
     require_torch,
     require_torch_gpu,
     require_torch_multi_accelerator,
@@ -106,6 +107,7 @@ class TestTrainerExt(TestCasePlus):
     def test_run_seq2seq_ddp(self):
         self.run_seq2seq_quick(distributed=True)
 
+    @require_non_xpu
     @require_apex
     @require_torch_gpu
     def test_run_seq2seq_apex(self):

tests/models/layoutlmv2/test_modeling_layoutlmv2.py

@@ -16,7 +16,14 @@
 
 import unittest
 
-from transformers.testing_utils import require_detectron2, require_torch, require_torch_multi_gpu, slow, torch_device
+from transformers.testing_utils import (
+    require_detectron2,
+    require_non_xpu,
+    require_torch,
+    require_torch_multi_gpu,
+    slow,
+    torch_device,
+)
 from transformers.utils import is_detectron2_available, is_torch_available
 
 from ...test_configuration_common import ConfigTester
@@ -251,6 +258,7 @@ class LayoutLMv2ModelTester:
         return config, inputs_dict
 
 
+@require_non_xpu
 @require_torch
 @require_detectron2
 class LayoutLMv2ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):

tests/test_modeling_common.py

@@ -76,6 +76,7 @@ from transformers.testing_utils import (
     require_accelerate,
     require_bitsandbytes,
     require_flash_attn,
+    require_non_xpu,
     require_read_token,
     require_safetensors,
     require_torch,
@@ -2884,6 +2885,7 @@ class ModelTesterMixin:
                 )
             self.assertTrue(torch.allclose(out_embeds, out_ids))
 
+    @require_non_xpu
     @require_torch_multi_gpu
     def test_multi_gpu_data_parallel_forward(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -4118,6 +4120,7 @@ class ModelTesterMixin:
                 with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
                     _ = model(**inputs_dict)
 
+    @require_non_xpu
     @require_torch_sdpa
     @require_torch_accelerator
     @slow

tests/trainer/test_trainer.py

@@ -66,6 +66,7 @@ from transformers.testing_utils import (
     require_intel_extension_for_pytorch,
     require_liger_kernel,
     require_lomo,
+    require_non_xpu,
     require_optuna,
     require_peft,
     require_ray,
@@ -884,6 +885,7 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
 
         # will add more specific tests once there are some bugs to fix
 
+    @require_non_xpu
     @require_torch_gpu
     @require_torch_tf32
     def test_tf32(self):
@@ -3196,6 +3198,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
         # perfect world: fp32_init/2 == fp16_eval
         self.assertAlmostEqual(fp16_eval, fp32_init / 2, delta=5_000)
 
+    @require_non_xpu
     @require_torch_non_multi_gpu
     @require_torchdynamo
     @require_torch_tensorrt_fx

tests/utils/test_cache_utils.py

@@ -22,6 +22,7 @@ from transformers import set_seed
 from transformers.testing_utils import (
     is_torch_available,
     require_auto_gptq,
+    require_non_xpu,
     require_read_token,
     require_torch,
     require_torch_gpu,
@@ -317,6 +318,7 @@ class CacheIntegrationTest(unittest.TestCase):
         ]
         self.assertListEqual(decoded, expected_text)
 
+    @require_non_xpu
     @require_auto_gptq
     def test_sink_cache_hard(self):
         tokenizer = AutoTokenizer.from_pretrained("TheBloke/LLaMa-7B-GPTQ")