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

Rename _supports_flash_attn_2 in examples and tests (#39471)

Browse Source 
* delete `_supports_flash_attn_2` from examples and tests

* simplify docs
This commit is contained in:
Raushan Turganbay
2025-07-21 14:02:57 +02:00
committed by GitHub
parent 3a152e3a5c
commit 8c102e2eb1
22 changed files with 39 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/causal_lm_tester.py
View File

@@ -422,7 +422,7 @@ class CausalLMModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterM
@slow
def test_flash_attn_2_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(reason="Model does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

4
tests/generation/test_utils.py
View File

@@ -2297,8 +2297,8 @@ class GenerationTesterMixin:
max_new_tokens = 3
support_flag = {
"sdpa": "_supports_sdpa",
"flash_attention_2": "_supports_flash_attn_2",
"flash_attention_3": "_supports_flash_attn_3",
"flash_attention_2": "_supports_flash_attn",
"flash_attention_3": "_supports_flash_attn",
}
set_model_tester_for_less_flaky_test(self)

4
tests/models/aimv2/test_modeling_aimv2.py
View File

@@ -478,7 +478,7 @@ class Aimv2ModelTest(Aimv2ModelTesterMixin, PipelineTesterMixin, unittest.TestCa
@slow
def test_flash_attn_2_inference_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -516,7 +516,7 @@ class Aimv2ModelTest(Aimv2ModelTesterMixin, PipelineTesterMixin, unittest.TestCa
@mark.flash_attn_test
def test_flash_attn_2_inference_equivalence_right_padding(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

2
tests/models/bamba/test_modeling_bamba.py
View File

@@ -524,7 +524,7 @@ class BambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
max_new_tokens = 30
for model_class in self.all_generative_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

4
tests/models/bark/test_modeling_bark.py
View File

@@ -934,7 +934,7 @@ class BarkFineModelTest(ModelTesterMixin, unittest.TestCase):
@slow
def test_flash_attn_2_inference_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(reason="Model does not support flash_attention_2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -991,7 +991,7 @@ class BarkFineModelTest(ModelTesterMixin, unittest.TestCase):
@slow
def test_flash_attn_2_inference_equivalence_right_padding(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(reason="Model does not support flash_attention_2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

4
tests/models/clip/test_modeling_clip.py
View File

@@ -715,7 +715,7 @@ class CLIPModelTest(CLIPModelTesterMixin, PipelineTesterMixin, unittest.TestCase
@slow
def test_flash_attn_2_inference_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -753,7 +753,7 @@ class CLIPModelTest(CLIPModelTesterMixin, PipelineTesterMixin, unittest.TestCase
@mark.flash_attn_test
def test_flash_attn_2_inference_equivalence_right_padding(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

2
tests/models/esm/test_modeling_esm.py
View File

@@ -318,7 +318,7 @@ class EsmModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
@slow
def test_flash_attn_2_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(reason="Model does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

2
tests/models/kyutai_speech_to_text/test_modeling_kyutai_speech_to_text.py
View File

@@ -548,7 +548,7 @@ class KyutaiSpeechToTextModelTest(ModelTesterMixin, GenerationTesterMixin, Pipel
max_new_tokens = 30
support_flag = {
"sdpa": "_supports_sdpa",
"flash_attention_2": "_supports_flash_attn_2",
"flash_attention_2": "_supports_flash_attn",
}
for model_class in self.all_generative_model_classes:

8
tests/models/musicgen/test_modeling_musicgen.py
View File

@@ -292,7 +292,7 @@ class MusicgenDecoderTest(ModelTesterMixin, GenerationTesterMixin, PipelineTeste
# Copied from tests.test_modeling_common.ModelTesterMixin.test_flash_attn_2_inference_equivalence
def test_flash_attn_2_inference_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -372,7 +372,7 @@ class MusicgenDecoderTest(ModelTesterMixin, GenerationTesterMixin, PipelineTeste
# Copied from tests.test_modeling_common.ModelTesterMixin.test_flash_attn_2_inference_equivalence_right_padding
def test_flash_attn_2_inference_equivalence_right_padding(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -948,7 +948,7 @@ class MusicgenTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
# Adapted from tests.test_modeling_common.ModelTesterMixin.test_flash_attn_2_inference_equivalence
def test_flash_attn_2_inference_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -1096,7 +1096,7 @@ class MusicgenTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
# Adapted from tests.test_modeling_common.ModelTesterMixin.test_flash_attn_2_inference_equivalence_right_padding
def test_flash_attn_2_inference_equivalence_right_padding(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

8
tests/models/musicgen_melody/test_modeling_musicgen_melody.py
View File

@@ -300,7 +300,7 @@ class MusicgenMelodyDecoderTest(ModelTesterMixin, GenerationTesterMixin, unittes
# Copied from tests.models.musicgen.test_modeling_musicgen.MusicgenDecoderTest.test_flash_attn_2_inference_equivalence
def test_flash_attn_2_inference_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -382,7 +382,7 @@ class MusicgenMelodyDecoderTest(ModelTesterMixin, GenerationTesterMixin, unittes
# Copied from tests.models.musicgen.test_modeling_musicgen.MusicgenDecoderTest.test_flash_attn_2_inference_equivalence_right_padding
def test_flash_attn_2_inference_equivalence_right_padding(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -948,7 +948,7 @@ class MusicgenMelodyTest(ModelTesterMixin, GenerationTesterMixin, PipelineTester
# Adapted from tests.test_modeling_common.ModelTesterMixin.test_flash_attn_2_inference_equivalence
def test_flash_attn_2_inference_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -1096,7 +1096,7 @@ class MusicgenMelodyTest(ModelTesterMixin, GenerationTesterMixin, PipelineTester
# Adapted from tests.test_modeling_common.ModelTesterMixin.test_flash_attn_2_inference_equivalence_right_padding
def test_flash_attn_2_inference_equivalence_right_padding(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

2
tests/models/siglip/test_modeling_siglip.py
View File

@@ -610,7 +610,7 @@ class SiglipModelTest(SiglipModelTesterMixin, PipelineTesterMixin, unittest.Test
@slow
def test_flash_attn_2_inference_equivalence(self):
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

2
tests/models/siglip2/test_modeling_siglip2.py
View File

@@ -99,7 +99,7 @@ class Siglip2ModelTesterMixin(ModelTesterMixin):
dtype = torch.float16
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
# Prepare inputs

2
tests/models/videomae/test_modeling_videomae.py
View File

@@ -359,7 +359,7 @@ class VideoMAEModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
self.skipTest(reason="Model architecture does not support attentions")
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

2
tests/models/vit_mae/test_modeling_vit_mae.py
View File

@@ -273,7 +273,7 @@ class ViTMAEModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
self.skipTest(reason="Model architecture does not support attentions")
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

4
tests/models/whisper/test_modeling_whisper.py
View File

@@ -867,7 +867,7 @@ class WhisperModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
import torch
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(reason="Model does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -913,7 +913,7 @@ class WhisperModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
import torch
for model_class in self.all_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(reason="Model does not support flash_attention_2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

29
tests/test_modeling_common.py
View File

@@ -3471,9 +3471,7 @@ class ModelTesterMixin:
self.skipTest(reason="Model architecture does not support attentions")
for model_class in self.all_model_classes:
if (attn_implementation == "flash_attention_2" and not model_class._supports_flash_attn_2) or (
attn_implementation == "flash_attention_3" and not model_class._supports_flash_attn_3
):
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support {attn_implementation}")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -3969,22 +3967,12 @@ class ModelTesterMixin:
model = model_class.from_pretrained(tmpdirname, torch_dtype=torch_dtype)
sub_models_supporting_fa = [
(
module._supports_flash_attn_3
if attn_implementation == "flash_attention_3"
else module._supports_flash_attn_2
)
module._supports_flash_attn
for name, module in model.named_modules()
if isinstance(module, PreTrainedModel) and name != ""
]
supports_fa_all_modules = (
all(sub_models_supporting_fa)
if len(sub_models_supporting_fa) > 0
else (
model._supports_flash_attn_3
if attn_implementation == "flash_attention_3"
else model._supports_flash_attn_2
)
all(sub_models_supporting_fa) if len(sub_models_supporting_fa) > 0 else model._supports_flash_attn
)
if not supports_fa_all_modules:
with self.assertRaises(ValueError):
@@ -4037,7 +4025,7 @@ class ModelTesterMixin:
self.skipTest(reason="Model architecture does not support attentions")
for model_class in self.all_generative_model_classes:
if not model_class._supports_flash_attn_2:
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support Flash Attention 2")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
model = model_class(config)
@@ -4104,9 +4092,8 @@ class ModelTesterMixin:
torch_dtype = torch.float16
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
config._attn_implementation = "flash_attention_2"
cls = self._torch_compile_train_cls
model = cls(config).to(device=torch_device, dtype=torch_dtype)
cls = self._torch_compile_train_cls # e.g. LlamaFroCausalLM
model = cls(config, attn_implementation="flash_attention_2").to(device=torch_device, dtype=torch_dtype)
inputs = {
"input_ids": torch.randint(low=1, high=model.config.vocab_size, size=(2, 10), device=torch_device),
@@ -4268,9 +4255,7 @@ class ModelTesterMixin:
self.skipTest(reason="Model architecture does not support attentions")
for model_class in self.all_generative_model_classes:
if (attn_implementation == "flash_attention_2" and not model_class._supports_flash_attn_2) or (
attn_implementation == "flash_attention_3" and not model_class._supports_flash_attn_3
):
if not model_class._supports_flash_attn:
self.skipTest(f"{model_class.__name__} does not support {attn_implementation}")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()