🚨 🚨 Inherited CausalLM Tests (#37590)
* stash commit * Experiment 1: Try just Gemma * Experiment 1: Just try Gemma * make fixup * Trigger tests * stash commit * Try adding Gemma3 as well * make fixup * Correct attrib names * Correct pipeline model mapping * Add in all_model_classes for Gemma1 again * Move the pipeline model mapping around again * make fixup * Revert Gemma3 changes since it's a VLM * Let's try Falcon * Correct attributes * Correct attributes * Let's try just overriding get_config() for now * Do Nemotron too * And Llama! * Do llama/persimmon * Correctly skip tests * Fix Persimmon * Include Phimoe * Fix Gemma2 * Set model_tester_class correctly * Add GLM * More models! * models models models * make fixup * Add Qwen3 + Qwen3MoE * Correct import * make fixup * Add the QuestionAnswering classes * Add the QuestionAnswering classes * Move pipeline mapping to the right place * Jetmoe too * Stop RoPE testing models with no RoPE * Fix up JetMOE a bit * Fix up JetMOE a bit * Can we just force pad_token_id all the time? * make fixup * fix starcoder2 * Move pipeline mapping * Fix RoPE skipping * Fix RecurrentGemma tests * Fix Falcon tests * Add MoE attributes * Fix values for RoPE testing * Make sure we set bos_token_id and eos_token_id in an appropriate range * make fixup * Fix GLM4 * Add mamba attributes * Revert bits of JetMOE * Re-add the JetMOE skips * Update tests/causal_lm_tester.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Add licence --------- Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
This commit is contained in:
Matt
@@ -16,19 +16,14 @@
|
||||
|
||||
import unittest
|
||||
|
||||
from parameterized import parameterized
|
||||
|
||||
from transformers import PhiConfig, is_torch_available, set_seed
|
||||
from transformers import PhiConfig, is_torch_available
|
||||
from transformers.testing_utils import (
|
||||
require_torch,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
|
||||
from ...generation.test_utils import GenerationTesterMixin
|
||||
from ...test_configuration_common import ConfigTester
|
||||
from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
|
||||
from ...test_pipeline_mixin import PipelineTesterMixin
|
||||
from ...causal_lm_tester import CausalLMModelTest, CausalLMModelTester
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
@@ -44,124 +39,17 @@ if is_torch_available():
|
||||
from transformers.models.phi.modeling_phi import PhiRotaryEmbedding
|
||||
|
||||
|
||||
class PhiModelTester:
|
||||
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,
|
||||
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.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
|
||||
|
||||
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 = random_attention_mask([self.batch_size, self.seq_length])
|
||||
|
||||
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
|
||||
|
||||
def get_config(self):
|
||||
return PhiConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
hidden_size=self.hidden_size,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=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,
|
||||
type_vocab_size=self.type_vocab_size,
|
||||
is_decoder=False,
|
||||
initializer_range=self.initializer_range,
|
||||
pad_token_id=self.pad_token_id,
|
||||
)
|
||||
|
||||
def create_and_check_model(
|
||||
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
):
|
||||
model = PhiModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=input_mask)
|
||||
result = model(input_ids)
|
||||
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
||||
|
||||
def 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
|
||||
class PhiModelTester(CausalLMModelTester):
|
||||
config_class = PhiConfig
|
||||
if is_torch_available():
|
||||
base_model_class = PhiModel
|
||||
causal_lm_class = PhiForCausalLM
|
||||
sequence_class = PhiForSequenceClassification
|
||||
token_class = PhiForTokenClassification
|
||||
|
||||
|
||||
@require_torch
|
||||
class PhiModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
||||
class PhiModelTest(CausalLMModelTest, unittest.TestCase):
|
||||
all_model_classes = (
|
||||
(PhiModel, PhiForCausalLM, PhiForSequenceClassification, PhiForTokenClassification)
|
||||
if is_torch_available()
|
||||
@@ -171,9 +59,8 @@ class PhiModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
|
||||
{
|
||||
"feature-extraction": PhiModel,
|
||||
"text-classification": PhiForSequenceClassification,
|
||||
"text-generation": PhiForCausalLM,
|
||||
"token-classification": PhiForTokenClassification,
|
||||
"zero-shot": PhiForSequenceClassification,
|
||||
"text-generation": PhiForCausalLM,
|
||||
}
|
||||
if is_torch_available()
|
||||
else {}
|
||||
@@ -181,6 +68,8 @@ class PhiModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
|
||||
|
||||
test_headmasking = False
|
||||
test_pruning = False
|
||||
model_tester_class = PhiModelTester
|
||||
rotary_embedding_layer = PhiRotaryEmbedding
|
||||
|
||||
# TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79292/workflows/fa2ba644-8953-44a6-8f67-ccd69ca6a476/jobs/1012905
|
||||
def is_pipeline_test_to_skip(
|
||||
@@ -195,146 +84,6 @@ class PhiModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
|
||||
):
|
||||
return True
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.setUp with Llama->Phi
|
||||
def setUp(self):
|
||||
self.model_tester = PhiModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=PhiConfig, hidden_size=37)
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_config
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_model
|
||||
def test_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_model(*config_and_inputs)
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model with Llama->Phi,llama->phi
|
||||
def test_phi_sequence_classification_model(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
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 = PhiForSequenceClassification(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))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model_for_single_label with Llama->Phi,llama->phi
|
||||
def test_phi_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 = PhiForSequenceClassification(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))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model_for_multi_label with Llama->Phi,llama->phi
|
||||
def test_phi_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 = PhiForSequenceClassification(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))
|
||||
|
||||
@parameterized.expand([("linear",), ("dynamic",)])
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_model_rope_scaling_from_config with Llama->Phi
|
||||
def test_model_rope_scaling_from_config(self, scaling_type):
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
short_input = ids_tensor([1, 10], config.vocab_size)
|
||||
long_input = ids_tensor([1, int(config.max_position_embeddings * 1.5)], config.vocab_size)
|
||||
|
||||
set_seed(42) # Fixed seed at init time so the two models get the same random weights
|
||||
original_model = PhiModel(config)
|
||||
original_model.to(torch_device)
|
||||
original_model.eval()
|
||||
original_short_output = original_model(short_input).last_hidden_state
|
||||
original_long_output = original_model(long_input).last_hidden_state
|
||||
|
||||
set_seed(42) # Fixed seed at init time so the two models get the same random weights
|
||||
config.rope_scaling = {"type": scaling_type, "factor": 10.0}
|
||||
scaled_model = PhiModel(config)
|
||||
scaled_model.to(torch_device)
|
||||
scaled_model.eval()
|
||||
scaled_short_output = scaled_model(short_input).last_hidden_state
|
||||
scaled_long_output = scaled_model(long_input).last_hidden_state
|
||||
|
||||
# Dynamic scaling does not change the RoPE embeddings until it receives an input longer than the original
|
||||
# maximum sequence length, so the outputs for the short input should match.
|
||||
if scaling_type == "dynamic":
|
||||
torch.testing.assert_close(original_short_output, scaled_short_output, rtol=1e-5, atol=1e-5)
|
||||
else:
|
||||
self.assertFalse(torch.allclose(original_short_output, scaled_short_output, atol=1e-5))
|
||||
|
||||
# The output should be different for long inputs
|
||||
self.assertFalse(torch.allclose(original_long_output, scaled_long_output, atol=1e-5))
|
||||
|
||||
# Copied from tests.models.gpt_neox.test_modeling_gpt_neox.GPTNeoXModelTest.test_model_rope_scaling with GPTNeoX->Phi
|
||||
def test_model_rope_scaling(self):
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
scaling_factor = 10
|
||||
short_input_length = 10
|
||||
long_input_length = int(config.max_position_embeddings * 1.5)
|
||||
|
||||
# Inputs
|
||||
x = torch.randn(
|
||||
1, dtype=torch.float32, device=torch_device
|
||||
) # used exclusively to get the dtype and the device
|
||||
position_ids_short = torch.arange(short_input_length, dtype=torch.long, device=torch_device)
|
||||
position_ids_short = position_ids_short.unsqueeze(0)
|
||||
position_ids_long = torch.arange(long_input_length, dtype=torch.long, device=torch_device)
|
||||
position_ids_long = position_ids_long.unsqueeze(0)
|
||||
|
||||
# Sanity check original RoPE
|
||||
original_rope = PhiRotaryEmbedding(config).to(torch_device)
|
||||
original_cos_short, original_sin_short = original_rope(x, position_ids_short)
|
||||
original_cos_long, original_sin_long = original_rope(x, position_ids_long)
|
||||
torch.testing.assert_close(original_cos_short, original_cos_long[:, :short_input_length, :])
|
||||
torch.testing.assert_close(original_sin_short, original_sin_long[:, :short_input_length, :])
|
||||
|
||||
# Sanity check linear RoPE scaling
|
||||
# New position "x" should match original position with index "x/scaling_factor"
|
||||
config.rope_scaling = {"type": "linear", "factor": scaling_factor}
|
||||
linear_scaling_rope = PhiRotaryEmbedding(config).to(torch_device)
|
||||
linear_cos_short, linear_sin_short = linear_scaling_rope(x, position_ids_short)
|
||||
linear_cos_long, linear_sin_long = linear_scaling_rope(x, position_ids_long)
|
||||
torch.testing.assert_close(linear_cos_short, linear_cos_long[:, :short_input_length, :])
|
||||
torch.testing.assert_close(linear_sin_short, linear_sin_long[:, :short_input_length, :])
|
||||
for new_position in range(0, long_input_length, scaling_factor):
|
||||
original_position = int(new_position // scaling_factor)
|
||||
torch.testing.assert_close(linear_cos_long[:, new_position, :], original_cos_long[:, original_position, :])
|
||||
torch.testing.assert_close(linear_sin_long[:, new_position, :], original_sin_long[:, original_position, :])
|
||||
|
||||
# Sanity check Dynamic NTK RoPE scaling
|
||||
# Scaling should only be observed after a long input is fed. We can observe that the frequencies increase
|
||||
# with scaling_factor (or that `inv_freq` decreases)
|
||||
config.rope_scaling = {"type": "dynamic", "factor": scaling_factor}
|
||||
ntk_scaling_rope = PhiRotaryEmbedding(config).to(torch_device)
|
||||
ntk_cos_short, ntk_sin_short = ntk_scaling_rope(x, position_ids_short)
|
||||
ntk_cos_long, ntk_sin_long = ntk_scaling_rope(x, position_ids_long)
|
||||
torch.testing.assert_close(ntk_cos_short, original_cos_short)
|
||||
torch.testing.assert_close(ntk_sin_short, original_sin_short)
|
||||
with self.assertRaises(AssertionError):
|
||||
torch.testing.assert_close(ntk_cos_long, original_cos_long)
|
||||
with self.assertRaises(AssertionError):
|
||||
torch.testing.assert_close(ntk_sin_long, original_sin_long)
|
||||
self.assertTrue((ntk_scaling_rope.inv_freq <= original_rope.inv_freq).all())
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch
|
||||
|
||||
Reference in New Issue
Block a user