mirror of
https://github.com/huggingface/transformers.git
synced 2025-07-31 02:02:21 +06:00
Tests: upgrade test_eager_matches_sdpa_generate (#34386)
This commit is contained in:
Joao Gante
GitHub
parent
8814043c8c
commit
186b8dc190
@ -15,6 +15,7 @@
|
||||
|
||||
|
||||
import copy
|
||||
import gc
|
||||
import inspect
|
||||
import tempfile
|
||||
import unittest
|
||||
@ -33,6 +34,7 @@ from transformers.testing_utils import (
|
||||
require_torch_gpu,
|
||||
require_torch_multi_accelerator,
|
||||
require_torch_multi_gpu,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -2046,6 +2048,86 @@ class GenerationTesterMixin:
|
||||
for model_class in self.all_generative_model_classes:
|
||||
self.assertTrue("GenerationMixin" in str(model_class.__bases__))
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
max_new_tokens = 30
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, original_inputs_dict = self.prepare_config_and_inputs_for_generate()
|
||||
inputs_dict = {}
|
||||
for input_name, input_data in original_inputs_dict.items():
|
||||
if isinstance(input_data, torch.Tensor) and input_data.dtype in [torch.float32, torch.bfloat16]:
|
||||
inputs_dict[input_name] = input_data.to(torch.float16)
|
||||
else:
|
||||
inputs_dict[input_name] = input_data
|
||||
main_input = inputs_dict[model_class.main_input_name]
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + main_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
del model
|
||||
gc.collect()
|
||||
|
||||
generate_kwargs = {
|
||||
"max_new_tokens": max_new_tokens,
|
||||
"do_sample": False,
|
||||
"return_dict_in_generate": True,
|
||||
"output_scores": True,
|
||||
}
|
||||
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
res_sdpa = model_sdpa.generate(**inputs_dict, **generate_kwargs)
|
||||
del model_sdpa
|
||||
gc.collect()
|
||||
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
res_eager = model_eager.generate(**inputs_dict, **generate_kwargs)
|
||||
del model_eager
|
||||
gc.collect()
|
||||
|
||||
# Eager and SDPA are very similar, but not exactly the same. Because we are using random models, this
|
||||
# test would be flaky if we only checked the sequences. Two situations in which this test passes:
|
||||
# 1. The sequences are the same
|
||||
# 2. The sequences are different, but the scores up until the first mismatch are nearly identical
|
||||
output_matches = res_eager.sequences == res_sdpa.sequences
|
||||
has_matching_outputs = output_matches.all()
|
||||
has_matching_scores = None
|
||||
if not has_matching_outputs:
|
||||
input_length = main_input.shape[1]
|
||||
for batch_idx in range(res_eager.sequences.shape[0]):
|
||||
batch_matches = output_matches[batch_idx]
|
||||
if batch_matches.all():
|
||||
continue
|
||||
first_mismatch_idx = batch_matches.int().argmin() # gets the index of the first False
|
||||
first_mismatch_idx -= input_length # scores doesn't include data regarding input tokens
|
||||
sdpa_first_mismatch_scores = res_sdpa.scores[first_mismatch_idx][batch_idx]
|
||||
eager_first_mismatch_scores = res_eager.scores[first_mismatch_idx][batch_idx]
|
||||
has_matching_scores = torch.allclose(
|
||||
sdpa_first_mismatch_scores, eager_first_mismatch_scores, rtol=1e-3, atol=1e-3
|
||||
)
|
||||
if not has_matching_scores:
|
||||
break
|
||||
|
||||
self.assertTrue(has_matching_outputs or has_matching_scores)
|
||||
|
||||
def _check_outputs(self, output, main_input, config, use_cache=False, num_return_sequences=1):
|
||||
# we can be sure what is batch size from main input but seq length depends on model type and whether input is text/audio/image
|
||||
# so we infer actual text seq length from model_tester, same was as it is done in `test_modeling_common.py` tests`
|
||||
|
||||
@ -22,7 +22,6 @@ from transformers.testing_utils import (
|
||||
CaptureLogger,
|
||||
require_torch,
|
||||
require_torch_accelerator,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -672,79 +671,6 @@ class BertModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin
|
||||
loaded = torch.jit.load(os.path.join(tmp, "bert.pt"), map_location=torch_device)
|
||||
loaded(inputs_dict["input_ids"].to(torch_device), inputs_dict["attention_mask"].to(torch_device))
|
||||
|
||||
# This test was copied from the common test_eager_matches_sdpa_generate(), but without low_cpu_mem_usage=True.
|
||||
# TODO: Remove this and use the parent method (in common tests) once BERT supports low_cpu_mem_usage=True.
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
max_new_tokens = 30
|
||||
|
||||
if len(self.all_generative_model_classes) == 0:
|
||||
self.skipTest(f"{self.__class__.__name__} tests a model that does support generate: skipping this test")
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
# low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
# low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
class_name = submodule.__class__.__name__
|
||||
if "SdpaAttention" in class_name or "SdpaSelfAttention" in class_name:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
class_name = submodule.__class__.__name__
|
||||
if "SdpaAttention" in class_name or "SdpaSelfAttention" in class_name:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
res_eager = model_eager.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
res_sdpa = model_sdpa.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
self.assertTrue(torch.allclose(res_eager, res_sdpa))
|
||||
|
||||
|
||||
@require_torch
|
||||
class BertModelIntegrationTest(unittest.TestCase):
|
||||
|
||||
@ -307,64 +307,6 @@ class CohereModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
|
||||
def test_torch_fx_output_loss(self):
|
||||
super().test_torch_fx_output_loss()
|
||||
|
||||
@require_bitsandbytes
|
||||
@require_torch_sdpa
|
||||
@require_torch_multi_gpu
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
"""
|
||||
Overwritting the common test as the test is flaky on tiny models
|
||||
"""
|
||||
max_new_tokens = 30
|
||||
|
||||
model_id = "CohereForAI/c4ai-command-r-v01-4bit"
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
|
||||
model_sdpa = CohereForCausalLM.from_pretrained(
|
||||
model_id, torch_dtype=torch.float16, low_cpu_mem_usage=True, device_map="auto"
|
||||
)
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = CohereForCausalLM.from_pretrained(
|
||||
model_id, torch_dtype=torch.float16, attn_implementation="eager", device_map="auto"
|
||||
)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
texts = [
|
||||
"hi here's a longer context, getting longer and",
|
||||
"Hello this is a very long sentence my friend, very long for real",
|
||||
"Today I am in Paris and",
|
||||
]
|
||||
|
||||
for padding_side in ["left", "right"]:
|
||||
tokenizer.padding_side = padding_side
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
|
||||
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
|
||||
with self.subTest(f"{padding_side}"):
|
||||
torch.testing.assert_close(
|
||||
res_eager,
|
||||
res_sdpa,
|
||||
msg=f"\n{tokenizer.batch_decode(res_eager)} \nvs\n{tokenizer.batch_decode(res_sdpa)}",
|
||||
)
|
||||
|
||||
|
||||
@require_torch
|
||||
@slow
|
||||
|
||||
@ -14,7 +14,6 @@
|
||||
# limitations under the License.
|
||||
"""Testing suite for the PyTorch Falcon model."""
|
||||
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
from parameterized import parameterized
|
||||
@ -27,7 +26,6 @@ from transformers import (
|
||||
set_seed,
|
||||
)
|
||||
from transformers.testing_utils import (
|
||||
is_flaky,
|
||||
require_bitsandbytes,
|
||||
require_torch,
|
||||
require_torch_sdpa,
|
||||
@ -520,78 +518,6 @@ class FalconModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
|
||||
torch.testing.assert_close(ntk_sin_long, original_sin_long)
|
||||
self.assertTrue((ntk_scaling_rope.inv_freq <= original_rope.inv_freq).all())
|
||||
|
||||
# TODO: @Fxmarty
|
||||
@is_flaky(max_attempts=3, description="flaky on some models.")
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
max_new_tokens = 30
|
||||
|
||||
if len(self.all_generative_model_classes) == 0:
|
||||
self.skipTest(f"{self.__class__.__name__} tests a model that does support generate: skipping this test")
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
# NOTE: This check is disabled for Falcon as the non-SDPA/SDPA implementation is in the same class (legacy reason).
|
||||
# for name, submodule in model_eager.named_modules():
|
||||
# if "SdpaAttention" in submodule.__class__.__name__:
|
||||
# raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
# has_sdpa = False
|
||||
# for name, submodule in model_sdpa.named_modules():
|
||||
# if "SdpaAttention" in submodule.__class__.__name__:
|
||||
# has_sdpa = True
|
||||
# break
|
||||
# if not has_sdpa:
|
||||
# raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
res_eager = model_eager.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
res_sdpa = model_sdpa.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
self.assertTrue(torch.allclose(res_eager, res_sdpa))
|
||||
|
||||
|
||||
@require_torch
|
||||
class FalconLanguageGenerationTest(unittest.TestCase):
|
||||
|
||||
@ -758,77 +758,6 @@ class GlmModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
|
||||
|
||||
self.assertTrue(len(fail_cases) == 0, "\n".join(fail_cases))
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
@is_flaky()
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
"""Overwrite to add flakyness: outputs sometimes start to diverge after some tokens"""
|
||||
|
||||
max_new_tokens = 30
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
class_name = submodule.__class__.__name__
|
||||
if "SdpaAttention" in class_name or "SdpaSelfAttention" in class_name:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
class_name = submodule.__class__.__name__
|
||||
if "SdpaAttention" in class_name or "SdpaSelfAttention" in class_name:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
res_eager = model_eager.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
res_sdpa = model_sdpa.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
self.assertTrue(torch.allclose(res_eager, res_sdpa))
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch_accelerator
|
||||
|
||||
@ -19,7 +19,7 @@ import unittest
|
||||
from parameterized import parameterized
|
||||
|
||||
from transformers import AutoTokenizer, GPTNeoXConfig, is_torch_available, set_seed
|
||||
from transformers.testing_utils import require_torch, require_torch_sdpa, slow, torch_device
|
||||
from transformers.testing_utils import require_torch, slow, torch_device
|
||||
|
||||
from ...generation.test_utils import GenerationTesterMixin
|
||||
from ...test_configuration_common import ConfigTester
|
||||
@ -434,68 +434,6 @@ class GPTNeoXModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
|
||||
torch.testing.assert_close(ntk_sin_long, original_sin_long)
|
||||
self.assertTrue((ntk_scaling_rope.inv_freq <= original_rope.inv_freq).all())
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
"""
|
||||
Based on tests.models.llama.test_modeling_llama.LlamaModelTest.test_eager_matches_sdpa_generate
|
||||
which also overwrites the common test as the test is flaky on tiny models.
|
||||
"""
|
||||
max_new_tokens = 30
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pythia-1b")
|
||||
|
||||
model_sdpa = GPTNeoXForCausalLM.from_pretrained(
|
||||
"EleutherAI/pythia-1b",
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = GPTNeoXForCausalLM.from_pretrained(
|
||||
"EleutherAI/pythia-1b",
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
texts = [
|
||||
"hi here's a longer context, getting longer and",
|
||||
"Hello this is a very long sentence my friend, very long for real",
|
||||
"Today I am in Paris and",
|
||||
]
|
||||
|
||||
for padding_side in ["left", "right"]:
|
||||
tokenizer.padding_side = padding_side
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
|
||||
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
|
||||
with self.subTest(f"{padding_side}"):
|
||||
torch.testing.assert_close(
|
||||
res_eager,
|
||||
res_sdpa,
|
||||
msg=f"\n{tokenizer.batch_decode(res_eager)} \nvs\n{tokenizer.batch_decode(res_sdpa)}",
|
||||
)
|
||||
|
||||
|
||||
@require_torch
|
||||
class GPTNeoXLanguageGenerationTest(unittest.TestCase):
|
||||
|
||||
@ -24,11 +24,9 @@ from parameterized import parameterized
|
||||
from transformers import AutoTokenizer, JetMoeConfig, is_torch_available
|
||||
from transformers.testing_utils import (
|
||||
backend_empty_cache,
|
||||
is_flaky,
|
||||
require_flash_attn,
|
||||
require_torch,
|
||||
require_torch_gpu,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -302,13 +300,6 @@ class JetMoeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
|
||||
test_disk_offload_bin = False
|
||||
test_disk_offload_safetensors = False
|
||||
|
||||
# TODO: @Fxmarty
|
||||
@is_flaky(max_attempts=3, description="flaky on some models.")
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
@parameterized.expand([(1, False), (1, True), (4, False)])
|
||||
def test_new_cache_format(self, num_beams, do_sample):
|
||||
pass
|
||||
|
||||
@ -32,7 +32,6 @@ from transformers.testing_utils import (
|
||||
require_torch,
|
||||
require_torch_accelerator,
|
||||
require_torch_gpu,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -651,67 +650,6 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
|
||||
if not has_flash:
|
||||
raise ValueError("The flash model should have flash attention layers")
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
"""
|
||||
Overwritting the common test as the test is flaky on tiny models
|
||||
"""
|
||||
max_new_tokens = 30
|
||||
|
||||
tokenizer = LlamaTokenizer.from_pretrained("saibo/llama-1B")
|
||||
|
||||
model_sdpa = LlamaForCausalLM.from_pretrained(
|
||||
"saibo/llama-1B",
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = LlamaForCausalLM.from_pretrained(
|
||||
"saibo/llama-1B",
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
texts = [
|
||||
"hi here's a longer context, getting longer and",
|
||||
"Hello this is a very long sentence my friend, very long for real",
|
||||
"Today I am in Paris and",
|
||||
]
|
||||
|
||||
for padding_side in ["left", "right"]:
|
||||
tokenizer.padding_side = padding_side
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
|
||||
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
|
||||
with self.subTest(f"{padding_side}"):
|
||||
torch.testing.assert_close(
|
||||
res_eager,
|
||||
res_sdpa,
|
||||
msg=f"\n{tokenizer.batch_decode(res_eager)} \nvs\n{tokenizer.batch_decode(res_sdpa)}",
|
||||
)
|
||||
|
||||
@unittest.skip("Broken by the loss update will fix soon @ArthurZucker")
|
||||
def test_torch_fx_output_loss(self, *args, **kwargs):
|
||||
pass
|
||||
|
||||
@ -24,7 +24,6 @@ from packaging import version
|
||||
from transformers import AutoTokenizer, MistralConfig, is_torch_available, set_seed
|
||||
from transformers.testing_utils import (
|
||||
backend_empty_cache,
|
||||
is_flaky,
|
||||
require_bitsandbytes,
|
||||
require_flash_attn,
|
||||
require_read_token,
|
||||
@ -332,13 +331,6 @@ class MistralModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
|
||||
):
|
||||
return True
|
||||
|
||||
# TODO: @Fxmarty
|
||||
@is_flaky(max_attempts=3, description="flaky on some models.")
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = MistralModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=MistralConfig, hidden_size=37)
|
||||
|
||||
@ -21,11 +21,9 @@ import pytest
|
||||
|
||||
from transformers import MixtralConfig, is_torch_available
|
||||
from transformers.testing_utils import (
|
||||
is_flaky,
|
||||
require_flash_attn,
|
||||
require_torch,
|
||||
require_torch_gpu,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -332,13 +330,6 @@ class MixtralModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMi
|
||||
):
|
||||
return True
|
||||
|
||||
# TODO: @Fxmarty
|
||||
@is_flaky(max_attempts=3, description="flaky on some models.")
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = MixtralModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=MixtralConfig, hidden_size=37)
|
||||
|
||||
@ -132,12 +132,6 @@ class MllamaForCausalLMModelTest(ModelTesterMixin, GenerationTesterMixin, unitte
|
||||
self.model_tester = MllamaText2TextModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=MllamaTextConfig, has_text_modality=True)
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
@is_flaky()
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
|
||||
class MllamaVisionText2TextModelTester:
|
||||
def __init__(
|
||||
@ -360,12 +354,6 @@ class MllamaForConditionalGenerationModelTest(ModelTesterMixin, GenerationTester
|
||||
|
||||
self.assertListEqual([layer_attention.shape for layer_attention in iter_attentions], expected_shapes)
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
@is_flaky()
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
@is_flaky()
|
||||
|
||||
@ -788,14 +788,10 @@ class MoshiTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
|
||||
@slow
|
||||
@is_flaky(max_attempts=5, description="flaky on some models.")
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
if not self.has_attentions:
|
||||
self.skipTest(reason="Model architecture does not support attentions")
|
||||
"""Overwritten -- mochi has custom inputs and custom output checks"""
|
||||
|
||||
max_new_tokens = 5
|
||||
|
||||
if len(self.all_generative_model_classes) == 0:
|
||||
self.skipTest(f"{self.__class__.__name__} tests a model that does support generate: skipping this test")
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
@ -819,74 +819,6 @@ class MusicgenDecoderTest(ModelTesterMixin, GenerationTesterMixin, PipelineTeste
|
||||
|
||||
self.assertTrue(len(fail_cases) == 0, "\n".join(fail_cases))
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
# Copied from tests.test_modeling_common.ModelTesterMixin.test_eager_matches_sdpa_generate
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
max_new_tokens = 30
|
||||
|
||||
# Ignore copy
|
||||
for model_class in self.greedy_sample_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
res_eager = model_eager.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
res_sdpa = model_sdpa.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
self.assertTrue(torch.allclose(res_eager, res_sdpa))
|
||||
|
||||
|
||||
def prepare_musicgen_inputs_dict(
|
||||
config,
|
||||
@ -2085,74 +2017,6 @@ class MusicgenTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
|
||||
|
||||
self.assertTrue(len(fail_cases) == 0, "\n".join(fail_cases))
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
# Copied from tests.test_modeling_common.ModelTesterMixin.test_eager_matches_sdpa_generate
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
max_new_tokens = 30
|
||||
|
||||
# Ignore copy
|
||||
for model_class in self.greedy_sample_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
res_eager = model_eager.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
res_sdpa = model_sdpa.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
self.assertTrue(torch.allclose(res_eager, res_sdpa))
|
||||
|
||||
def test_requires_grad_with_frozen_encoders(self):
|
||||
config = self.model_tester.get_config()
|
||||
for model_class in self.all_model_classes:
|
||||
|
||||
@ -1866,74 +1866,6 @@ class MusicgenMelodyTest(ModelTesterMixin, GenerationTesterMixin, PipelineTester
|
||||
|
||||
self.assertTrue(len(fail_cases) == 0, "\n".join(fail_cases))
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
# Copied from tests.test_modeling_common.ModelTesterMixin.test_eager_matches_sdpa_generate
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
max_new_tokens = 30
|
||||
|
||||
# Ignore copy
|
||||
for model_class in self.greedy_sample_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
res_eager = model_eager.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
res_sdpa = model_sdpa.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
self.assertTrue(torch.allclose(res_eager, res_sdpa))
|
||||
|
||||
def test_requires_grad_with_frozen_encoders(self):
|
||||
config = self.model_tester.get_config()
|
||||
for model_class in self.all_model_classes:
|
||||
|
||||
@ -24,10 +24,8 @@ from transformers.generation.configuration_utils import GenerationConfig
|
||||
from transformers.models.auto.tokenization_auto import AutoTokenizer
|
||||
from transformers.models.gpt_neox.tokenization_gpt_neox_fast import GPTNeoXTokenizerFast
|
||||
from transformers.testing_utils import (
|
||||
is_flaky,
|
||||
require_tokenizers,
|
||||
require_torch,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -317,13 +315,6 @@ class OlmoModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin
|
||||
def test_save_load_fast_init_from_base(self):
|
||||
pass
|
||||
|
||||
# TODO: @Fxmarty
|
||||
@is_flaky(max_attempts=3, description="flaky on some models.")
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
@parameterized.expand([("linear",), ("dynamic",)])
|
||||
def test_model_rope_scaling(self, scaling_type):
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
@ -22,10 +22,8 @@ from transformers import OlmoeConfig, is_torch_available, set_seed
|
||||
from transformers.models.auto.tokenization_auto import AutoTokenizer
|
||||
from transformers.models.gpt_neox.tokenization_gpt_neox_fast import GPTNeoXTokenizerFast
|
||||
from transformers.testing_utils import (
|
||||
is_flaky,
|
||||
require_tokenizers,
|
||||
require_torch,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -330,13 +328,6 @@ class OlmoeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
|
||||
def test_save_load_fast_init_from_base(self):
|
||||
pass
|
||||
|
||||
# TODO: @Fxmarty
|
||||
@is_flaky(max_attempts=3, description="flaky on some models.")
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
@parameterized.expand([("linear",), ("dynamic",)])
|
||||
def test_model_rope_scaling(self, scaling_type):
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
@ -25,7 +25,6 @@ from transformers.testing_utils import (
|
||||
require_torch,
|
||||
require_torch_accelerator,
|
||||
require_torch_fp16,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -339,68 +338,6 @@ class OPTModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
|
||||
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
||||
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
"""
|
||||
Overwritting the common test as the test is flaky on tiny models
|
||||
"""
|
||||
max_new_tokens = 30
|
||||
|
||||
tokenizer = GPT2Tokenizer.from_pretrained("facebook/opt-350M")
|
||||
|
||||
texts = [
|
||||
"hi here's a longer context, getting longer and",
|
||||
"Hello this is a very long sentence my friend, very long for real",
|
||||
"Today I am in Paris and",
|
||||
]
|
||||
|
||||
model_sdpa = OPTForCausalLM.from_pretrained(
|
||||
"facebook/opt-350M",
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="sdpa",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = OPTForCausalLM.from_pretrained(
|
||||
"facebook/opt-350M",
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for _, submodule in model_eager.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for _, submodule in model_sdpa.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
for padding_side in ["left", "right"]:
|
||||
tokenizer.padding_side = padding_side
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
|
||||
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
|
||||
with self.subTest(f"{padding_side}"):
|
||||
torch.testing.assert_close(
|
||||
res_eager,
|
||||
res_sdpa,
|
||||
msg=f"\n{tokenizer.batch_decode(res_eager)} \nvs\n{tokenizer.batch_decode(res_sdpa)}",
|
||||
)
|
||||
|
||||
@unittest.skip(reason="Does not work on the tiny model as we keep hitting edge cases.")
|
||||
def test_model_parallelism(self):
|
||||
super().test_model_parallelism()
|
||||
|
||||
@ -343,14 +343,6 @@ class Qwen2ModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
|
||||
):
|
||||
return True
|
||||
|
||||
# Ignore copy
|
||||
# TODO: @Fxmarty
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
@unittest.skip(reason="Currently failing.")
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = Qwen2ModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=Qwen2Config, hidden_size=37)
|
||||
|
||||
@ -368,12 +368,6 @@ class Qwen2MoeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterM
|
||||
):
|
||||
return True
|
||||
|
||||
# Ignore copy
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
super().test_eager_matches_sdpa_generate()
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = Qwen2MoeModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=Qwen2MoeConfig, hidden_size=37)
|
||||
|
||||
@ -21,11 +21,9 @@ from parameterized import parameterized
|
||||
|
||||
from transformers import StableLmConfig, is_torch_available, set_seed
|
||||
from transformers.testing_utils import (
|
||||
is_flaky,
|
||||
require_bitsandbytes,
|
||||
require_flash_attn,
|
||||
require_torch,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
@ -558,67 +556,3 @@ class StableLmModelIntegrationTest(unittest.TestCase):
|
||||
input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
|
||||
generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
|
||||
self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-3:].tolist())
|
||||
|
||||
# Copied from transformers.tests.models.llama.test_modeling_llama.LlamaModelTest.test_eager_matches_sdpa_generate with Llama->StableLm,saibo/llama-1B->stabilityai/stablelm-3b-4e1t
|
||||
# TODO: @Fxmarty
|
||||
@is_flaky(max_attempts=3, description="flaky on some models.")
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
"""
|
||||
Overwritting the common test as the test is flaky on tiny models
|
||||
"""
|
||||
max_new_tokens = 30
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablelm-3b-4e1t")
|
||||
|
||||
model_sdpa = StableLmForCausalLM.from_pretrained(
|
||||
"stabilityai/stablelm-3b-4e1t",
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
model_eager = StableLmForCausalLM.from_pretrained(
|
||||
"stabilityai/stablelm-3b-4e1t",
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
if "SdpaAttention" in submodule.__class__.__name__:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
texts = [
|
||||
"hi here's a longer context, getting longer and",
|
||||
"Hello this is a very long sentence my friend, very long for real",
|
||||
"Today I am in Paris and",
|
||||
]
|
||||
|
||||
for padding_side in ["left", "right"]:
|
||||
tokenizer.padding_side = padding_side
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
|
||||
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
res_eager = model_eager.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
res_sdpa = model_sdpa.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
|
||||
|
||||
with self.subTest(f"{padding_side}"):
|
||||
torch.testing.assert_close(
|
||||
res_eager,
|
||||
res_sdpa,
|
||||
msg=f"\n{tokenizer.batch_decode(res_eager)} \nvs\n{tokenizer.batch_decode(res_sdpa)}",
|
||||
)
|
||||
|
||||
@ -14,11 +14,10 @@
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
from transformers import XLMRobertaXLConfig, is_torch_available
|
||||
from transformers.testing_utils import require_torch, require_torch_sdpa, slow, torch_device
|
||||
from transformers.testing_utils import require_torch, slow, torch_device
|
||||
|
||||
from ...generation.test_utils import GenerationTesterMixin
|
||||
from ...test_configuration_common import ConfigTester
|
||||
@ -523,84 +522,6 @@ class XLMRobertaXLModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTes
|
||||
self.assertEqual(position_ids.shape, expected_positions.shape)
|
||||
self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
|
||||
|
||||
# TODO: Remove this and use the parent method (in common tests) once XLM RoBERTa XL supports low_cpu_mem_usage=True.
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
# Copied from tests.test_modeling_common.ModelTesterMixin.test_eager_matches_sdpa_generate
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
if not self.has_attentions:
|
||||
self.skipTest(reason="Model architecture does not support attentions")
|
||||
|
||||
max_new_tokens = 30
|
||||
|
||||
if len(self.all_generative_model_classes) == 0:
|
||||
self.skipTest(f"{self.__class__.__name__} tests a model that does support generate: skipping this test")
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
|
||||
# Ignore copy
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=False,
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_sdpa.config._attn_implementation == "sdpa")
|
||||
|
||||
# Ignore copy
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=False,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
self.assertTrue(model_eager.config._attn_implementation == "eager")
|
||||
|
||||
for name, submodule in model_eager.named_modules():
|
||||
class_name = submodule.__class__.__name__
|
||||
if "SdpaAttention" in class_name or "SdpaSelfAttention" in class_name:
|
||||
raise ValueError("The eager model should not have SDPA attention layers")
|
||||
|
||||
has_sdpa = False
|
||||
for name, submodule in model_sdpa.named_modules():
|
||||
class_name = submodule.__class__.__name__
|
||||
if "SdpaAttention" in class_name or "SdpaSelfAttention" in class_name:
|
||||
has_sdpa = True
|
||||
break
|
||||
if not has_sdpa:
|
||||
raise ValueError("The SDPA model should have SDPA attention layers")
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
res_eager = model_eager.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
res_sdpa = model_sdpa.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
self.assertTrue(torch.allclose(res_eager, res_sdpa))
|
||||
|
||||
|
||||
@require_torch
|
||||
class XLMRobertaModelXLIntegrationTest(unittest.TestCase):
|
||||
|
||||
@ -4469,62 +4469,6 @@ class ModelTesterMixin:
|
||||
with torch.no_grad():
|
||||
_ = model(**inputs_dict)
|
||||
|
||||
@require_torch_sdpa
|
||||
@slow
|
||||
def test_eager_matches_sdpa_generate(self):
|
||||
if not self.has_attentions:
|
||||
self.skipTest(reason="Model architecture does not support attentions")
|
||||
|
||||
max_new_tokens = 30
|
||||
|
||||
if len(self.all_generative_model_classes) == 0:
|
||||
self.skipTest(f"{self.__class__.__name__} tests a model that does support generate: skipping this test")
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
if not model_class._supports_sdpa:
|
||||
self.skipTest(f"{model_class.__name__} does not support SDPA")
|
||||
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
|
||||
model_sdpa = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
model_eager = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
low_cpu_mem_usage=True,
|
||||
attn_implementation="eager",
|
||||
).to(torch_device)
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
res_eager = model_eager.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
res_sdpa = model_sdpa.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=max_new_tokens, do_sample=False
|
||||
)
|
||||
|
||||
self.assertTrue(torch.allclose(res_eager, res_sdpa))
|
||||
|
||||
@require_torch_sdpa
|
||||
def test_sdpa_matches_eager_sliding_window(self):
|
||||
if not self.has_attentions:
|
||||
|
||||
Loading…
Reference in New Issue
Block a user