[ gemma] Adds support for Gemma 💎 (#29167)
* inital commit * update * update conversion checkpoint * update conversion script * nits * some fixes * nits * merge * fix permute * nits * fix * nits * nits * nits * fix rope * fix both rope * nites * style * make sure flax works * fix flax init code * fix foward * nits * print flax generation out * current code * nits * SIIIIIIIIIIIIIIIIIII * update * add new tokenizer * correct fast tokenizer * fix conversion * more comments * fix modeling and conversion * nits and nits * nits testing * add some tokenization tests * add some edge cases * add slow tests and fix them * fixup * fix copies for modeling * fix copies * add 7B slow tests * fix * fix * fix tests * make tokenizer cis go green * styling * last tokenizer nits * update jax tests * fix flax for 7b * add jit testing 🤗 * cleanups * isolated nit, inv_freq for rotary_emb.inv_freq * propagate to jax * Apply suggestions from code review Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com> * adjust test * fix conversion script * change name * correct file names * update conversion script * Fix bos and eos token ids in the model configuration (#3) * update modelling * update conversion script * add static cache for gemma * fix sdpa generate * fix batched * multiple fixes * fix FA2 * final fix * Rename a few missing strings and filenames (#4) * merge with upstream main * fix copies * fix copies * fix fixup * fix fixup * fix * fix * final tests * fix fx gemma tests * fix fx bf16/fp16 tests * update slow fx tests * fx slow tests: one logits, one generation * move jit test standalone * Apply suggestions from code review * nits * tokenizer updates * more tokenization updates: custom GemmaSentencepieceExtrator * style * Update src/transformers/cache_utils.py * Update src/transformers/models/gemma/__init__.py * Update tests/models/gemma/test_modeling_flax_gemma.py * small nits * style * update tokenization test * fix the rotary embedding * with style * fix slow tests * WARNING this commit might be very important for precisions * Update tests/models/gemma/test_modeling_flax_gemma.py * Update src/transformers/models/gemma/configuration_gemma.py Co-authored-by: Lysandre Debut <hi@lysand.re> * Update src/transformers/models/gemma/modeling_flax_gemma.py Co-authored-by: Lysandre Debut <hi@lysand.re> * small nits here and there! * forgotten nit * remove on the fly computation of inv_freq * revert previous change, let's be safe and for now re-compute freq cis to make sure it's in float * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update src/transformers/models/gemma/convert_gemma_weights_to_hf.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update src/transformers/models/gemma/convert_gemma_weights_to_hf.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_flax_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_tokenization_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_tokenization_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_tokenization_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_tokenization_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Update tests/models/gemma/test_modeling_gemma.py Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * nit conversion script link * fix some tests * add not doctest and pr doctest * repo consistency * fix last CIs 🚀 * update all readmes --------- Co-authored-by: younesbelkada <younesbelkada@gmail.com> Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com> Co-authored-by: sanchit-gandhi <sanchit@huggingface.co> Co-authored-by: Lysandre Debut <hi@lysand.re>
This commit is contained in:
Arthur
0
tests/models/gemma/__init__.py
Normal file
0
tests/models/gemma/__init__.py
Normal file
267
tests/models/gemma/test_modeling_flax_gemma.py
Normal file
267
tests/models/gemma/test_modeling_flax_gemma.py
Normal file
@@ -0,0 +1,267 @@
|
||||
# Copyright 2024 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.
|
||||
|
||||
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
|
||||
from transformers import AutoTokenizer, GemmaConfig, is_flax_available
|
||||
from transformers.testing_utils import require_flax, slow
|
||||
|
||||
from ...generation.test_flax_utils import FlaxGenerationTesterMixin
|
||||
from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_flax_available():
|
||||
import jax
|
||||
import jax.numpy as jnp
|
||||
|
||||
from transformers.models.gemma.modeling_flax_gemma import (
|
||||
FlaxGemmaForCausalLM,
|
||||
FlaxGemmaModel,
|
||||
)
|
||||
|
||||
|
||||
class FlaxGemmaModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=2,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_input_mask=True,
|
||||
use_token_type_ids=False,
|
||||
use_labels=True,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=2,
|
||||
num_attention_heads=4,
|
||||
num_key_value_heads=2,
|
||||
intermediate_size=37,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=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_token_type_ids = use_token_type_ids
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.initializer_range = initializer_range
|
||||
self.scope = None
|
||||
self.bos_token_id = vocab_size - 1
|
||||
self.eos_token_id = vocab_size - 1
|
||||
self.pad_token_id = vocab_size - 1
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
input_mask = None
|
||||
if self.use_input_mask:
|
||||
input_mask = np.tril(np.ones((self.batch_size, self.seq_length)))
|
||||
|
||||
config = GemmaConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
hidden_size=self.hidden_size,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
num_key_value_heads=self.num_key_value_heads,
|
||||
head_dim=self.hidden_size // self.num_attention_heads,
|
||||
intermediate_size=self.intermediate_size,
|
||||
hidden_act=self.hidden_act,
|
||||
hidden_dropout_prob=self.hidden_dropout_prob,
|
||||
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
use_cache=True,
|
||||
is_decoder=False,
|
||||
initializer_range=self.initializer_range,
|
||||
)
|
||||
|
||||
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_flax
|
||||
class FlaxGemmaModelTest(FlaxModelTesterMixin, FlaxGenerationTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (FlaxGemmaModel, FlaxGemmaForCausalLM) if is_flax_available() else ()
|
||||
all_generative_model_classes = (FlaxGemmaForCausalLM,) if is_flax_available() else ()
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = FlaxGemmaModelTester(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_model_from_pretrained(self):
|
||||
for model_class_name in self.all_model_classes:
|
||||
model = model_class_name.from_pretrained("google/gemma-2b", from_pt=True)
|
||||
outputs = model(np.ones((1, 1)))
|
||||
self.assertIsNotNone(outputs)
|
||||
|
||||
|
||||
@slow
|
||||
@require_flax
|
||||
class FlaxGemmaIntegrationTest(unittest.TestCase):
|
||||
input_text = ["The capital of France is", "To play the perfect cover drive"]
|
||||
model_id = "google/gemma-2b"
|
||||
revision = "flax"
|
||||
|
||||
def setUp(self):
|
||||
self.model, self.params = FlaxGemmaForCausalLM.from_pretrained(
|
||||
self.model_id, revision=self.revision, _do_init=False
|
||||
)
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(self.model_id)
|
||||
self.tokenizer.padding_side = "left"
|
||||
|
||||
def test_logits(self):
|
||||
inputs = self.tokenizer(self.input_text, return_tensors="np", padding=True)
|
||||
# fmt: off
|
||||
EXPECTED_MEAN = [
|
||||
[-16.427, -21.386, -35.491, -36.258, -31.401, -36.370, -37.598],
|
||||
[-21.386, -32.150, -33.155, -34.344, -34.706, -34.678, -38.495],
|
||||
]
|
||||
EXPECTED_SLICE = [-33.462, -16.481, -30.837, -32.195, -33.113]
|
||||
# fmt: on
|
||||
|
||||
logits = self.model(**inputs, params=self.params).logits
|
||||
|
||||
diff_mean = jnp.abs(logits.mean(-1) - np.array(EXPECTED_MEAN)).max()
|
||||
diff_slice = jnp.abs(logits[0, -1, 475:480] - np.array(EXPECTED_SLICE)).max()
|
||||
|
||||
self.assertAlmostEqual(diff_mean, 0, places=3)
|
||||
self.assertAlmostEqual(diff_slice, 0, places=3)
|
||||
|
||||
def test_generation(self):
|
||||
EXPECTED_TEXTS = [
|
||||
"The capital of France is a city of contrasts. It is a city of history, of art, of culture, of fashion",
|
||||
"To play the perfect cover drive, you need to have a good technique and a good mindset.\n\nThe cover drive is a shot",
|
||||
]
|
||||
inputs = self.tokenizer(self.input_text, return_tensors="np", padding=True)
|
||||
|
||||
output = self.model.generate(**inputs, params=self.params, max_new_tokens=20, do_sample=False)
|
||||
output_text = self.tokenizer.batch_decode(output.sequences, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
def test_jit_generation(self):
|
||||
EXPECTED_TEXTS = [
|
||||
"The capital of France is a city of contrasts. It is a city of history, culture, and art, but it is",
|
||||
"To play the perfect cover drive, you need to have a good technique and a good mindset.\n\nThe cover drive is a shot",
|
||||
]
|
||||
inputs = self.tokenizer(self.input_text, return_tensors="np", padding=True)
|
||||
|
||||
def generate(input_ids, attention_mask):
|
||||
outputs = self.model.generate(
|
||||
input_ids, attention_mask=attention_mask, params=self.params, max_new_tokens=20, do_sample=False
|
||||
)
|
||||
return outputs
|
||||
|
||||
jit_generate = jax.jit(generate)
|
||||
output_sequences = jit_generate(**inputs).sequences
|
||||
output_text = self.tokenizer.batch_decode(output_sequences, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
656
tests/models/gemma/test_modeling_gemma.py
Normal file
656
tests/models/gemma/test_modeling_gemma.py
Normal file
@@ -0,0 +1,656 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2024 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.
|
||||
""" Testing suite for the PyTorch Gemma model. """
|
||||
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
import pytest
|
||||
from parameterized import parameterized
|
||||
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer, GemmaConfig, is_torch_available
|
||||
from transformers.testing_utils import (
|
||||
require_bitsandbytes,
|
||||
require_flash_attn,
|
||||
require_torch,
|
||||
require_torch_gpu,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
|
||||
from ...generation.test_utils import GenerationTesterMixin
|
||||
from ...test_configuration_common import ConfigTester
|
||||
from ...test_modeling_common import ModelTesterMixin, ids_tensor
|
||||
from ...test_pipeline_mixin import PipelineTesterMixin
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
from transformers import GemmaForCausalLM, GemmaForSequenceClassification, GemmaModel
|
||||
|
||||
|
||||
class GemmaModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_input_mask=True,
|
||||
use_token_type_ids=False,
|
||||
use_labels=True,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=2,
|
||||
num_attention_heads=4,
|
||||
num_key_value_heads=2,
|
||||
intermediate_size=37,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=512,
|
||||
type_vocab_size=16,
|
||||
type_sequence_label_size=2,
|
||||
initializer_range=0.02,
|
||||
num_labels=3,
|
||||
num_choices=4,
|
||||
pad_token_id=0,
|
||||
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_token_type_ids = use_token_type_ids
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.type_sequence_label_size = type_sequence_label_size
|
||||
self.initializer_range = initializer_range
|
||||
self.num_labels = num_labels
|
||||
self.num_choices = num_choices
|
||||
self.pad_token_id = pad_token_id
|
||||
self.scope = scope
|
||||
self.head_dim = self.hidden_size // self.num_attention_heads
|
||||
|
||||
# Copied from tests.models.mistral.test_modeling_mistral.MistralModelTester.prepare_config_and_inputs
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
input_mask = None
|
||||
if self.use_input_mask:
|
||||
input_mask = torch.tril(torch.ones(self.batch_size, self.seq_length)).to(torch_device)
|
||||
|
||||
token_type_ids = None
|
||||
if self.use_token_type_ids:
|
||||
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
||||
|
||||
sequence_labels = None
|
||||
token_labels = None
|
||||
choice_labels = None
|
||||
if self.use_labels:
|
||||
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
||||
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
||||
choice_labels = ids_tensor([self.batch_size], self.num_choices)
|
||||
|
||||
config = self.get_config()
|
||||
|
||||
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
|
||||
# Ignore copy
|
||||
def get_config(self):
|
||||
return GemmaConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
hidden_size=self.hidden_size,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
num_key_value_heads=self.num_key_value_heads,
|
||||
intermediate_size=self.intermediate_size,
|
||||
hidden_act=self.hidden_act,
|
||||
hidden_dropout_prob=self.hidden_dropout_prob,
|
||||
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
type_vocab_size=self.type_vocab_size,
|
||||
is_decoder=False,
|
||||
initializer_range=self.initializer_range,
|
||||
pad_token_id=self.pad_token_id,
|
||||
head_dim=self.head_dim,
|
||||
)
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->Gemma
|
||||
def create_and_check_model(
|
||||
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
):
|
||||
model = GemmaModel(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))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Gemma
|
||||
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 = GemmaModel(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->Gemma
|
||||
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 = GemmaForCausalLM(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->Gemma
|
||||
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 = GemmaForCausalLM(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()
|
||||
(
|
||||
config,
|
||||
input_ids,
|
||||
token_type_ids,
|
||||
input_mask,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
) = config_and_inputs
|
||||
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
|
||||
return config, inputs_dict
|
||||
|
||||
|
||||
@require_torch
|
||||
class GemmaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (GemmaModel, GemmaForCausalLM, GemmaForSequenceClassification) if is_torch_available() else ()
|
||||
all_generative_model_classes = (GemmaForCausalLM,) if is_torch_available() else ()
|
||||
pipeline_model_mapping = (
|
||||
{
|
||||
"feature-extraction": GemmaModel,
|
||||
"text-classification": GemmaForSequenceClassification,
|
||||
"text-generation": GemmaForCausalLM,
|
||||
"zero-shot": GemmaForSequenceClassification,
|
||||
}
|
||||
if is_torch_available()
|
||||
else {}
|
||||
)
|
||||
test_headmasking = False
|
||||
test_pruning = False
|
||||
|
||||
# TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79245/workflows/9490ef58-79c2-410d-8f51-e3495156cf9c/jobs/1012146
|
||||
def is_pipeline_test_to_skip(
|
||||
self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
|
||||
):
|
||||
return True
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = GemmaModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=GemmaConfig, hidden_size=37)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_model(*config_and_inputs)
|
||||
|
||||
def test_model_various_embeddings(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
for type in ["absolute", "relative_key", "relative_key_query"]:
|
||||
config_and_inputs[0].position_embedding_type = type
|
||||
self.model_tester.create_and_check_model(*config_and_inputs)
|
||||
|
||||
def test_Gemma_sequence_classification_model(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
print(config)
|
||||
config.num_labels = 3
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
|
||||
model = GemmaForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
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))
|
||||
|
||||
def test_Gemma_sequence_classification_model_for_single_label(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.num_labels = 3
|
||||
config.problem_type = "single_label_classification"
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
|
||||
model = GemmaForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
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))
|
||||
|
||||
def test_Gemma_sequence_classification_model_for_multi_label(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.num_labels = 3
|
||||
config.problem_type = "multi_label_classification"
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor(
|
||||
[self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
|
||||
).to(torch.float)
|
||||
model = GemmaForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
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))
|
||||
|
||||
@unittest.skip("TODO @gante fix this for Llama")
|
||||
@parameterized.expand([(1, False), (1, True), (4, False)])
|
||||
def test_new_cache_format(self, num_beams, do_sample):
|
||||
pass
|
||||
|
||||
@unittest.skip("Gemma buffers include complex numbers, which breaks this test")
|
||||
def test_save_load_fast_init_from_base(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("Gemma uses GQA on all models so the KV cache is a non standard format")
|
||||
def test_past_key_values_format(self):
|
||||
pass
|
||||
|
||||
@require_flash_attn
|
||||
@require_torch_gpu
|
||||
@pytest.mark.flash_attn_test
|
||||
@slow
|
||||
def test_flash_attn_2_generate_padding_right(self):
|
||||
import torch
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, low_cpu_mem_usage=True).to(
|
||||
torch_device
|
||||
)
|
||||
|
||||
dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
|
||||
dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [1, 1, 1, 0]]).to(torch_device)
|
||||
|
||||
model.generate(dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False)
|
||||
|
||||
model = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
attn_implementation="flash_attention_2",
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
with self.assertRaises(ValueError):
|
||||
_ = model.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
|
||||
)
|
||||
|
||||
@require_flash_attn
|
||||
@require_torch_gpu
|
||||
@pytest.mark.flash_attn_test
|
||||
@slow
|
||||
def test_flash_attn_2_generate_use_cache(self):
|
||||
import torch
|
||||
|
||||
max_new_tokens = 30
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
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))
|
||||
# NOTE: Gemma apparently does not support right padding + use_cache with FA2.
|
||||
dummy_attention_mask[:, -1] = 1
|
||||
|
||||
model = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
attn_implementation="flash_attention_2",
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
_ = model.generate(
|
||||
dummy_input,
|
||||
attention_mask=dummy_attention_mask,
|
||||
max_new_tokens=max_new_tokens,
|
||||
do_sample=False,
|
||||
use_cache=True,
|
||||
)
|
||||
|
||||
@require_flash_attn
|
||||
@require_torch_gpu
|
||||
@pytest.mark.flash_attn_test
|
||||
@slow
|
||||
def test_flash_attn_2_inference_padding_right(self):
|
||||
self.skipTest("Gemma flash attention does not support right padding")
|
||||
|
||||
|
||||
@require_torch_gpu
|
||||
@slow
|
||||
class GemmaIntegrationTest(unittest.TestCase):
|
||||
input_text = ["Hello I am doing", "Hi today"]
|
||||
|
||||
def test_model_2b_fp32(self):
|
||||
model_id = "google/gemma-2b"
|
||||
EXPECTED_TEXTS = [
|
||||
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
|
||||
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True).to(torch_device)
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
def test_model_2b_fp16(self):
|
||||
model_id = "google/gemma-2b"
|
||||
EXPECTED_TEXTS = [
|
||||
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
|
||||
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16).to(
|
||||
torch_device
|
||||
)
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
def test_model_2b_fp16_static_cache(self):
|
||||
model_id = "google/gemma-2b"
|
||||
EXPECTED_TEXTS = [
|
||||
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
|
||||
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Kaju Kat",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16).to(
|
||||
torch_device
|
||||
)
|
||||
|
||||
model.generation_config.cache_implementation = "static"
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
def test_model_2b_bf16(self):
|
||||
model_id = "google/gemma-2b"
|
||||
EXPECTED_TEXTS = [
|
||||
"Hello I am doing a project on the 1990s and I need to know what the most popular music",
|
||||
"Hi today I am going to share with you a very easy and simple recipe of <strong><em>Khichdi",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16).to(
|
||||
torch_device
|
||||
)
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
@require_bitsandbytes
|
||||
def test_model_2b_4bit(self):
|
||||
model_id = "google/gemma-2b"
|
||||
EXPECTED_TEXTS = [
|
||||
"Hello I am doing a project and I need to make a 3d model of a house. I have been using",
|
||||
"Hi today I'd like to share with you my experience with the new wattpad wattpad wattpad wattpad wattpad wattpad wattpad",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, load_in_4bit=True)
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
@unittest.skip("The test will not fit our CI runners")
|
||||
def test_model_7b_fp32(self):
|
||||
model_id = "google/gemma-7b"
|
||||
EXPECTED_TEXTS = [
|
||||
"Hello my name is ***** ***** I will be assisting you today. I am sorry to hear about your issue. I will",
|
||||
"Hi,\n\nI have a problem with my 2005 1.6 16",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True).to(torch_device)
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
def test_model_7b_fp16(self):
|
||||
model_id = "google/gemma-7b"
|
||||
EXPECTED_TEXTS = [
|
||||
"""Hello I am doing a project on a 1999 4.0L 4x4. I""",
|
||||
"Hi today I am going to show you how to make a simple and easy to make a DIY 3D",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16).to(
|
||||
torch_device
|
||||
)
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
def test_model_7b_bf16(self):
|
||||
model_id = "google/gemma-7b"
|
||||
EXPECTED_TEXTS = [
|
||||
"""Hello I am doing a project on a 1991 240sx and I am trying to find""",
|
||||
"Hi today I am going to show you how to make a very simple and easy to make a very simple and",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16).to(
|
||||
torch_device
|
||||
)
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
def test_model_7b_fp16_static_cache(self):
|
||||
model_id = "google/gemma-7b"
|
||||
EXPECTED_TEXTS = [
|
||||
"""Hello I am doing a project on a 1999 4.0L 4x4. I""",
|
||||
"Hi today I am going to show you how to make a simple and easy to make a DIY 3D",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16).to(
|
||||
torch_device
|
||||
)
|
||||
|
||||
model.generation_config.cache_implementation = "static"
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
|
||||
@require_bitsandbytes
|
||||
def test_model_7b_4bit(self):
|
||||
model_id = "google/gemma-7b"
|
||||
EXPECTED_TEXTS = [
|
||||
"Hello I am doing a project for my school and I am trying to make a program that will take a number and then",
|
||||
"""Hi today I am going to talk about the new update for the game called "The new update" and I""",
|
||||
]
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, load_in_4bit=True)
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
|
||||
|
||||
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
|
||||
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(output_text, EXPECTED_TEXTS)
|
||||
497
tests/models/gemma/test_tokenization_gemma.py
Normal file
497
tests/models/gemma/test_tokenization_gemma.py
Normal file
@@ -0,0 +1,497 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2024 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.
|
||||
|
||||
import os
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
from datasets import load_dataset
|
||||
|
||||
from transformers import (
|
||||
AddedToken,
|
||||
GemmaTokenizer,
|
||||
GemmaTokenizerFast,
|
||||
is_torch_available,
|
||||
)
|
||||
from transformers.convert_slow_tokenizer import convert_slow_tokenizer
|
||||
from transformers.testing_utils import (
|
||||
get_tests_dir,
|
||||
nested_simplify,
|
||||
require_jinja,
|
||||
require_sentencepiece,
|
||||
require_tokenizers,
|
||||
require_torch,
|
||||
slow,
|
||||
)
|
||||
|
||||
from ...test_tokenization_common import TokenizerTesterMixin
|
||||
|
||||
|
||||
SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
pass
|
||||
|
||||
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class GemmaTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
||||
tokenizer_class = GemmaTokenizer
|
||||
rust_tokenizer_class = GemmaTokenizerFast
|
||||
|
||||
test_rust_tokenizer = False
|
||||
test_sentencepiece = True
|
||||
from_pretrained_kwargs = {}
|
||||
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
# We have a SentencePiece fixture for testing
|
||||
tokenizer = GemmaTokenizer(SAMPLE_VOCAB, keep_accents=True)
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
tokenizer.save_pretrained(self.tmpdirname)
|
||||
|
||||
@require_torch
|
||||
def test_batch_tokenization(self):
|
||||
if not self.test_seq2seq:
|
||||
return
|
||||
|
||||
tokenizers = self.get_tokenizers()
|
||||
for tokenizer in tokenizers:
|
||||
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
||||
# Longer text that will definitely require truncation.
|
||||
text = [
|
||||
" UN Chief Says There Is No Military Solution in Syria",
|
||||
" Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for"
|
||||
" Syria is that 'there is no military solution' to the nearly five-year conflict and more weapons"
|
||||
" will only worsen the violence and misery for millions of people.",
|
||||
]
|
||||
try:
|
||||
batch = tokenizer(
|
||||
text=text,
|
||||
max_length=3,
|
||||
max_target_length=10,
|
||||
return_tensors="pt",
|
||||
)
|
||||
except NotImplementedError:
|
||||
return
|
||||
self.assertEqual(batch.input_ids.shape[1], 3)
|
||||
# max_target_length will default to max_length if not specified
|
||||
batch = tokenizer(text, max_length=3, return_tensors="pt")
|
||||
self.assertEqual(batch.input_ids.shape[1], 3)
|
||||
|
||||
batch_encoder_only = tokenizer(text=text, max_length=3, max_target_length=10, return_tensors="pt")
|
||||
self.assertEqual(batch_encoder_only.input_ids.shape[1], 3)
|
||||
self.assertEqual(batch_encoder_only.attention_mask.shape[1], 3)
|
||||
self.assertNotIn("decoder_input_ids", batch_encoder_only)
|
||||
|
||||
@unittest.skip("Unfortunately way too slow to build a BPE with SentencePiece.")
|
||||
def test_save_slow_from_fast_and_reload_fast(self):
|
||||
pass
|
||||
|
||||
def test_special_tokens_initialization(self):
|
||||
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
|
||||
with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
|
||||
added_tokens = [AddedToken("<special>", lstrip=True)]
|
||||
|
||||
tokenizer_r = self.rust_tokenizer_class.from_pretrained(
|
||||
pretrained_name, additional_special_tokens=added_tokens, **kwargs
|
||||
)
|
||||
r_output = tokenizer_r.encode("Hey this is a <special> token")
|
||||
|
||||
special_token_id = tokenizer_r.encode("<special>", add_special_tokens=False)[0]
|
||||
|
||||
self.assertTrue(special_token_id in r_output)
|
||||
|
||||
if self.test_slow_tokenizer:
|
||||
tokenizer_cr = self.rust_tokenizer_class.from_pretrained(
|
||||
pretrained_name,
|
||||
additional_special_tokens=added_tokens,
|
||||
**kwargs, # , from_slow=True <- unfortunately too slow to convert
|
||||
)
|
||||
tokenizer_p = self.tokenizer_class.from_pretrained(
|
||||
pretrained_name, additional_special_tokens=added_tokens, **kwargs
|
||||
)
|
||||
|
||||
p_output = tokenizer_p.encode("Hey this is a <special> token")
|
||||
|
||||
cr_output = tokenizer_cr.encode("Hey this is a <special> token")
|
||||
|
||||
self.assertEqual(p_output, r_output)
|
||||
self.assertEqual(cr_output, r_output)
|
||||
self.assertTrue(special_token_id in p_output)
|
||||
self.assertTrue(special_token_id in cr_output)
|
||||
|
||||
@slow
|
||||
def test_tokenizer_integration(self):
|
||||
expected_encoding = {'input_ids': [[2, 158434, 591, 84193, 3836, 685, 6599, 31223, 235290, 140247, 578, 6599, 31223, 235290, 145139, 235290, 3491, 235275, 6572, 3311, 235290, 38197, 109959, 591, 25894, 235269, 162174, 235290, 235284, 235269, 1791, 6362, 12481, 235269, 1576, 18622, 235269, 2900, 1136, 86684, 235269, 29092, 4632, 16994, 604, 13146, 14944, 40371, 591, 19700, 235327, 235275, 578, 13146, 14944, 25511, 591, 235300, 12474, 235275, 675, 1163, 235248, 235304, 235284, 235340, 229903, 5377, 575, 235248, 235274, 235276, 235276, 235340, 17044, 578, 5271, 1061, 118345, 1865, 125247, 235269, 8745, 111226, 578, 176888, 235265], [2, 25894, 603, 6869, 577, 953, 235290, 8297, 5271, 209099, 41642, 774, 748, 78253, 2793, 731, 51506, 34346, 611, 2145, 2731, 578, 1833, 4807, 575, 832, 16630, 235265], [2, 651, 4320, 8426, 25341, 36271, 1163, 573, 27894, 5929, 235265]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]} # fmt: skip
|
||||
self.tokenizer_integration_test_util(
|
||||
expected_encoding=expected_encoding,
|
||||
model_name="hf-internal-testing/dummy-gemma",
|
||||
revision="",
|
||||
padding=False,
|
||||
)
|
||||
|
||||
@unittest.skip("worker 'gw4' crashed on CI, passing locally.")
|
||||
def test_pickle_subword_regularization_tokenizer(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("worker 'gw4' crashed on CI, passing locally.")
|
||||
def test_subword_regularization_tokenizer(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("This test will be removed from main @LysandreJik")
|
||||
def test_pretrained_model_lists(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("Skipping")
|
||||
def test_torch_encode_plus_sent_to_model(self):
|
||||
pass
|
||||
|
||||
|
||||
@require_torch
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class GemmaIntegrationTest(unittest.TestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
checkpoint_name = "hf-internal-testing/dummy-gemma"
|
||||
cls.tokenizer: GemmaTokenizer = GemmaTokenizer.from_pretrained(
|
||||
checkpoint_name, eos_token="<s>"
|
||||
) # add this token
|
||||
cls.rust_tokenizer = GemmaTokenizerFast.from_pretrained(
|
||||
checkpoint_name, eos_token="<s>", from_slow=True
|
||||
) # add this token
|
||||
return cls
|
||||
|
||||
@require_torch
|
||||
def integration_tests(self):
|
||||
inputs = self.tokenizer(
|
||||
["The following string should be properly encoded: Hello.", "But ird and ปี ird ด"],
|
||||
return_tensors="pt",
|
||||
)
|
||||
|
||||
self.assertEqual(
|
||||
nested_simplify(inputs),
|
||||
{
|
||||
"input_ids": [
|
||||
[2, 450, 1494, 1347, 881, 367, 6284, 18511, 29901, 15043, 29889],
|
||||
[2, 1205, 29871, 1823, 322, 29871, 31010, 30691, 1678, 1823, 1678, 30718],
|
||||
],
|
||||
"attention_mask": [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]],
|
||||
},
|
||||
)
|
||||
|
||||
def test_fast_special_tokens(self):
|
||||
slow_tokenizer = self.tokenizer
|
||||
fast_tokenizer = self.rust_tokenizer
|
||||
slow = slow_tokenizer.encode("A sample test", add_special_tokens=True)
|
||||
assert slow == [2, 235280, 6453, 2121]
|
||||
|
||||
fast_tokenizer.add_eos_token = False
|
||||
fast = fast_tokenizer.encode("A sample test", add_special_tokens=True)
|
||||
assert fast == [2, 235280, 6453, 2121]
|
||||
|
||||
fast_tokenizer.add_eos_token = True
|
||||
fast = fast_tokenizer.encode("A sample test", add_special_tokens=True)
|
||||
assert fast == [2, 235280, 6453, 2121, 204]
|
||||
|
||||
slow_tokenizer.add_eos_token = True
|
||||
slow = slow_tokenizer.encode("A sample test", add_special_tokens=True)
|
||||
assert slow == [2, 235280, 6453, 2121, 204]
|
||||
|
||||
self.tokenizer.add_eos_token = False
|
||||
self.rust_tokenizer.add_eos_token = False
|
||||
|
||||
@unittest.skip("Not super important and always failing. Let's skip it")
|
||||
@slow
|
||||
def test_conversion(self):
|
||||
# This is excruciatingly slow since it has to recreate the entire merge
|
||||
# list from the original vocabulary in spm
|
||||
self.rust_tokenizer.save_pretrained("./out")
|
||||
with tempfile.TemporaryDirectory() as dirname:
|
||||
self.rust_tokenizer.save_pretrained(dirname)
|
||||
|
||||
with open(os.path.join(dirname, "tokenizer.json"), "r") as f:
|
||||
old_serialized = f.read()
|
||||
|
||||
new_tokenizer = convert_slow_tokenizer(self.tokenizer)
|
||||
with tempfile.NamedTemporaryFile() as f:
|
||||
new_tokenizer.save(f.name)
|
||||
# Re-opening since `f` is in bytes.
|
||||
new_serialized = open(f.name, "r").read()
|
||||
with open("out_tokenizer.json", "w") as g:
|
||||
g.write(new_serialized)
|
||||
|
||||
self.assertEqual(old_serialized, new_serialized)
|
||||
|
||||
def test_simple_encode_decode(self):
|
||||
pyth_tokenizer = self.tokenizer
|
||||
rust_tokenizer = self.rust_tokenizer
|
||||
|
||||
self.tokenizer.add_eos_token = False
|
||||
self.rust_tokenizer.add_eos_token = False
|
||||
|
||||
self.assertEqual(pyth_tokenizer.encode("This is a test"), [2, 1596, 603, 476, 2121])
|
||||
self.assertEqual(rust_tokenizer.encode("This is a test"), [2, 1596, 603, 476, 2121])
|
||||
self.assertEqual(pyth_tokenizer.decode([2, 1596, 603, 476, 2121], skip_special_tokens=True), "This is a test")
|
||||
self.assertEqual(rust_tokenizer.decode([2, 1596, 603, 476, 2121], skip_special_tokens=True), "This is a test")
|
||||
|
||||
# bytefallback showcase
|
||||
self.assertEqual(pyth_tokenizer.encode("生活的真谛是"), [2, 122182, 235710, 245467, 235427] ) # fmt: skip
|
||||
self.assertEqual(rust_tokenizer.encode("生活的真谛是"), [2, 122182, 235710, 245467, 235427] ) # fmt: skip
|
||||
self.assertEqual(
|
||||
pyth_tokenizer.decode([2, 122182, 235710, 245467, 235427], skip_special_tokens=True),
|
||||
"生活的真谛是",
|
||||
)
|
||||
self.assertEqual(
|
||||
rust_tokenizer.decode([2, 122182, 235710, 245467, 235427], skip_special_tokens=True),
|
||||
"生活的真谛是",
|
||||
)
|
||||
|
||||
# Inner spaces showcase
|
||||
self.assertEqual(pyth_tokenizer.encode("Hi Hello"), [2, 2151, 139, 4521])
|
||||
self.assertEqual(rust_tokenizer.encode("Hi Hello"), [2, 2151, 139, 4521])
|
||||
self.assertEqual(pyth_tokenizer.decode([2, 2151, 139, 4521], skip_special_tokens=True), "Hi Hello")
|
||||
self.assertEqual(rust_tokenizer.decode([2, 2151, 139, 4521], skip_special_tokens=True), "Hi Hello")
|
||||
|
||||
self.assertEqual(pyth_tokenizer.encode("Hi Hello"), [2, 2151, 140, 4521])
|
||||
self.assertEqual(rust_tokenizer.encode("Hi Hello"), [2, 2151, 140, 4521])
|
||||
self.assertEqual(pyth_tokenizer.decode([2, 2151, 140, 4521], skip_special_tokens=True), "Hi Hello")
|
||||
self.assertEqual(rust_tokenizer.decode([2, 2151, 140, 4521], skip_special_tokens=True), "Hi Hello")
|
||||
|
||||
self.assertEqual(pyth_tokenizer.encode(""), [2])
|
||||
self.assertEqual(rust_tokenizer.encode(""), [2])
|
||||
|
||||
self.assertEqual(pyth_tokenizer.encode(" "), [2, 235248])
|
||||
self.assertEqual(rust_tokenizer.encode(" "), [2, 235248])
|
||||
|
||||
self.assertEqual(pyth_tokenizer.encode(" "), [2, 139])
|
||||
self.assertEqual(rust_tokenizer.encode(" "), [2, 139])
|
||||
|
||||
self.assertEqual(pyth_tokenizer.encode(" Hello"), [2, 25957])
|
||||
self.assertEqual(rust_tokenizer.encode(" Hello"), [2, 25957])
|
||||
|
||||
def test_no_differences_decode(self):
|
||||
self.tokenizer.add_eos_token = False
|
||||
self.rust_tokenizer.add_eos_token = False
|
||||
pyth_tokenizer = self.tokenizer
|
||||
rust_tokenizer = self.rust_tokenizer
|
||||
|
||||
self.assertEqual(pyth_tokenizer.decode([869]), "og")
|
||||
self.assertEqual(rust_tokenizer.decode([869]), "og")
|
||||
|
||||
self.assertEqual(pyth_tokenizer.decode([30112, 869]), " expenditureog")
|
||||
self.assertEqual(rust_tokenizer.decode([30112, 869]), " expenditureog")
|
||||
|
||||
def test_no_differences_special_tokens(self):
|
||||
pyth_tokenizer = self.tokenizer
|
||||
rust_tokenizer = self.rust_tokenizer
|
||||
self.assertEqual(pyth_tokenizer.encode(""), [2])
|
||||
self.assertEqual(rust_tokenizer.encode(""), [2])
|
||||
|
||||
self.assertEqual(pyth_tokenizer.encode("<s>"), [2, 204])
|
||||
self.assertEqual(rust_tokenizer.encode("<s>"), [2, 204])
|
||||
|
||||
@unittest.skipIf(
|
||||
os.getenv("RUN_TOKENIZER_INTEGRATION", "0") == "0",
|
||||
"RUN_TOKENIZER_INTEGRATION=1 to run tokenizer integration tests",
|
||||
)
|
||||
def test_integration_test_xnli(self):
|
||||
import tqdm
|
||||
|
||||
pyth_tokenizer = self.tokenizer
|
||||
rust_tokenizer = self.rust_tokenizer
|
||||
|
||||
dataset = load_dataset("code_x_glue_ct_code_to_text", "go")
|
||||
for item in tqdm.tqdm(dataset["validation"]):
|
||||
string = item["code"]
|
||||
encoded1 = pyth_tokenizer.encode(string)
|
||||
encoded2 = rust_tokenizer.encode(string)
|
||||
|
||||
self.assertEqual(encoded1, encoded2)
|
||||
|
||||
decoded1 = pyth_tokenizer.decode(encoded1, skip_special_tokens=True)
|
||||
decoded2 = rust_tokenizer.decode(encoded1, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(decoded1, decoded2)
|
||||
|
||||
dataset = load_dataset("xnli", "all_languages")
|
||||
|
||||
for item in tqdm.tqdm(dataset["train"]):
|
||||
for string in item["premise"].values():
|
||||
encoded1 = pyth_tokenizer.encode(string)
|
||||
encoded2 = rust_tokenizer.encode(string)
|
||||
|
||||
self.assertEqual(encoded1, encoded2)
|
||||
|
||||
decoded1 = pyth_tokenizer.decode(encoded1, skip_special_tokens=True)
|
||||
decoded2 = rust_tokenizer.decode(encoded2, skip_special_tokens=True)
|
||||
|
||||
self.assertEqual(decoded1, decoded2)
|
||||
|
||||
def test_special_token_special_word(self):
|
||||
# the word inform should be split as ['in', 'form']
|
||||
tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
|
||||
tokenizer.add_tokens([AddedToken("<REPR_END>", rstrip=True, lstrip=True)], special_tokens=False)
|
||||
out1 = tokenizer.decode(
|
||||
tokenizer.encode("<REPR_END>inform", add_special_tokens=False), spaces_between_special_tokens=False
|
||||
)
|
||||
self.assertEqual(out1, "<REPR_END>inform")
|
||||
out2 = tokenizer.decode(
|
||||
tokenizer.encode("<REPR_END>inform", add_special_tokens=False), spaces_between_special_tokens=True
|
||||
)
|
||||
# decoding strips the added prefix space.
|
||||
self.assertEqual(out2, "<REPR_END> inform")
|
||||
input_ids = tokenizer.encode("<REPR_END>inform", add_special_tokens=False)
|
||||
self.assertEqual(input_ids, [256000, 43910])
|
||||
|
||||
out2 = tokenizer.decode(
|
||||
tokenizer.encode(" <REPR_END> inform", add_special_tokens=False), spaces_between_special_tokens=False
|
||||
)
|
||||
# TODO @ArthurZ currently we strip left and right, so this will not keep the spaces
|
||||
self.assertEqual(out2, "<REPR_END>inform")
|
||||
|
||||
### Let's make sure decoding does not add extra spaces here and there
|
||||
# TODO @ArthurZ this should be affected by the lstrip/rstrip/single word /normalize refactoring
|
||||
# Since currently we always strip left and right of the token, results are as such
|
||||
input_ids = tokenizer.encode("<s> Hello<s>how", add_special_tokens=False)
|
||||
self.assertEqual(input_ids, [204, 25957, 204, 1139])
|
||||
tokens = tokenizer.tokenize("<s> Hello<s>how", add_special_tokens=False)
|
||||
self.assertEqual(tokens, ["<s>", "▁Hello", "<s>", "how"])
|
||||
decoded_tokens = tokenizer.decode(input_ids)
|
||||
self.assertEqual(decoded_tokens, "<s> Hello<s>how")
|
||||
|
||||
# Let's make sure that if there are any spaces, we don't remove them!
|
||||
input_ids = tokenizer.encode(" <s> Hello<s> how", add_special_tokens=False)
|
||||
self.assertEqual(input_ids, [235248, 204, 25957, 204, 1368])
|
||||
tokens = tokenizer.tokenize(" <s> Hello<s> how", add_special_tokens=False)
|
||||
self.assertEqual(tokens, ["▁", "<s>", "▁Hello", "<s>", "▁how"])
|
||||
decoded_tokens = tokenizer.decode(input_ids)
|
||||
self.assertEqual(decoded_tokens, " <s> Hello<s> how")
|
||||
|
||||
def test_some_edge_cases(self):
|
||||
tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
|
||||
|
||||
sp_tokens = tokenizer.sp_model.encode("<s>>", out_type=str)
|
||||
self.assertEqual(sp_tokens, ["<s>", ">"])
|
||||
tokens = tokenizer.tokenize("<s>>")
|
||||
self.assertEqual(sp_tokens, tokens)
|
||||
self.assertEqual(tokens, ["<s>", ">"])
|
||||
|
||||
tokens = tokenizer.tokenize("")
|
||||
self.assertEqual(tokens, [])
|
||||
self.assertEqual(tokens, tokenizer.sp_model.encode("", out_type=str))
|
||||
|
||||
tokens = tokenizer.tokenize(" ")
|
||||
self.assertEqual(tokens, ["▁"])
|
||||
# a dummy prefix space is not added by the sp_model as it was de-activated
|
||||
self.assertEqual(tokens, tokenizer.sp_model.encode(" ", out_type=str))
|
||||
|
||||
tokens = tokenizer.tokenize("▁")
|
||||
self.assertEqual(tokens, ["▁"])
|
||||
# a dummy prefix space is not added by the sp_model as it was de-activated
|
||||
self.assertEqual(tokens, tokenizer.sp_model.encode("▁", out_type=str))
|
||||
|
||||
tokens = tokenizer.tokenize(" ▁")
|
||||
self.assertEqual(tokens, ["▁▁"])
|
||||
# a dummy prefix space is not added by the sp_model as it was de-activated
|
||||
self.assertEqual(tokens, tokenizer.sp_model.encode("▁▁", out_type=str))
|
||||
|
||||
@require_jinja
|
||||
def test_tokenization_for_chat(self):
|
||||
tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
|
||||
|
||||
test_chats = [
|
||||
[{"role": "user", "content": "Hello!"}],
|
||||
[
|
||||
{"role": "user", "content": "Hello!"},
|
||||
{"role": "assistant", "content": "Nice to meet you."},
|
||||
],
|
||||
[{"role": "user", "content": "Hello!"}],
|
||||
]
|
||||
# Matt: The third test case tests the default system message, but if this is ever changed in the
|
||||
# class/repo code then that test will fail, and the case will need to be updated.
|
||||
tokenized_chats = [tokenizer.apply_chat_template(test_chat) for test_chat in test_chats]
|
||||
expected_tokens = [[235322, 235371, 571, 235298, 2997, 73786, 1645, 108, 4521, 149907, 235371, 571, 235298, 615, 73786, 108], [235322, 235371, 571, 235298, 2997, 73786, 1645, 108, 4521, 149907, 235371, 571, 235298, 615, 73786, 108, 235322, 235371, 571, 235298, 2997, 73786, 105776, 108, 7731, 577, 4664, 692, 35606, 235371, 571, 235298, 615, 73786, 108], [235322, 235371, 571, 235298, 2997, 73786, 1645, 108, 4521, 149907, 235371, 571, 235298, 615, 73786, 108]] # fmt: skip
|
||||
for tokenized_chat, expected_tokens in zip(tokenized_chats, expected_tokens):
|
||||
self.assertListEqual(tokenized_chat, expected_tokens)
|
||||
|
||||
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class CommonSpmIntegrationTests(unittest.TestCase):
|
||||
"""
|
||||
A class that regroups important test to make sure that we properly handle the special tokens.
|
||||
"""
|
||||
|
||||
def test_edge_case_tabulation(self):
|
||||
fast_tokenizer = GemmaTokenizerFast.from_pretrained("hf-internal-testing/dummy-gemma")
|
||||
slow_tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
|
||||
input_text = "Hey<eos>. \t\t \n\nyou é @#😈 🤗! , 1234 15 5,61"
|
||||
EXPECTED_IDS = [ 2, 6750, 1, 235265, 235248, 255969, 235248, 109, 4747, 139, 235335, 139, 216311, 241316, 139, 239880, 235341, 144, 235269, 235248, 235274, 235284, 235304, 235310, 235248, 235274, 235308, 235248, 235308, 235269, 235318, 235274] # fmt: skip
|
||||
EXPECTED_TOKENS = [ "Hey", "<eos>", ".", "▁", "\t\t", "▁", "\n\n", "you", "▁▁", "é", "▁▁", "@#", "😈", "▁▁", "🤗", "!", "▁▁▁▁▁▁▁", ",", "▁", "1", "2", "3", "4", "▁", "1", "5", "▁", "5", ",", "6", "1"] # fmt: skip
|
||||
|
||||
tokens = fast_tokenizer.tokenize(input_text)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(tokens, EXPECTED_TOKENS)
|
||||
|
||||
tokens = slow_tokenizer.tokenize(input_text)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(tokens, EXPECTED_TOKENS)
|
||||
|
||||
input_ids = fast_tokenizer.encode(input_text)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(input_ids, EXPECTED_IDS)
|
||||
|
||||
input_ids = slow_tokenizer.encode(input_text)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(input_ids, EXPECTED_IDS)
|
||||
|
||||
text = fast_tokenizer.decode(EXPECTED_IDS)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(text, "<bos>Hey<eos>. \t\t \n\nyou é @#😈 🤗! , 1234 15 5,61")
|
||||
|
||||
text = slow_tokenizer.decode(EXPECTED_IDS)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(text, "<bos>Hey<eos>. \t\t \n\nyou é @#😈 🤗! , 1234 15 5,61")
|
||||
|
||||
input_text = "\t\t\t\t \n\n61"
|
||||
EXPECTED_IDS = [2, 255971, 235248, 109, 235318, 235274]
|
||||
EXPECTED_TOKENS = ["\t\t\t\t", "▁", "\n\n", "6", "1"]
|
||||
|
||||
tokens = fast_tokenizer.tokenize(input_text)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(tokens, EXPECTED_TOKENS)
|
||||
|
||||
tokens = slow_tokenizer.tokenize(input_text)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(tokens, EXPECTED_TOKENS)
|
||||
|
||||
input_ids = fast_tokenizer.encode(input_text)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(input_ids, EXPECTED_IDS)
|
||||
|
||||
input_ids = slow_tokenizer.encode(input_text)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(input_ids, EXPECTED_IDS)
|
||||
|
||||
text = fast_tokenizer.decode(EXPECTED_IDS)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(text, "<bos>\t\t\t\t \n\n61")
|
||||
|
||||
text = slow_tokenizer.decode(EXPECTED_IDS)
|
||||
with self.subTest("test fast edge case fast"):
|
||||
self.assertEqual(text, "<bos>\t\t\t\t \n\n61")
|
||||
Reference in New Issue
Block a user