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

Deprecate TF + JAX (#38758)

Browse Source 
* Scatter deprecation warnings around

* Delete the tests

* Make logging work properly!
This commit is contained in:
Matt
2025-06-11 17:28:06 +01:00
committed by GitHub
parent 337757cbd5
commit 9f563ada70
121 changed files with 27 additions and 50030 deletions
Download Patch File Download Diff File Expand all files Collapse all files

217
tests/models/xglm/test_modeling_flax_xglm.py
View File

@@ -1,217 +0,0 @@
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
from transformers import XGLMConfig, XGLMTokenizer, is_flax_available
from transformers.testing_utils import require_flax, require_sentencepiece, slow
from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor, random_attention_mask
if is_flax_available():
import jax
import jax.numpy as jnp
import numpy as np
from transformers.models.xglm.modeling_flax_xglm import FlaxXGLMForCausalLM, FlaxXGLMModel
@require_flax
class FlaxXGLMModelTester:
def __init__(
self,
parent,
batch_size=14,
seq_length=7,
is_training=True,
use_input_mask=True,
use_labels=True,
vocab_size=99,
d_model=32,
num_hidden_layers=2,
num_attention_heads=4,
ffn_dim=37,
activation_function="gelu",
activation_dropout=0.1,
attention_dropout=0.1,
max_position_embeddings=512,
initializer_range=0.02,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = d_model
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.ffn_dim = ffn_dim
self.activation_function = activation_function
self.activation_dropout = activation_dropout
self.attention_dropout = attention_dropout
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.scope = None
self.bos_token_id = 0
self.eos_token_id = 2
self.pad_token_id = 1
def prepare_config_and_inputs(self):
input_ids = np.clip(ids_tensor([self.batch_size, self.seq_length], self.vocab_size), 3, self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
config = XGLMConfig(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
num_layers=self.num_hidden_layers,
attention_heads=self.num_attention_heads,
ffn_dim=self.ffn_dim,
activation_function=self.activation_function,
activation_dropout=self.activation_dropout,
attention_dropout=self.attention_dropout,
max_position_embeddings=self.max_position_embeddings,
initializer_range=self.initializer_range,
use_cache=True,
bos_token_id=self.bos_token_id,
eos_token_id=self.eos_token_id,
pad_token_id=self.pad_token_id,
)
return (config, input_ids, input_mask)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, input_ids, attention_mask = config_and_inputs
inputs_dict = {"input_ids": input_ids, "attention_mask": attention_mask}
return config, inputs_dict
def check_use_cache_forward(self, model_class_name, config, input_ids, attention_mask):
max_decoder_length = 20
model = model_class_name(config)
past_key_values = model.init_cache(input_ids.shape[0], max_decoder_length)
attention_mask = jnp.ones((input_ids.shape[0], max_decoder_length), dtype="i4")
position_ids = jnp.broadcast_to(
jnp.arange(input_ids.shape[-1] - 1)[None, :], (input_ids.shape[0], input_ids.shape[-1] - 1)
)
outputs_cache = model(
input_ids[:, :-1],
attention_mask=attention_mask,
past_key_values=past_key_values,
position_ids=position_ids,
)
position_ids = jnp.array(input_ids.shape[0] * [[input_ids.shape[-1] - 1]], dtype="i4")
outputs_cache_next = model(
input_ids[:, -1:],
attention_mask=attention_mask,
past_key_values=outputs_cache.past_key_values,
position_ids=position_ids,
)
outputs = model(input_ids)
diff = np.max(np.abs(outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5]))
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
def check_use_cache_forward_with_attn_mask(self, model_class_name, config, input_ids, attention_mask):
max_decoder_length = 20
model = model_class_name(config)
attention_mask_cache = jnp.concatenate(
[attention_mask, jnp.zeros((attention_mask.shape[0], max_decoder_length - attention_mask.shape[1]))],
axis=-1,
)
past_key_values = model.init_cache(input_ids.shape[0], max_decoder_length)
position_ids = jnp.broadcast_to(
jnp.arange(input_ids.shape[-1] - 1)[None, :], (input_ids.shape[0], input_ids.shape[-1] - 1)
)
outputs_cache = model(
input_ids[:, :-1],
attention_mask=attention_mask_cache,
past_key_values=past_key_values,
position_ids=position_ids,
)
position_ids = jnp.array(input_ids.shape[0] * [[input_ids.shape[-1] - 1]], dtype="i4")
outputs_cache_next = model(
input_ids[:, -1:],
past_key_values=outputs_cache.past_key_values,
attention_mask=attention_mask_cache,
position_ids=position_ids,
)
outputs = model(input_ids, attention_mask=attention_mask)
diff = np.max(np.abs(outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5]))
self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
@require_sentencepiece
@require_flax
class FlaxXGLMModelTest(FlaxModelTesterMixin, unittest.TestCase):
all_model_classes = (FlaxXGLMModel, FlaxXGLMForCausalLM) if is_flax_available() else ()
def setUp(self):
self.model_tester = FlaxXGLMModelTester(self)
def test_use_cache_forward(self):
for model_class_name in self.all_model_classes:
config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs()
self.model_tester.check_use_cache_forward(model_class_name, config, input_ids, attention_mask)
def test_use_cache_forward_with_attn_mask(self):
for model_class_name in self.all_model_classes:
config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs()
self.model_tester.check_use_cache_forward_with_attn_mask(
model_class_name, config, input_ids, attention_mask
)
@slow
def test_batch_generation(self):
tokenizer = XGLMTokenizer.from_pretrained("XGLM", padding_side="left")
inputs = tokenizer(["Hello this is a long string", "Hey"], return_tensors="np", padding=True, truncation=True)
model = FlaxXGLMForCausalLM.from_pretrained("facebook/xglm-564M")
model.config.num_beams = 1
model.config.do_sample = False
jit_generate = jax.jit(model.generate)
output_sequences = jit_generate(inputs["input_ids"], attention_mask=inputs["attention_mask"]).sequences
output_string = tokenizer.batch_decode(output_sequences, skip_special_tokens=True)
expected_string = [
"Hello this is a long string of questions, but I'm not sure if I'm",
"Hey, I'm a newbie to the forum and I'",
]
self.assertListEqual(output_string, expected_string)
@slow
def test_model_from_pretrained(self):
for model_class_name in self.all_model_classes:
model = model_class_name.from_pretrained("facebook/xglm-564M")
outputs = model(np.ones((1, 1)))
self.assertIsNotNone(outputs)

259
tests/models/xglm/test_modeling_tf_xglm.py
View File

@@ -1,259 +0,0 @@
# Copyright 2022 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
import unittest
from transformers import XGLMConfig, XGLMTokenizer, is_tf_available
from transformers.testing_utils import require_tf, slow
from ...test_configuration_common import ConfigTester
from ...test_modeling_tf_common import TFModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
from ...test_pipeline_mixin import PipelineTesterMixin
if is_tf_available():
import tensorflow as tf
from transformers.models.xglm.modeling_tf_xglm import (
TFXGLMForCausalLM,
TFXGLMModel,
)
@require_tf
class TFXGLMModelTester:
config_cls = XGLMConfig
config_updates = {}
hidden_act = "gelu"
def __init__(
self,
parent,
batch_size=14,
seq_length=7,
is_training=True,
use_input_mask=True,
use_labels=True,
vocab_size=99,
d_model=32,
num_hidden_layers=2,
num_attention_heads=4,
ffn_dim=37,
activation_function="gelu",
activation_dropout=0.1,
attention_dropout=0.1,
max_position_embeddings=512,
initializer_range=0.02,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = d_model
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.ffn_dim = ffn_dim
self.activation_function = activation_function
self.activation_dropout = activation_dropout
self.attention_dropout = attention_dropout
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.scope = None
self.bos_token_id = 0
self.eos_token_id = 2
self.pad_token_id = 1
def get_large_model_config(self):
return XGLMConfig.from_pretrained("facebook/xglm-564M")
def prepare_config_and_inputs(self):
input_ids = tf.clip_by_value(
ids_tensor([self.batch_size, self.seq_length], self.vocab_size), clip_value_min=0, clip_value_max=3
)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
config = self.get_config()
head_mask = floats_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
return (
config,
input_ids,
input_mask,
head_mask,
)
def get_config(self):
return XGLMConfig(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
num_layers=self.num_hidden_layers,
attention_heads=self.num_attention_heads,
ffn_dim=self.ffn_dim,
activation_function=self.activation_function,
activation_dropout=self.activation_dropout,
attention_dropout=self.attention_dropout,
max_position_embeddings=self.max_position_embeddings,
initializer_range=self.initializer_range,
use_cache=True,
bos_token_id=self.bos_token_id,
eos_token_id=self.eos_token_id,
pad_token_id=self.pad_token_id,
return_dict=True,
)
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
input_mask,
head_mask,
) = config_and_inputs
inputs_dict = {
"input_ids": input_ids,
"head_mask": head_mask,
}
return config, inputs_dict
@require_tf
class TFXGLMModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (TFXGLMModel, TFXGLMForCausalLM) if is_tf_available() else ()
all_generative_model_classes = (TFXGLMForCausalLM,) if is_tf_available() else ()
pipeline_model_mapping = (
{"feature-extraction": TFXGLMModel, "text-generation": TFXGLMForCausalLM} if is_tf_available() else {}
)
test_onnx = False
test_missing_keys = False
test_pruning = False
def setUp(self):
self.model_tester = TFXGLMModelTester(self)
self.config_tester = ConfigTester(self, config_class=XGLMConfig, n_embd=37)
def test_config(self):
self.config_tester.run_common_tests()
@slow
def test_model_from_pretrained(self):
model_name = "facebook/xglm-564M"
model = TFXGLMModel.from_pretrained(model_name)
self.assertIsNotNone(model)
@unittest.skip(reason="Currently, model embeddings are going to undergo a major refactor.")
def test_resize_token_embeddings(self):
super().test_resize_token_embeddings()
@require_tf
class TFXGLMModelLanguageGenerationTest(unittest.TestCase):
@slow
def test_lm_generate_xglm(self, verify_outputs=True):
model = TFXGLMForCausalLM.from_pretrained("facebook/xglm-564M")
input_ids = tf.convert_to_tensor([[2, 268, 9865]], dtype=tf.int32) # The dog
# </s> The dog is a very friendly dog. He is very affectionate and loves to play with other
expected_output_ids = [2, 268, 9865, 67, 11, 1988, 57252, 9865, 5, 984, 67, 1988, 213838, 1658, 53, 70446, 33, 6657, 278, 1581] # fmt: skip
output_ids = model.generate(input_ids, do_sample=False, num_beams=1)
if verify_outputs:
self.assertListEqual(output_ids[0].numpy().tolist(), expected_output_ids)
@slow
def test_xglm_sample(self):
tokenizer = XGLMTokenizer.from_pretrained("facebook/xglm-564M")
model = TFXGLMForCausalLM.from_pretrained("facebook/xglm-564M")
tf.random.set_seed(0)
tokenized = tokenizer("Today is a nice day and", return_tensors="tf")
input_ids = tokenized.input_ids
# forces the generation to happen on CPU, to avoid GPU-related quirks (and assure same output regardless of the available devices)
with tf.device(":/CPU:0"):
output_ids = model.generate(input_ids, do_sample=True, seed=[7, 0])
output_str = tokenizer.decode(output_ids[0], skip_special_tokens=True)
EXPECTED_OUTPUT_STR = (
"Today is a nice day and warm evening here over Southern Alberta!! Today when they closed schools due"
)
self.assertEqual(output_str, EXPECTED_OUTPUT_STR)
@slow
def test_batch_generation(self):
model = TFXGLMForCausalLM.from_pretrained("facebook/xglm-564M")
tokenizer = XGLMTokenizer.from_pretrained("facebook/xglm-564M")
tokenizer.padding_side = "left"
# use different length sentences to test batching
sentences = [
"This is an extremely long sentence that only exists to test the ability of the model to cope with "
"left-padding, such as in batched generation. The output for the sequence below should be the same "
"regardless of whether left padding is applied or not. When",
"Hello, my dog is a little",
]
inputs = tokenizer(sentences, return_tensors="tf", padding=True)
input_ids = inputs["input_ids"]
outputs = model.generate(input_ids=input_ids, attention_mask=inputs["attention_mask"], max_new_tokens=12)
inputs_non_padded = tokenizer(sentences[0], return_tensors="tf").input_ids
output_non_padded = model.generate(input_ids=inputs_non_padded, max_new_tokens=12)
inputs_padded = tokenizer(sentences[1], return_tensors="tf").input_ids
output_padded = model.generate(input_ids=inputs_padded, max_new_tokens=12)
batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
non_padded_sentence = tokenizer.decode(output_non_padded[0], skip_special_tokens=True)
padded_sentence = tokenizer.decode(output_padded[0], skip_special_tokens=True)
expected_output_sentence = [
"This is an extremely long sentence that only exists to test the ability of the model to cope with "
"left-padding, such as in batched generation. The output for the sequence below should be the same "
"regardless of whether left padding is applied or not. When left padding is applied, the sequence will be "
"a single",
"Hello, my dog is a little bit of a shy one, but he is very friendly",
]
self.assertListEqual(expected_output_sentence, batch_out_sentence)
self.assertListEqual(expected_output_sentence, [non_padded_sentence, padded_sentence])
@slow
def test_loss_with_padding(self):
tokenizer = XGLMTokenizer.from_pretrained("facebook/xglm-564M")
model = TFXGLMForCausalLM.from_pretrained("facebook/xglm-564M")
tokenizer.padding_side = "right"
sequence = "Sequence"
tokenized_non_padded = tokenizer(sequence, return_tensors="tf")
labels_non_padded = tokenized_non_padded.input_ids
loss_non_padded = model(tokenized_non_padded, labels=labels_non_padded).loss
tokenized_padded = tokenizer(sequence, padding="max_length", max_length=16, return_tensors="tf")
labels_padded = tokenized_padded.input_ids
labels_padded = tf.where(labels_padded == tokenizer.pad_token_id, -100, labels_padded)
loss_padded = model(tokenized_padded, labels=labels_padded).loss
tf.debugging.assert_near(loss_non_padded, loss_padded, atol=1e-3)