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

Chameleon: add model (#31534)

Browse Source 
* Chameleon model integration

Co-authored-by: Jacob Kahn <jacobkahn1@gmail.com>
Co-authored-by: Leonid Shamis <leonid.shamis@gmail.com>

* fix 7B, again. mask away image tokens

* Apply suggestions from code review

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* remove pretrained_config_map

* make fixup passing up to utils/check_config_docstrings.py; vqgan moved to the modeling file

* remove tokenizer (use llama's); remove codechameleon tests

* a few copied from statements and minor changes

* copied from in ChameleonModel

* some copies in ChameleonForCausalLM

* a few more copies

* VQModel moved to ChameleonModel (as opposed to being in the processor)

* ChameleonProcessor ready

* Fix chameleon weights convert

* update conversion script

* clean-up processing

* update modeling a bit

* update

* update (throws error...)

* correct conversion ready

* fix tests

* fix docs

* docs

* ve swin norm

* fix device for vocab map

* add normalization

* update

* update script with rope rotations

* final fix on model conversion

* add slow tests

* more info in docs

* fix repo consistency tests

* fix repo tests

* fix-copies

* hope this will make CI happy

* fix for 30b model

* Update docs/source/en/index.md

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update docs/source/en/model_doc/chameleon.md

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/chameleon/modeling_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update docs/source/en/model_doc/chameleon.md

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update docs/source/en/model_doc/chameleon.md

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update docs/source/en/model_doc/chameleon.md

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update docs/source/en/model_doc/chameleon.md

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/auto/configuration_auto.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/chameleon/image_processing_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/chameleon/image_processing_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/chameleon/image_processing_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/chameleon/image_processing_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/chameleon/modeling_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/chameleon/processing_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/chameleon/processing_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/chameleon/test_modeling_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/chameleon/test_modeling_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/chameleon/test_modeling_chameleon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* address comments

* remove assertion in conversion script

* add image processor test

* not copied

* port changes for qk layernorm

* fix-copies

* read token decorator for tests

* [run-slow] chameleon

* one more read-token

* address some comments

* qk norm changes

* tests and repo check

* moved rope permutations to conversion, YAY!

* fix past kv check

* docs

* layernorm done!

* let's be consistent in naming

* fix slow tests

* weird thing with slow CI, but let's see

* once more try

* remove past-kv as tuple following llama

* ignore

* style

---------

Co-authored-by: Pablo Montalvo <39954772+molbap@users.noreply.github.com>
Co-authored-by: ArthurZucker <arthur.zucker@gmail.com>
Co-authored-by: jacobkahn <jacobkahn1@gmail.com>
Co-authored-by: Leonid Shamis <leonid.shamis@gmail.com>
Co-authored-by: Leonid Shamis <lshamis@meta.com>
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Joao Gante <joao@huggingface.co>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
This commit is contained in:
Raushan Turganbay
2024-07-17 10:41:43 +05:00
committed by GitHub
parent 4037a2b5b1
commit 24cfcc2114
24 changed files with 3950 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
tests/models/chameleon/__init__.py Normal file
View File

205
tests/models/chameleon/test_image_processing_chameleon.py Normal file
View File

@@ -0,0 +1,205 @@
# coding=utf-8
# Copyright 2024 HuggingFace Inc.
#
# 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.testing_utils import require_torch, require_vision
from transformers.utils import is_torch_available, is_vision_available
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
if is_torch_available():
import torch
if is_vision_available():
from PIL import Image
from transformers import ChameleonImageProcessor
class ChameleonImageProcessingTester(unittest.TestCase):
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
image_size=18,
min_resolution=30,
max_resolution=200,
do_resize=True,
size=None,
do_center_crop=True,
crop_size=None,
do_normalize=True,
image_mean=[1.0, 1.0, 1.0],
image_std=[1.0, 1.0, 1.0],
do_convert_rgb=True,
):
size = size if size is not None else {"shortest_edge": 18}
crop_size = crop_size if crop_size is not None else {"height": 18, "width": 18}
self.parent = parent
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.min_resolution = min_resolution
self.max_resolution = max_resolution
self.do_resize = do_resize
self.size = size
self.do_center_crop = do_center_crop
self.crop_size = crop_size
self.do_normalize = do_normalize
self.image_mean = image_mean
self.image_std = image_std
self.do_convert_rgb = do_convert_rgb
def prepare_image_processor_dict(self):
return {
"do_resize": self.do_resize,
"size": self.size,
"do_center_crop": self.do_center_crop,
"crop_size": self.crop_size,
"do_normalize": self.do_normalize,
"image_mean": self.image_mean,
"image_std": self.image_std,
"do_convert_rgb": self.do_convert_rgb,
}
# Copied from tests.models.clip.test_image_processing_clip.CLIPImageProcessingTester.expected_output_image_shape
def expected_output_image_shape(self, images):
return self.num_channels, self.crop_size["height"], self.crop_size["width"]
# Copied from tests.models.clip.test_image_processing_clip.CLIPImageProcessingTester.prepare_image_inputs
def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
return prepare_image_inputs(
batch_size=self.batch_size,
num_channels=self.num_channels,
min_resolution=self.min_resolution,
max_resolution=self.max_resolution,
equal_resolution=equal_resolution,
numpify=numpify,
torchify=torchify,
)
@require_torch
@require_vision
class ChameleonImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = ChameleonImageProcessor if is_vision_available() else None
# Copied from tests.models.clip.test_image_processing_clip.CLIPImageProcessingTest.setUp with CLIP->Chameleon
def setUp(self):
super().setUp()
self.image_processor_tester = ChameleonImageProcessingTester(self)
@property
# Copied from tests.models.clip.test_image_processing_clip.CLIPImageProcessingTest.image_processor_dict
def image_processor_dict(self):
return self.image_processor_tester.prepare_image_processor_dict()
def test_image_processor_properties(self):
image_processing = self.image_processing_class(**self.image_processor_dict)
self.assertTrue(hasattr(image_processing, "do_resize"))
self.assertTrue(hasattr(image_processing, "size"))
self.assertTrue(hasattr(image_processing, "do_center_crop"))
self.assertTrue(hasattr(image_processing, "center_crop"))
self.assertTrue(hasattr(image_processing, "do_normalize"))
self.assertTrue(hasattr(image_processing, "image_mean"))
self.assertTrue(hasattr(image_processing, "image_std"))
self.assertTrue(hasattr(image_processing, "do_convert_rgb"))
def test_image_processor_from_dict_with_kwargs(self):
image_processor = self.image_processing_class.from_dict(self.image_processor_dict)
self.assertEqual(image_processor.size, {"shortest_edge": 18})
self.assertEqual(image_processor.crop_size, {"height": 18, "width": 18})
image_processor = self.image_processing_class.from_dict(self.image_processor_dict, size=42, crop_size=84)
self.assertEqual(image_processor.size, {"shortest_edge": 42})
self.assertEqual(image_processor.crop_size, {"height": 84, "width": 84})
def test_call_pil(self):
# Initialize image_processing
image_processing = self.image_processing_class(**self.image_processor_dict)
# create random PIL images
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True)
for image in image_inputs:
self.assertIsInstance(image, Image.Image)
# Test not batched input
encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
expected_output_image_shape = (1, 3, 18, 18)
self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
# Test batched
encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
expected_output_image_shape = (7, 3, 18, 18)
self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
def test_call_numpy(self):
# Initialize image_processing
image_processing = self.image_processing_class(**self.image_processor_dict)
# create random numpy tensors
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, numpify=True)
for image in image_inputs:
self.assertIsInstance(image, np.ndarray)
# Test not batched input
encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
expected_output_image_shape = (1, 3, 18, 18)
self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
# Test batched
encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
expected_output_image_shape = (7, 3, 18, 18)
self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
def test_call_pytorch(self):
# Initialize image_processing
image_processing = self.image_processing_class(**self.image_processor_dict)
# create random PyTorch tensors
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, torchify=True)
for image in image_inputs:
self.assertIsInstance(image, torch.Tensor)
# Test not batched input
encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
expected_output_image_shape = (1, 3, 18, 18)
self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
# Test batched
encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
expected_output_image_shape = (7, 3, 18, 18)
self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
def test_nested_input(self):
image_processing = self.image_processing_class(**self.image_processor_dict)
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True)
# Test batched as a list of images
encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
expected_output_image_shape = (7, 3, 18, 18)
self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape)
# Test batched as a nested list of images, where each sublist is one batch
image_inputs_nested = [image_inputs[:3], image_inputs[3:]]
encoded_images_nested = image_processing(image_inputs_nested, return_tensors="pt").pixel_values
expected_output_image_shape = (7, 3, 18, 18)
self.assertEqual(tuple(encoded_images_nested.shape), expected_output_image_shape)
# Image processor should return same pixel values, independently of input format
self.assertTrue((encoded_images_nested == encoded_images).all())

448
tests/models/chameleon/test_modeling_chameleon.py Normal file
View File

@@ -0,0 +1,448 @@
# 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 chameleon model."""
import unittest
import pytest
import requests
from parameterized import parameterized
from transformers import ChameleonConfig, is_torch_available, is_vision_available, set_seed
from transformers.testing_utils import (
require_bitsandbytes,
require_flash_attn,
require_read_token,
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_vision_available():
from PIL import Image
if is_torch_available():
import torch
from transformers import (
ChameleonForCausalLM,
ChameleonModel,
ChameleonProcessor,
)
class ChameleonModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=False,
use_input_mask=True,
use_labels=True,
vocab_size=99,
image_token_id=98,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=2,
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,
vq_num_embeds=12,
vq_embed_dim=12,
vq_channel_multiplier=[1, 2],
vq_img_token_start_id=10, # has to be less than vocab size when added with vq_num_embeds
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.image_token_id = image_token_id
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.vq_num_embeds = vq_num_embeds
self.vq_embed_dim = vq_embed_dim
self.vq_channel_multiplier = vq_channel_multiplier
self.vq_img_token_start_id = vq_img_token_start_id
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)
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, input_mask, sequence_labels, token_labels, choice_labels
def get_config(self):
# create dummy vocab map for image2bpe mapping if it needs remapping
# we assume that vocab size is big enough to accoun for image tokens somewhere in the beginning
# same way as in real ckpt, when img tokens are in first half of embeds
# we will need "vq_num_embeds" amount of tokens
vocab_map = {i: chr(i) for i in range(self.vocab_size)}
vocab_map[self.image_token_id] = "<image>"
start = self.vq_img_token_start_id
end = self.vq_img_token_start_id + self.vq_num_embeds
for i in range(start, end):
vocab_map[i] = f"IMGIMGBS{i}" # dummy str for each token, anything starting with IMGIMG
return ChameleonConfig(
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,
vocabulary_map={v: k for k, v in vocab_map.items()},
vq_config=self.get_vq_config(),
)
def get_vq_config(self):
return {
"embed_dim": self.vq_embed_dim,
"num_embeddings": self.vq_num_embeds,
"latent_channels": self.vq_embed_dim,
"in_channels": 3,
"base_channels": 32, # we have a GroupNorm of 32 groups, so can't do less
"channel_multiplier": self.vq_channel_multiplier,
}
def create_and_check_model(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
model = ChameleonModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = ChameleonForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
model = ChameleonForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_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 ChameleonModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (ChameleonModel, ChameleonForCausalLM) if is_torch_available() else ()
all_generative_model_classes = (ChameleonForCausalLM,) if is_torch_available() else ()
pipeline_model_mapping = (
{
"feature-extraction": ChameleonModel,
"text-generation": ChameleonForCausalLM,
}
if is_torch_available()
else {}
)
test_headmasking = False
test_pruning = False
fx_compatible = False
def setUp(self):
self.model_tester = ChameleonModelTester(self)
self.config_tester = ConfigTester(self, config_class=ChameleonConfig, 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)
@parameterized.expand([("linear",), ("dynamic",)])
def test_model_rope_scaling(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 = ChameleonModel(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 = ChameleonModel(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":
self.assertTrue(torch.allclose(original_short_output, scaled_short_output, 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))
@require_flash_attn
@require_read_token
@require_torch_gpu
@require_bitsandbytes
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_generate_padding_right(self):
"""
Overwritting the common test as the test is flaky on tiny models
"""
model = ChameleonForCausalLM.from_pretrained(
"facebook/chameleon-7b",
load_in_4bit=True,
device_map={"": 0},
)
processor = ChameleonProcessor.from_pretrained("facebook/chameleon-7b")
texts = ["hi", "Hello this is a very long sentence"]
processor.tokenizer.padding_side = "right"
inputs = processor(texts, return_tensors="pt", padding=True).to(0)
output_native = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_native = processor.tokenizer.batch_decode(output_native)
model = ChameleonForCausalLM.from_pretrained(
"facebook/chameleon-7b",
load_in_4bit=True,
attn_implementation="flash_attention_2",
)
output_fa_2 = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_fa_2 = processor.tokenizer.batch_decode(output_fa_2)
self.assertListEqual(output_native, output_fa_2)
@unittest.skip("Chameleon forces some token ids to be -inf!")
def test_batching_equivalence(self):
pass
@require_torch
class ChameleonIntegrationTest(unittest.TestCase):
@slow
@require_bitsandbytes
@require_read_token
def test_model_7b(self):
model = ChameleonForCausalLM.from_pretrained("facebook/chameleon-7b", load_in_4bit=True, device_map="auto")
processor = ChameleonProcessor.from_pretrained("facebook/chameleon-7b")
image = Image.open(
requests.get("https://nineplanets.org/wp-content/uploads/2020/12/the-big-dipper-1.jpg", stream=True).raw
)
prompt = "<image>Describe what do you see here and tell me about the history behind it?"
inputs = processor(prompt, images=image, return_tensors="pt").to(model.device, torch.float16)
# greedy generation outputs
EXPECTED_TEXT_COMPLETION = ['Describe what do you see here and tell me about the history behind it?The image depicts a star map, with a bright blue line extending across the center of the image. The line is labeled "390 light years" and is accompanied by a small black and'] # fmt: skip
generated_ids = model.generate(**inputs, max_new_tokens=40, do_sample=False)
text = processor.batch_decode(generated_ids, skip_special_tokens=True)
self.assertEqual(EXPECTED_TEXT_COMPLETION, text)
@slow
@require_bitsandbytes
@require_read_token
def test_model_7b_batched(self):
model = ChameleonForCausalLM.from_pretrained("facebook/chameleon-7b", load_in_4bit=True, device_map="auto")
processor = ChameleonProcessor.from_pretrained("facebook/chameleon-7b")
image = Image.open(
requests.get("https://nineplanets.org/wp-content/uploads/2020/12/the-big-dipper-1.jpg", stream=True).raw
)
image_2 = Image.open(
requests.get("https://www.kxan.com/wp-content/uploads/sites/40/2020/10/ORION.jpg", stream=True).raw
)
prompts = [
"<image>Describe what do you see here and tell me about the history behind it?",
"What constellation is this image showing?<image>",
]
inputs = processor(prompts, images=[image, image_2], padding=True, return_tensors="pt").to(
model.device, torch.float16
)
# greedy generation outputs
EXPECTED_TEXT_COMPLETION = [
'Describe what do you see here and tell me about the history behind it?The image depicts a star map, with a bright blue dot in the center representing the star Alpha Centauri. The star map is a representation of the night sky, showing the positions of stars in',
'What constellation is this image showing?The image is showing the constellation of Orion.'
] # fmt: skip
generated_ids = model.generate(**inputs, max_new_tokens=40, do_sample=False)
text = processor.batch_decode(generated_ids, skip_special_tokens=True)
self.assertEqual(EXPECTED_TEXT_COMPLETION, text)
@slow
@require_bitsandbytes
@require_read_token
def test_model_7b_multi_image(self):
model = ChameleonForCausalLM.from_pretrained("facebook/chameleon-7b", load_in_4bit=True, device_map="auto")
processor = ChameleonProcessor.from_pretrained("facebook/chameleon-7b")
image = Image.open(
requests.get("https://nineplanets.org/wp-content/uploads/2020/12/the-big-dipper-1.jpg", stream=True).raw
)
image_2 = Image.open(
requests.get("https://www.kxan.com/wp-content/uploads/sites/40/2020/10/ORION.jpg", stream=True).raw
)
prompt = "What do these two images have in common?<image><image>"
inputs = processor(prompt, images=[image, image_2], return_tensors="pt").to(model.device, torch.float16)
# greedy generation outputs
EXPECTED_TEXT_COMPLETION = ['What do these two images have in common?The two images show a connection between two things that are not necessarily related. The first image shows a group of stars, while the second image shows a network of lines connecting two points. The connection between'] # fmt: skip
generated_ids = model.generate(**inputs, max_new_tokens=40, do_sample=False)
text = processor.batch_decode(generated_ids, skip_special_tokens=True)
self.assertEqual(EXPECTED_TEXT_COMPLETION, text)

4
tests/test_modeling_common.py
View File

@@ -259,9 +259,11 @@ class ModelTesterMixin:
# make sure we don't have nans
out_2 = out2.cpu().numpy()
out_2[np.isnan(out_2)] = 0
out_2 = out_2[~np.isneginf(out_2)]
out_1 = out1.cpu().numpy()
out_1[np.isnan(out_1)] = 0
out_1 = out_1[~np.isneginf(out_1)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
@@ -660,6 +662,8 @@ class ModelTesterMixin:
out_2 = second.cpu().numpy()
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
out_1 = out_1[~np.isneginf(out_1)]
out_2 = out_2[~np.isneginf(out_2)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)