riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-08-02 19:21:31 +06:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Purge unused ModelTester code (#37085)

Browse Source 
* Purge correctly this time

* Remove more methods from recent PRs

* make fixup
This commit is contained in:
Matt 2025-04-03 17:48:35 +01:00 committed by GitHub
parent 1b29409d89
commit 2d46a08b63
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
72 changed files with 3 additions and 4286 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
tests/models/albert/test_modeling_albert.py
View File

@ -198,16 +198,6 @@ class AlbertModelTester:
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
def create_and_check_for_token_classification(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = AlbertForTokenClassification(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_for_multiple_choice(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):

13
tests/models/aria/test_modeling_aria.py
View File

@ -168,19 +168,6 @@ class AriaVisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_aria_model_fp16_forward(self, config, input_ids, pixel_values, attention_mask):
model = AriaForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class AriaForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

27
tests/models/aya_vision/test_modeling_aya_vision.py
View File

@ -154,33 +154,6 @@ class AyaVisionVisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_model_fp16_forward(self, config, input_ids, pixel_values, attention_mask):
model = AyaVisionForConditionalGeneration(config=config)
model.to(torch_device)
model.half()
model.eval()
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values,
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
def create_and_check_model_fp16_autocast_forward(self, config, input_ids, pixel_values, attention_mask):
config.torch_dtype = torch.float16
model = AyaVisionForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values,
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class AyaVisionModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):

18
tests/models/big_bird/test_modeling_big_bird.py
View File

@ -235,24 +235,6 @@ class BigBirdModelTester:
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = BigBirdForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_for_masked_lm(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):

18
tests/models/biogpt/test_modeling_biogpt.py
View File

@ -136,24 +136,6 @@ class BioGptModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = BioGptForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_biogpt_model_attention_mask_past(
self, config, input_ids, input_mask, head_mask, token_type_ids, *args
):

8
tests/models/bloom/test_modeling_bloom.py
View File

@ -275,14 +275,6 @@ class BloomModelTester:
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_question_answering_model(self, config, input_ids, input_mask, *args):
model = BloomForQuestionAnswering(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_forward_and_backwards(
self, config, input_ids, input_mask, *args, gradient_checkpointing=False
):

77
tests/models/chameleon/test_modeling_chameleon.py
View File

@ -180,83 +180,6 @@ class ChameleonModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = ChameleonForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
model = ChameleonForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/cohere/test_modeling_cohere.py
View File

@ -146,116 +146,6 @@ class CohereModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = self.model_class(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = self.for_causal_lm_class(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = self.for_causal_lm_class(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

29
tests/models/convnextv2/test_modeling_convnextv2.py
View File

@ -119,35 +119,6 @@ class ConvNextV2ModelTester:
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
def create_and_check_backbone(self, config, pixel_values, labels):
model = ConvNextV2Backbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify hidden states
self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[1], 4, 4])
# verify channels
self.parent.assertEqual(len(model.channels), len(config.out_features))
self.parent.assertListEqual(model.channels, config.hidden_sizes[1:])
# verify backbone works with out_features=None
config.out_features = None
model = ConvNextV2Backbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify feature maps
self.parent.assertEqual(len(result.feature_maps), 1)
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[-1], 1, 1])
# verify channels
self.parent.assertEqual(len(model.channels), 1)
self.parent.assertListEqual(model.channels, [config.hidden_sizes[-1]])
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, labels = config_and_inputs

9
tests/models/ctrl/test_modeling_ctrl.py
View File

@ -180,15 +180,6 @@ class CTRLModelTester:
return config, inputs_dict
def create_and_check_ctrl_for_sequence_classification(self, config, input_ids, head_mask, token_type_ids, *args):
config.num_labels = self.num_labels
model = CTRLForSequenceClassification(config)
model.to(torch_device)
model.eval()
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
result = model(input_ids, token_type_ids=token_type_ids, labels=sequence_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
@require_torch
class CTRLModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):

26
tests/models/data2vec/test_modeling_data2vec_audio.py
View File

@ -184,32 +184,6 @@ class Data2VecAudioModelTester:
(self.batch_size, self.adapter_output_seq_length, config.output_hidden_size),
)
def create_and_check_batch_inference(self, config, input_values, *args):
# test does not pass for models making use of `group_norm`
# check: https://github.com/pytorch/fairseq/issues/3227
model = Data2VecAudioModel(config=config)
model.to(torch_device)
model.eval()
input_values = input_values[:3]
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
# pad input
for i in range(len(input_lengths)):
input_values[i, input_lengths[i] :] = 0.0
attention_mask[i, input_lengths[i] :] = 0.0
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
for i in range(input_values.shape[0]):
input_slice = input_values[i : i + 1, : input_lengths[i]]
output = model(input_slice).last_hidden_state
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
def check_ctc_loss(self, config, input_values, *args):
model = Data2VecAudioForCTC(config=config)
model.to(torch_device)

112
tests/models/dbrx/test_modeling_dbrx.py
View File

@ -191,118 +191,6 @@ class DbrxModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Dbrx
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = DbrxModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Dbrx
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = DbrxForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = DbrxForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common with Llama->Dbrx
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

110
tests/models/deepseek_v3/test_modeling_deepseek_v3.py
View File

@ -193,116 +193,6 @@ class DeepseekV3ModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = DeepseekV3Model(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = DeepseekV3ForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = DeepseekV3ForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/diffllama/test_modeling_diffllama.py
View File

@ -159,116 +159,6 @@ class DiffLlamaModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = DiffLlamaModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = DiffLlamaForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = DiffLlamaForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/falcon/test_modeling_falcon.py
View File

@ -152,116 +152,6 @@ class FalconModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = FalconModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = FalconForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = FalconForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

39
tests/models/fnet/test_modeling_fnet.py
View File

@ -22,7 +22,7 @@ from transformers.models.auto import get_values
from transformers.testing_utils import require_tokenizers, require_torch, slow, torch_device
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
from ...test_modeling_common import ModelTesterMixin, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
@ -41,10 +41,6 @@ if is_torch_available():
FNetModel,
FNetTokenizerFast,
)
from transformers.models.fnet.modeling_fnet import (
FNetBasicFourierTransform,
is_scipy_available,
)
# Override ConfigTester
@ -133,26 +129,6 @@ class FNetModelTester:
tpu_short_seq_length=self.seq_length,
)
@require_torch
def create_and_check_fourier_transform(self, config):
hidden_states = floats_tensor([self.batch_size, self.seq_length, config.hidden_size])
transform = FNetBasicFourierTransform(config)
fftn_output = transform(hidden_states)
config.use_tpu_fourier_optimizations = True
if is_scipy_available():
transform = FNetBasicFourierTransform(config)
dft_output = transform(hidden_states)
config.max_position_embeddings = 4097
transform = FNetBasicFourierTransform(config)
fft_output = transform(hidden_states)
if is_scipy_available():
self.parent.assertTrue(torch.allclose(fftn_output[0][0], dft_output[0][0], atol=1e-4))
self.parent.assertTrue(torch.allclose(fft_output[0][0], dft_output[0][0], atol=1e-4))
self.parent.assertTrue(torch.allclose(fftn_output[0][0], fft_output[0][0], atol=1e-4))
def create_and_check_model(self, config, input_ids, token_type_ids, sequence_labels, token_labels, choice_labels):
model = FNetModel(config=config)
model.to(torch_device)
@ -185,19 +161,6 @@ class FNetModelTester:
result = model(input_ids, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_for_next_sentence_prediction(
self, config, input_ids, token_type_ids, sequence_labels, token_labels, choice_labels
):
model = FNetForNextSentencePrediction(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
token_type_ids=token_type_ids,
next_sentence_label=sequence_labels,
)
self.parent.assertEqual(result.logits.shape, (self.batch_size, 2))
def create_and_check_for_question_answering(
self, config, input_ids, token_type_ids, sequence_labels, token_labels, choice_labels
):

123
tests/models/fuyu/test_modeling_fuyu.py
View File

@ -22,7 +22,7 @@ import requests
from parameterized import parameterized
from transformers import FuyuConfig, is_torch_available, is_vision_available
from transformers.testing_utils import require_torch, require_torch_accelerator, slow, torch_device
from transformers.testing_utils import require_torch, require_torch_accelerator, slow
from transformers.utils import cached_property
from ...generation.test_utils import GenerationTesterMixin
@ -39,8 +39,6 @@ if is_torch_available() and is_vision_available():
if is_torch_available():
import torch
from transformers import FuyuForCausalLM
@ -133,125 +131,6 @@ class FuyuModelTester:
pad_token_id=self.pad_token_id,
)
def create_and_check_model(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
):
model = FuyuForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = FuyuForCausalLM(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = FuyuForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = FuyuForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/gemma/test_modeling_gemma.py
View File

@ -168,116 +168,6 @@ class GemmaModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = self.model_class(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = self.for_causal_lm_class(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = self.for_causal_lm_class(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common with Llama->Gemma
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

107
tests/models/glm/test_modeling_glm.py
View File

@ -156,113 +156,6 @@ class GlmModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = self.model_class(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = self.for_causal_lm_class(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = self.for_causal_lm_class(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common with Llama->Glm
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

27
tests/models/got_ocr2/test_modeling_got_ocr2.py
View File

@ -138,33 +138,6 @@ class GotOcr2VisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_model_fp16_forward(self, config, input_ids, pixel_values, attention_mask):
model = GotOcr2ForConditionalGeneration(config=config)
model.to(torch_device)
model.half()
model.eval()
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
def create_and_check_model_fp16_autocast_forward(self, config, input_ids, pixel_values, attention_mask):
config.torch_dtype = torch.float16
model = GotOcr2ForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class GotOcr2ModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):

110
tests/models/granite/test_modeling_granite.py
View File

@ -146,116 +146,6 @@ class GraniteModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = GraniteModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = GraniteForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = GraniteForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/granitemoe/test_modeling_granitemoe.py
View File

@ -145,116 +145,6 @@ class GraniteMoeModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = GraniteMoeModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = GraniteMoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = GraniteMoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/granitemoeshared/test_modeling_granitemoeshared.py
View File

@ -148,116 +148,6 @@ class GraniteMoeSharedModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = GraniteMoeSharedModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = GraniteMoeSharedForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = GraniteMoeSharedForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/jetmoe/test_modeling_jetmoe.py
View File

@ -150,116 +150,6 @@ class JetMoeModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = JetMoeModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = JetMoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = JetMoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/llama/test_modeling_llama.py
View File

@ -151,116 +151,6 @@ class LlamaModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = LlamaModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = LlamaForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = LlamaForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

13
tests/models/llava/test_modeling_llava.py
View File

@ -160,19 +160,6 @@ class LlavaVisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_llava_model_fp16_forward(self, config, input_ids, pixel_values, attention_mask):
model = LlavaForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class LlavaForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

33
tests/models/llava_next/test_modeling_llava_next.py
View File

@ -174,39 +174,6 @@ class LlavaNextVisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_llava_next_model_fp16_forward(
self, config, input_ids, pixel_values, attention_mask, image_sizes
):
model = LlavaNextForConditionalGeneration(config=config)
model.to(torch_device)
model.half()
model.eval()
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_sizes=image_sizes,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
def create_and_check_llava_next_model_fp16_autocast_forward(
self, config, input_ids, pixel_values, attention_mask, image_sizes
):
config.torch_dtype = torch.float16
model = LlavaNextForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_sizes=image_sizes,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class LlavaNextForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

35
tests/models/llava_next_video/test_modeling_llava_next_video.py
View File

@ -188,41 +188,6 @@ class LlavaNextVideoVisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_llava_next_video_model_fp16_forward(
self, config, input_ids, pixel_values, pixel_values_videos, attention_mask, image_sizes
):
model = LlavaNextVideoForConditionalGeneration(config=config)
model.to(torch_device)
model.half()
model.eval()
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_sizes=image_sizes,
pixel_values=pixel_values.to(torch.bfloat16),
pixel_values_videos=pixel_values_videos.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
def create_and_check_llava_next_video_model_fp16_autocast_forward(
self, config, input_ids, pixel_values, pixel_values_videos, attention_mask, image_sizes
):
config.torch_dtype = torch.float16
model = LlavaNextVideoForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_sizes=image_sizes,
pixel_values=pixel_values.to(torch.bfloat16),
pixel_values_videos=pixel_values_videos.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class LlavaNextVideoForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

33
tests/models/llava_onevision/test_modeling_llava_onevision.py
View File

@ -174,39 +174,6 @@ class LlavaOnevisionVisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_llava_onevision_model_fp16_forward(
self, config, input_ids, pixel_values, attention_mask, image_sizes
):
model = LlavaOnevisionForConditionalGeneration(config=config)
model.to(torch_device)
model.half()
model.eval()
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_sizes=image_sizes,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
def create_and_check_llava_onevision_model_fp16_autocast_forward(
self, config, input_ids, pixel_values, attention_mask, image_sizes
):
config.torch_dtype = torch.float16
model = LlavaOnevisionForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_sizes=image_sizes,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class LlavaOnevisionForConditionalGenerationModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

9
tests/models/megatron_bert/test_modeling_megatron_bert.py
View File

@ -157,15 +157,6 @@ class MegatronBertModelTester:
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_for_causal_lm(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = MegatronBertForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_megatron_bert_for_next_sequence_prediction(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):

113
tests/models/mistral/test_modeling_mistral.py
View File

@ -159,119 +159,6 @@ class MistralModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Mistral
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = MistralModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Mistral
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = MistralForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Mistral
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = MistralForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

27
tests/models/mistral3/test_modeling_mistral3.py
View File

@ -160,33 +160,6 @@ class Mistral3VisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_model_fp16_forward(self, config, input_ids, pixel_values, attention_mask):
model = Mistral3ForConditionalGeneration(config=config)
model.to(torch_device)
model.half()
model.eval()
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
def create_and_check_model_fp16_autocast_forward(self, config, input_ids, pixel_values, attention_mask):
config.torch_dtype = torch.float16
model = Mistral3ForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class Mistral3ModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):

113
tests/models/mixtral/test_modeling_mixtral.py
View File

@ -157,119 +157,6 @@ class MixtralModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Mixtral
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = MixtralModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Mixtral
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = MixtralForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Mixtral
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = MixtralForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common with Llama->Mixtral
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

35
tests/models/moonshine/test_modeling_moonshine.py
View File

@ -109,41 +109,6 @@ class MoonshineModelTester:
eos_token_id=self.eos_token_id,
)
def create_and_check_model(self, config, input_values, attention_mask):
model = MoonshineModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_values, attention_mask=attention_mask)
self.parent.assertEqual(
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
)
def create_and_check_batch_inference(self, config, input_values, *args):
# test does not pass for models making use of `group_norm`
# check: https://github.com/pytorch/fairseq/issues/3227
model = MoonshineModel(config=config)
model.to(torch_device)
model.eval()
input_values = input_values[:3]
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
# pad input
for i in range(len(input_lengths)):
input_values[i, input_lengths[i] :] = 0.0
attention_mask[i, input_lengths[i] :] = 0.0
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
for i in range(input_values.shape[0]):
input_slice = input_values[i : i + 1, : input_lengths[i]]
output = model(input_slice).last_hidden_state
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
def check_output_attentions(self, config, input_values, attention_mask):
model = MoonshineModel(config=config)
model.config.layerdrop = 1.0

8
tests/models/mpt/test_modeling_mpt.py
View File

@ -287,14 +287,6 @@ class MptModelTester:
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_question_answering_model(self, config, input_ids, input_mask, *args):
model = MptForQuestionAnswering(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
def create_and_check_forward_and_backwards(
self, config, input_ids, input_mask, *args, gradient_checkpointing=False
):

32
tests/models/mra/test_modeling_mra.py
View File

@ -169,38 +169,6 @@ class MraModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = MraModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_masked_lm(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):

110
tests/models/olmo/test_modeling_olmo.py
View File

@ -146,116 +146,6 @@ class OlmoModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = OlmoModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = OlmoForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = OlmoForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/olmo2/test_modeling_olmo2.py
View File

@ -145,116 +145,6 @@ class Olmo2ModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = Olmo2Model(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = Olmo2ForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = Olmo2ForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/olmoe/test_modeling_olmoe.py
View File

@ -159,116 +159,6 @@ class OlmoeModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = OlmoeModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = OlmoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = OlmoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/persimmon/test_modeling_persimmon.py
View File

@ -151,116 +151,6 @@ class PersimmonModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = PersimmonModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = PersimmonForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = PersimmonForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

110
tests/models/phi/test_modeling_phi.py
View File

@ -146,116 +146,6 @@ class PhiModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = PhiModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = PhiForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = PhiForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

113
tests/models/phi3/test_modeling_phi3.py
View File

@ -198,119 +198,6 @@ class Phi3ModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Phi3
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = Phi3Model(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Phi3
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = Phi3ForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Phi3
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = Phi3ForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

13
tests/models/phi4_multimodal/test_modeling_phi4_multimodal.py
View File

@ -190,19 +190,6 @@ class Phi4MultimodalModelTester:
}
return config, inputs_dict
def create_and_check_model(self, config, input_ids, attention_mask):
model = Phi4MultimodalForCausalLM(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
return_dict=True,
)["logits"]
self.parent.assertEqual(logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class Phi4MultimodalModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

113
tests/models/phimoe/test_modeling_phimoe.py
View File

@ -205,119 +205,6 @@ class PhimoeModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Phimoe
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = PhimoeModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Phimoe
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = PhimoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Phimoe
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = PhimoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

26
tests/models/pixtral/test_modeling_pixtral.py
View File

@ -96,32 +96,6 @@ class PixtralVisionModelTester:
initializer_range=self.initializer_range,
)
def create_and_check_model(self, config, pixel_values):
model = PixtralVisionModel(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = (self.image_size, self.image_size)
patch_size = (self.patch_size, self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
def create_and_check_model_with_projection(self, config, pixel_values):
model = PixtralVisionModel(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = (self.image_size, self.image_size)
patch_size = (self.patch_size, self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
self.parent.assertEqual(result.image_embeds.shape, (self.batch_size, self.projection_dim))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, image_sizes = config_and_inputs

18
tests/models/pop2piano/test_modeling_pop2piano.py
View File

@ -358,24 +358,6 @@ class Pop2PianoModelTester:
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def create_and_check_generate_with_past_key_values(
self,
config,
input_ids,
decoder_input_ids,
attention_mask,
decoder_attention_mask,
lm_labels,
):
model = Pop2PianoForConditionalGeneration(config=config).to(torch_device).eval()
torch.manual_seed(0)
output_without_past_cache = model.generate(
input_ids[:1], num_beams=2, max_length=5, do_sample=True, use_cache=False
)
torch.manual_seed(0)
output_with_past_cache = model.generate(input_ids[:1], num_beams=2, max_length=5, do_sample=True)
self.parent.assertTrue(torch.all(output_with_past_cache == output_without_past_cache))
def create_and_check_model_fp16_forward(
self,
config,

18
tests/models/pvt/test_modeling_pvt.py
View File

@ -121,24 +121,6 @@ class PvtModelTester:
result = model(pixel_values)
self.parent.assertIsNotNone(result.last_hidden_state)
def create_and_check_for_image_classification(self, config, pixel_values, labels):
config.num_labels = self.type_sequence_label_size
model = PvtForImageClassification(config)
model.to(torch_device)
model.eval()
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
# test greyscale images
config.num_channels = 1
model = PvtForImageClassification(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, labels = config_and_inputs

47
tests/models/pvt_v2/test_modeling_pvt_v2.py
View File

@ -128,53 +128,6 @@ class PvtV2ModelTester(ModelTesterMixin):
result = model(pixel_values)
self.parent.assertIsNotNone(result.last_hidden_state)
def create_and_check_backbone(self, config, pixel_values, labels):
model = PvtV2Backbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify feature maps
self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[1], 4, 4])
# verify channels
self.parent.assertEqual(len(model.channels), len(config.out_features))
self.parent.assertListEqual(model.channels, config.hidden_sizes[1:])
# verify backbone works with out_features=None
config.out_features = None
model = PvtV2Backbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify feature maps
self.parent.assertEqual(len(result.feature_maps), 1)
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[-1], 1, 1])
# verify channels
self.parent.assertEqual(len(model.channels), 1)
self.parent.assertListEqual(model.channels, [config.hidden_sizes[-1]])
def create_and_check_for_image_classification(self, config, pixel_values, labels):
config.num_labels = self.num_labels
model = PvtV2ForImageClassification(config)
model.to(torch_device)
model.eval()
result = model(pixel_values, labels=labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
# test greyscale images
config.num_channels = 1
model = PvtV2ForImageClassification(config)
model.to(torch_device)
model.eval()
pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
result = model(pixel_values)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, labels = config_and_inputs

113
tests/models/qwen2/test_modeling_qwen2.py
View File

@ -169,119 +169,6 @@ class Qwen2ModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Qwen2
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = Qwen2Model(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Qwen2
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = Qwen2ForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Qwen2
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = Qwen2ForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

33
tests/models/qwen2_5_vl/test_modeling_qwen2_5_vl.py
View File

@ -189,39 +189,6 @@ class Qwen2_5_VLVisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_qwen2_5_vl_model_fp16_forward(
self, config, input_ids, pixel_values, attention_mask, image_grid_thw
):
model = Qwen2_5_VLForConditionalGeneration(config=config)
model.to(torch_device)
model.half()
model.eval()
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_grid_thw=image_grid_thw,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
def create_and_check_qwen2_5_vl_model_fp16_autocast_forward(
self, config, input_ids, pixel_values, attention_mask, image_grid_thw
):
config.torch_dtype = torch.float16
model = Qwen2_5_VLForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_grid_thw=image_grid_thw,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class Qwen2_5_VLModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

13
tests/models/qwen2_audio/test_modeling_qwen2_audio.py
View File

@ -131,19 +131,6 @@ class Qwen2AudioModelTester:
}
return config, inputs_dict
def create_and_check_qwen2audio_model_fp16_forward(self, config, input_ids, pixel_values, attention_mask):
model = Qwen2AudioForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class Qwen2AudioForConditionalGenerationModelTest(ModelTesterMixin, unittest.TestCase):

113
tests/models/qwen2_moe/test_modeling_qwen2_moe.py
View File

@ -197,119 +197,6 @@ class Qwen2MoeModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Qwen2Moe
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = Qwen2MoeModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Qwen2Moe
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = Qwen2MoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Qwen2Moe
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = Qwen2MoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

33
tests/models/qwen2_vl/test_modeling_qwen2_vl.py
View File

@ -184,39 +184,6 @@ class Qwen2VLVisionText2TextModelTester:
}
return config, inputs_dict
def create_and_check_qwen2_vl_model_fp16_forward(
self, config, input_ids, pixel_values, attention_mask, image_grid_thw
):
model = Qwen2VLForConditionalGeneration(config=config)
model.to(torch_device)
model.half()
model.eval()
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_grid_thw=image_grid_thw,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
def create_and_check_qwen2_vl_model_fp16_autocast_forward(
self, config, input_ids, pixel_values, attention_mask, image_grid_thw
):
config.torch_dtype = torch.float16
model = Qwen2VLForConditionalGeneration(config=config)
model.to(torch_device)
model.eval()
with torch.autocast(device_type="cuda", dtype=torch.float16):
logits = model(
input_ids=input_ids,
attention_mask=attention_mask,
image_grid_thw=image_grid_thw,
pixel_values=pixel_values.to(torch.bfloat16),
return_dict=True,
)["logits"]
self.parent.assertFalse(torch.isnan(logits).any().item())
@require_torch
class Qwen2VLModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

113
tests/models/qwen3/test_modeling_qwen3.py
View File

@ -172,119 +172,6 @@ class Qwen3ModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Qwen3
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = Qwen3Model(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Qwen3
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = Qwen3ForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Qwen3
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = Qwen3ForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

113
tests/models/qwen3_moe/test_modeling_qwen3_moe.py
View File

@ -191,119 +191,6 @@ class Qwen3MoeModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Qwen3Moe
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = Qwen3MoeModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Qwen3Moe
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = Qwen3MoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Qwen3Moe
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = Qwen3MoeForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

113
tests/models/recurrent_gemma/test_modeling_recurrent_gemma.py
View File

@ -152,119 +152,6 @@ class RecurrentGemmaModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->RecurrentGemma
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = RecurrentGemmaModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->RecurrentGemma
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = RecurrentGemmaForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->RecurrentGemma
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = RecurrentGemmaForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common with Llama->RecurrentGemma
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

18
tests/models/rembert/test_modeling_rembert.py
View File

@ -200,24 +200,6 @@ class RemBertModelTester:
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = RemBertForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_for_masked_lm(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):

27
tests/models/roc_bert/test_modeling_roc_bert.py
View File

@ -259,33 +259,6 @@ class RoCBertModelTester:
)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
input_shape_ids,
input_pronunciation_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = RoCBertForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
input_shape_ids=input_shape_ids,
input_pronunciation_ids=input_pronunciation_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
labels=token_labels,
)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_for_masked_lm(
self,
config,

18
tests/models/roformer/test_modeling_roformer.py
View File

@ -200,24 +200,6 @@ class RoFormerModelTester:
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = RoFormerForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_for_generate_causal_lm(
self,
config,

31
tests/models/rt_detr/test_modeling_rt_detr_resnet.py
View File

@ -16,7 +16,7 @@
import unittest
from transformers import RTDetrResNetConfig
from transformers.testing_utils import require_torch, torch_device
from transformers.testing_utils import require_torch
from transformers.utils.import_utils import is_torch_available
from ...test_backbone_common import BackboneTesterMixin
@ -84,35 +84,6 @@ class RTDetrResNetModelTester:
out_indices=self.out_indices,
)
def create_and_check_backbone(self, config, pixel_values, labels):
model = RTDetrResNetBackbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify feature maps
self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[1], 4, 4])
# verify channels
self.parent.assertEqual(len(model.channels), len(config.out_features))
self.parent.assertListEqual(model.channels, config.hidden_sizes[1:])
# verify backbone works with out_features=None
config.out_features = None
model = RTDetrResNetBackbone(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# verify feature maps
self.parent.assertEqual(len(result.feature_maps), 1)
self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[-1], 1, 1])
# verify channels
self.parent.assertEqual(len(model.channels), 1)
self.parent.assertListEqual(model.channels, [config.hidden_sizes[-1]])
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values, labels = config_and_inputs

13
tests/models/rwkv/test_modeling_rwkv.py
View File

@ -195,19 +195,6 @@ class RwkvModelTester:
self.parent.assertTrue(torch.allclose(torch.cat([output_one, output_two], dim=1), output_whole, atol=1e-5))
def create_and_check_forward_and_backwards(
self, config, input_ids, input_mask, head_mask, token_type_ids, *args, gradient_checkpointing=False
):
model = RwkvForCausalLM(config)
model.to(torch_device)
if gradient_checkpointing:
model.gradient_checkpointing_enable()
result = model(input_ids, labels=input_ids)
self.parent.assertEqual(result.loss.shape, ())
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
result.loss.backward()
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

26
tests/models/sew/test_modeling_sew.py
View File

@ -147,32 +147,6 @@ class SEWModelTester:
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
)
def create_and_check_batch_inference(self, config, input_values, *args):
# test does not pass for models making use of `group_norm`
# check: https://github.com/pytorch/fairseq/issues/3227
model = SEWModel(config=config)
model.to(torch_device)
model.eval()
input_values = input_values[:3]
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
# pad input
for i in range(len(input_lengths)):
input_values[i, input_lengths[i] :] = 0.0
attention_mask[i, input_lengths[i] :] = 0.0
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
for i in range(input_values.shape[0]):
input_slice = input_values[i : i + 1, : input_lengths[i]]
output = model(input_slice).last_hidden_state
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
def check_ctc_loss(self, config, input_values, *args):
model = SEWForCTC(config=config)
model.to(torch_device)

26
tests/models/sew_d/test_modeling_sew_d.py
View File

@ -168,32 +168,6 @@ class SEWDModelTester:
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
)
def create_and_check_batch_inference(self, config, input_values, *args):
# test does not pass for models making use of `group_norm`
# check: https://github.com/pytorch/fairseq/issues/3227
model = SEWDModel(config=config)
model.to(torch_device)
model.eval()
input_values = input_values[:3]
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
# pad input
for i in range(len(input_lengths)):
input_values[i, input_lengths[i] :] = 0.0
attention_mask[i, input_lengths[i] :] = 0.0
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
for i in range(input_values.shape[0]):
input_slice = input_values[i : i + 1, : input_lengths[i]]
output = model(input_slice).last_hidden_state
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
def check_ctc_loss(self, config, input_values, *args):
model = SEWDForCTC(config=config)
model.to(torch_device)

110
tests/models/stablelm/test_modeling_stablelm.py
View File

@ -153,116 +153,6 @@ class StableLmModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = StableLmModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = StableLmForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = StableLmForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(

113
tests/models/starcoder2/test_modeling_starcoder2.py
View File

@ -157,119 +157,6 @@ class Starcoder2ModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Starcoder2
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = Starcoder2Model(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Starcoder2
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = Starcoder2ForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Starcoder2
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = Starcoder2ForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()

13
tests/models/timm_backbone/test_modeling_timm_backbone.py
View File

@ -27,8 +27,6 @@ from ...test_modeling_common import ModelTesterMixin, floats_tensor
if is_torch_available():
import torch
from transformers import TimmBackbone, TimmBackboneConfig
from ...test_pipeline_mixin import PipelineTesterMixin
@ -76,17 +74,6 @@ class TimmBackboneModelTester:
backbone=self.backbone,
)
def create_and_check_model(self, config, pixel_values):
model = TimmBackbone(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values)
self.parent.assertEqual(
result.feature_map[-1].shape,
(self.batch_size, model.channels[-1], 14, 14),
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values = config_and_inputs

11
tests/models/timm_wrapper/test_modeling_timm_wrapper.py
View File

@ -75,17 +75,6 @@ class TimmWrapperModelTester:
def get_config(self):
return TimmWrapperConfig.from_pretrained(self.model_name)
def create_and_check_model(self, config, pixel_values):
model = TimmWrapperModel(config=config)
model.to(torch_device)
model.eval()
with torch.no_grad():
result = model(pixel_values)
self.parent.assertEqual(
result.feature_map[-1].shape,
(self.batch_size, model.channels[-1], 14, 14),
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values = config_and_inputs

37
tests/models/umt5/test_modeling_umt5.py
View File

@ -229,43 +229,6 @@ class UMT5ModelTester:
# There should be a self attn key, a self attn value, a cross attn key and a cross attn value stored in each decoder_past tuple
self.parent.assertEqual(len(decoder_past[0]), 4)
def create_and_check_decoder_model_past(
self,
config,
input_ids,
decoder_input_ids,
attention_mask,
decoder_attention_mask,
lm_labels,
):
model = UMT5Model(config=config).get_decoder().to(torch_device).eval()
# first forward pass
outputs = model(input_ids, use_cache=True)
outputs_use_cache_conf = model(input_ids)
outputs_no_past = model(input_ids, use_cache=False)
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
output, past_key_values = outputs.to_tuple()
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
output_from_no_past = model(next_input_ids)["last_hidden_state"]
output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def create_and_check_model_fp16_forward(
self,
config,

26
tests/models/wav2vec2_bert/test_modeling_wav2vec2_bert.py
View File

@ -246,32 +246,6 @@ class Wav2Vec2BertModelTester:
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
)
def create_and_check_batch_inference(self, config, input_features, *args):
# test does not pass for models making use of `group_norm`
# check: https://github.com/pytorch/fairseq/issues/3227
model = Wav2Vec2BertModel(config=config)
model.to(torch_device)
model.eval()
input_features = input_features[:3]
attention_mask = torch.ones(input_features.shape, device=torch_device, dtype=torch.bool)
input_lengths = [input_features.shape[-1] // i for i in [4, 2, 1]]
# pad input
for i in range(len(input_lengths)):
input_features[i, input_lengths[i] :] = 0.0
attention_mask[i, input_lengths[i] :] = 0.0
batch_outputs = model(input_features, attention_mask=attention_mask).last_hidden_state
for i in range(input_features.shape[0]):
input_slice = input_features[i : i + 1, : input_lengths[i]]
output = model(input_slice).last_hidden_state
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
def check_ctc_loss(self, config, input_features, *args):
model = Wav2Vec2BertForCTC(config=config)
model.to(torch_device)

26
tests/models/wav2vec2_conformer/test_modeling_wav2vec2_conformer.py
View File

@ -240,32 +240,6 @@ class Wav2Vec2ConformerModelTester:
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
)
def create_and_check_batch_inference(self, config, input_values, *args):
# test does not pass for models making use of `group_norm`
# check: https://github.com/pytorch/fairseq/issues/3227
model = Wav2Vec2ConformerModel(config=config)
model.to(torch_device)
model.eval()
input_values = input_values[:3]
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
# pad input
for i in range(len(input_lengths)):
input_values[i, input_lengths[i] :] = 0.0
attention_mask[i, input_lengths[i] :] = 0.0
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
for i in range(input_values.shape[0]):
input_slice = input_values[i : i + 1, : input_lengths[i]]
output = model(input_slice).last_hidden_state
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
def check_ctc_loss(self, config, input_values, *args):
model = Wav2Vec2ConformerForCTC(config=config)
model.to(torch_device)

26
tests/models/wavlm/test_modeling_wavlm.py
View File

@ -157,32 +157,6 @@ class WavLMModelTester:
result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
)
def create_and_check_batch_inference(self, config, input_values, *args):
# test does not pass for models making use of `group_norm`
# check: https://github.com/pytorch/fairseq/issues/3227
model = WavLMModel(config=config)
model.to(torch_device)
model.eval()
input_values = input_values[:3]
attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.bool)
input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]]
# pad input
for i in range(len(input_lengths)):
input_values[i, input_lengths[i] :] = 0.0
attention_mask[i, input_lengths[i] :] = 0.0
batch_outputs = model(input_values, attention_mask=attention_mask).last_hidden_state
for i in range(input_values.shape[0]):
input_slice = input_values[i : i + 1, : input_lengths[i]]
output = model(input_slice).last_hidden_state
batch_output = batch_outputs[i : i + 1, : output.shape[1]]
self.parent.assertTrue(torch.allclose(output, batch_output, atol=1e-3))
def check_ctc_loss(self, config, input_values, *args):
model = WavLMForCTC(config=config)
model.to(torch_device)

32
tests/models/yoso/test_modeling_yoso.py
View File

@ -168,38 +168,6 @@ class YosoModelTester:
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = YosoModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_masked_lm(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):