EngineX-Cambricon/enginex-mlu370-any2any
9
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

init

Browse Source
This commit is contained in:
luopingyi
2025-10-09 16:47:16 +08:00
parent c8feb4deb5
commit e27e3f16bb
5248 changed files with 1778505 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
transformers/tests/models/aya_vision/__init__.py Normal file
View File

509
transformers/tests/models/aya_vision/test_modeling_aya_vision.py Normal file
View File

@@ -0,0 +1,509 @@
# Copyright 2025 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 GotOcr2 model."""
import unittest
import pytest
from transformers import (
AutoProcessor,
AyaVisionConfig,
is_torch_available,
)
from transformers.testing_utils import (
Expectations,
cleanup,
get_device_properties,
require_deterministic_for_xpu,
require_read_token,
require_torch,
require_torch_accelerator,
slow,
torch_device,
)
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import (
AyaVisionForConditionalGeneration,
AyaVisionModel,
)
class AyaVisionVisionText2TextModelTester:
def __init__(
self,
parent,
batch_size=3,
seq_length=7,
vision_feature_layer=-1,
downsample_factor=2,
ignore_index=-100,
bos_token_id=0,
eos_token_id=0,
pad_token_id=0,
image_token_index=2,
num_channels=3,
image_size=64,
model_type="aya_vision",
is_training=True,
text_config={
"model_type": "cohere2",
"vocab_size": 99,
"hidden_size": 128,
"intermediate_size": 37,
"num_hidden_layers": 2,
"num_attention_heads": 4,
"output_channels": 64,
"hidden_act": "silu",
"max_position_embeddings": 512,
"tie_word_embeddings": True,
"bos_token_id": 0,
"eos_token_id": 0,
"pad_token_id": 0,
},
vision_config={
"model_type": "siglip_vision_model",
"hidden_size": 32,
"num_hidden_layers": 2,
"num_attention_heads": 4,
"intermediate_size": 128,
"image_size": 64,
"patch_size": 8,
"vision_use_head": False,
},
):
self.parent = parent
self.ignore_index = ignore_index
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.image_token_index = image_token_index
self.model_type = model_type
self.text_config = text_config
self.vision_config = vision_config
self.batch_size = batch_size
self.vision_feature_layer = vision_feature_layer
self.downsample_factor = downsample_factor
self.is_training = is_training
self.num_channels = num_channels
self.image_size = image_size
self.image_seq_length = (image_size // (vision_config["patch_size"] * downsample_factor)) ** 2
self.seq_length = seq_length + self.image_seq_length
self.num_hidden_layers = text_config["num_hidden_layers"]
self.vocab_size = text_config["vocab_size"]
self.hidden_size = text_config["hidden_size"]
self.num_attention_heads = text_config["num_attention_heads"]
def get_config(self):
return AyaVisionConfig(
text_config=self.text_config,
vision_config=self.vision_config,
model_type=self.model_type,
bos_token_id=self.bos_token_id,
eos_token_id=self.eos_token_id,
pad_token_id=self.pad_token_id,
image_token_index=self.image_token_index,
vision_feature_layer=self.vision_feature_layer,
downsample_factor=self.downsample_factor,
)
def prepare_config_and_inputs(self):
config = self.get_config()
pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
return config, pixel_values
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
config, pixel_values = config_and_inputs
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
# input_ids[:, -1] = self.pad_token_id
input_ids[input_ids == self.image_token_index] = self.pad_token_id
input_ids[:, : self.image_seq_length] = self.image_token_index
inputs_dict = {
"pixel_values": pixel_values,
"input_ids": input_ids,
"attention_mask": attention_mask,
}
return config, inputs_dict
@require_torch
class AyaVisionModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
AyaVisionModel,
AyaVisionForConditionalGeneration,
)
if is_torch_available()
else ()
)
all_generative_model_classes = (AyaVisionForConditionalGeneration,) if is_torch_available() else ()
pipeline_model_mapping = (
{
"image-text-to-text": AyaVisionForConditionalGeneration,
}
if is_torch_available()
else {}
)
fx_compatible = False
test_pruning = False
test_torchscript = False
test_head_masking = False
_is_composite = True
def setUp(self):
self.model_tester = AyaVisionVisionText2TextModelTester(self)
self.config_tester = ConfigTester(self, config_class=AyaVisionConfig, has_text_modality=False)
def test_config(self):
self.config_tester.run_common_tests()
@unittest.skip(reason="SiglipVisionModel does not support standalone training")
def test_training(self):
pass
@unittest.skip(reason="SiglipVisionModel does not support standalone training")
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(reason="SiglipVisionModel does not support standalone training")
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(reason="SiglipVisionModel does not support standalone training")
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@unittest.skip(reason="Siglip uses a non-standard initialization scheme")
def test_initialization(self):
pass
@unittest.skip(reason="Compile not yet supported because in LLava models")
@pytest.mark.torch_compile_test
def test_sdpa_can_compile_dynamic(self):
pass
# todo: yoni - fix or improve the test
@unittest.skip("Difference is slightly higher than the threshold")
def test_batching_equivalence(self):
pass
@require_read_token
@require_torch
class AyaVisionIntegrationTest(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model_checkpoint = "CohereForAI/aya-vision-8b"
cls.model = None
@classmethod
def tearDownClass(cls):
del cls.model
cleanup(torch_device, gc_collect=True)
def tearDown(self):
cleanup(torch_device, gc_collect=True)
@classmethod
def get_model(cls):
# Use 4-bit on T4
device_type, major, _ = get_device_properties()
load_in_4bit = (device_type == "cuda") and (major < 8)
dtype = None if load_in_4bit else torch.float16
if cls.model is None:
cls.model = AyaVisionForConditionalGeneration.from_pretrained(
cls.model_checkpoint,
device_map=torch_device,
dtype=dtype,
load_in_4bit=load_in_4bit,
)
return cls.model
@slow
@require_torch_accelerator
def test_small_model_integration_forward(self):
processor = AutoProcessor.from_pretrained(self.model_checkpoint)
model = self.get_model()
messages = [
{
"role": "user",
"content": [
{"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
{"type": "text", "text": "Please describe the image explicitly."},
],
}
]
inputs = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
).to(torch_device, dtype=torch.float16)
# Forward
with torch.inference_mode():
output = model(**inputs)
actual_logits = output.logits[0, -1, :5].cpu()
EXPECTED_LOGITS = Expectations(
{
("xpu", 3): [0.4109, 0.1532, 0.8018, 2.1328, 0.5483],
# 4-bit
("cuda", 7): [0.1097, 0.3481, 3.8340, 9.7969, 2.0488],
("cuda", 8): [1.6396, 0.6094, 3.1992, 8.5234, 2.1875],
}
) # fmt: skip
expected_logits = torch.tensor(EXPECTED_LOGITS.get_expectation(), dtype=torch.float16)
self.assertTrue(
torch.allclose(actual_logits, expected_logits, atol=0.1),
f"Actual logits: {actual_logits}"
f"\nExpected logits: {expected_logits}"
f"\nDifference: {torch.abs(actual_logits - expected_logits)}",
)
@slow
@require_torch_accelerator
@require_deterministic_for_xpu
def test_small_model_integration_generate_text_only(self):
processor = AutoProcessor.from_pretrained(self.model_checkpoint)
model = self.get_model()
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": "Write a haiku"},
],
}
]
inputs = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
).to(torch_device, dtype=torch.float16)
with torch.no_grad():
generate_ids = model.generate(**inputs, max_new_tokens=25, do_sample=False)
decoded_output = processor.decode(
generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
)
expected_outputs = Expectations(
{
("xpu", 3): "Whispers on the breeze,\nLeaves dance under moonlit skies,\nNature's quiet song.",
# 4-bit
("cuda", 7): "Sure, here's a haiku for you:\n\nMorning dew sparkles,\nPetals unfold in sunlight,\n",
("cuda", 8): "Whispers on the breeze,\nLeaves dance under moonlit skies,\nNature's quiet song.",
}
) # fmt: skip
expected_output = expected_outputs.get_expectation()
self.assertEqual(decoded_output, expected_output)
@slow
@require_torch_accelerator
@require_deterministic_for_xpu
def test_small_model_integration_generate_chat_template(self):
processor = AutoProcessor.from_pretrained(self.model_checkpoint)
model = self.get_model()
messages = [
{
"role": "user",
"content": [
{"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
{"type": "text", "text": "Please describe the image explicitly."},
],
}
]
inputs = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
).to(torch_device, dtype=torch.float16)
with torch.no_grad():
generate_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
decoded_output = processor.decode(
generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
)
expected_outputs = Expectations(
{
("xpu", 3): 'The image depicts a cozy scene of two cats resting on a bright pink blanket. The cats,',
# 4-bit
("cuda", 7): 'The image depicts two cats comfortably resting on a pink blanket spread across a sofa. The cats,',
("cuda", 8): 'The image depicts a cozy scene of two cats resting on a bright pink blanket. The cats,',
}
) # fmt: skip
expected_output = expected_outputs.get_expectation()
self.assertEqual(decoded_output, expected_output)
@slow
@require_torch_accelerator
def test_small_model_integration_batched_generate(self):
processor = AutoProcessor.from_pretrained(self.model_checkpoint)
model = self.get_model()
# Prepare inputs
messages = [
[
{
"role": "user",
"content": [
{"type": "image", "url": "https://llava-vl.github.io/static/images/view.jpg"},
{"type": "text", "text": "Write a haiku for this image"},
],
},
],
[
{
"role": "user",
"content": [
{
"type": "image",
"url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/australia.jpg",
},
{"type": "text", "text": "Describe this image"},
],
},
],
]
inputs = processor.apply_chat_template(
messages, padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
).to(model.device, dtype=torch.float16)
output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
# Check first output
decoded_output = processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
expected_outputs = Expectations(
{
("xpu", 3): "Wooden path to water,\nMountains echo in stillness,\nPeaceful forest lake.",
# 4-bit
("cuda", 7): "Wooden bridge stretches\nMirrored lake below, mountains rise\nPeaceful, serene",
("cuda", 8): 'Wooden path to water,\nMountains echo in stillness,\nPeaceful forest scene.',
}
) # fmt: skip
expected_output = expected_outputs.get_expectation()
self.assertEqual(
decoded_output,
expected_output,
f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
)
# Check second output
decoded_output = processor.decode(output[1, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
expected_outputs = Expectations(
{
("xpu", 3): 'This image captures a vibrant street scene in a bustling urban area, likely in an Asian city. The focal point is a',
# 4-bit
("cuda", 7): 'This vibrant image captures a bustling street scene in a multicultural urban area, featuring a traditional Chinese gate adorned with intricate red and',
("cuda", 8): 'This image captures a vibrant street scene in a bustling urban area, likely in an Asian city. The focal point is a',
}
) # fmt: skip
expected_output = expected_outputs.get_expectation()
self.assertEqual(
decoded_output,
expected_output,
f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
)
@slow
@require_torch_accelerator
@require_deterministic_for_xpu
def test_small_model_integration_batched_generate_multi_image(self):
processor = AutoProcessor.from_pretrained(self.model_checkpoint)
model = self.get_model()
# Prepare inputs
messages = [
[
{
"role": "user",
"content": [
{"type": "image", "url": "https://llava-vl.github.io/static/images/view.jpg"},
{"type": "text", "text": "Write a haiku for this image"},
],
},
],
[
{
"role": "user",
"content": [
{
"type": "image",
"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
},
{
"type": "image",
"url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg",
},
{
"type": "text",
"text": "These images depict two different landmarks. Can you identify them?",
},
],
},
],
]
inputs = processor.apply_chat_template(
messages, padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
).to(model.device, dtype=torch.float16)
output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
# Check first output
decoded_output = processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
# Batching seems to alter the output slightly, but it is also the case in the original implementation. This seems to be expected: https://github.com/huggingface/transformers/issues/23017#issuecomment-1649630232
expected_outputs = Expectations(
{
("xpu", 3): "Wooden path to water,\nMountains echo in stillness,\nPeaceful forest lake.",
("cuda", 7): 'Wooden bridge stretches\nMirrored lake below, mountains rise\nPeaceful, serene',
("cuda", 8): 'Wooden path to water,\nMountains echo in stillness,\nPeaceful forest scene.',
}
) # fmt: skip
expected_output = expected_outputs.get_expectation()
self.assertEqual(
decoded_output,
expected_output,
f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
)
# Check second output
decoded_output = processor.decode(output[1, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
expected_outputs = Expectations(
{
("xpu", 3): "The first image showcases the Statue of Liberty, a colossal neoclassical sculpture on Liberty Island in New York Harbor. Standing at ",
("cuda", 7): 'The first image showcases the Statue of Liberty, a monumental sculpture located on Liberty Island in New York Harbor. Standing atop a',
("cuda", 8): 'The first image showcases the Statue of Liberty, a colossal neoclassical sculpture on Liberty Island in New York Harbor. Standing at ',
}
) # fmt: skip
expected_output = expected_outputs.get_expectation()
self.assertEqual(
decoded_output,
expected_output,
f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
)

165
transformers/tests/models/aya_vision/test_processing_aya_vision.py Normal file
View File

@@ -0,0 +1,165 @@
# Copyright 2025 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 shutil
import tempfile
import unittest
from transformers import AutoProcessor, AutoTokenizer, AyaVisionProcessor
from transformers.testing_utils import require_torch, require_vision
from transformers.utils import is_torch_available, is_vision_available
from ...test_processing_common import ProcessorTesterMixin, url_to_local_path
if is_torch_available():
import torch
if is_vision_available():
from transformers import GotOcr2ImageProcessor
@require_vision
class AyaVisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):
processor_class = AyaVisionProcessor
@classmethod
def setUpClass(cls):
cls.tmpdirname = tempfile.mkdtemp()
image_processor = GotOcr2ImageProcessor(
do_resize=True,
size={"height": 20, "width": 20},
max_patches=2,
do_rescale=True,
rescale_factor=1 / 255,
do_normalize=True,
image_mean=[0.485, 0.456, 0.406],
image_std=[0.229, 0.224, 0.225],
do_convert_rgb=True,
)
tokenizer = AutoTokenizer.from_pretrained(
"hf-internal-testing/namespace-CohereForAI-repo_name_aya-vision-8b", padding_side="left"
)
processor_kwargs = cls.prepare_processor_dict()
processor = AyaVisionProcessor.from_pretrained(
"hf-internal-testing/namespace-CohereForAI-repo_name_aya-vision-8b",
image_processor=image_processor,
tokenizer=tokenizer,
**processor_kwargs,
)
processor.save_pretrained(cls.tmpdirname)
cls.image_token = processor.image_token
@staticmethod
def prepare_processor_dict():
return {"patch_size": 10, "img_size": 20}
def get_tokenizer(self, **kwargs):
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
def get_image_processor(self, **kwargs):
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
def get_processor(self, **kwargs):
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs)
@classmethod
def tearDownClass(cls):
shutil.rmtree(cls.tmpdirname, ignore_errors=True)
# Copied from tests.models.llava.test_processing_llava.LlavaProcessorTest.test_get_num_vision_tokens
def test_get_num_vision_tokens(self):
"Tests general functionality of the helper used internally in vLLM"
processor = self.get_processor()
output = processor._get_num_multimodal_tokens(image_sizes=[(100, 100), (300, 100), (500, 30)])
self.assertTrue("num_image_tokens" in output)
self.assertEqual(len(output["num_image_tokens"]), 3)
self.assertTrue("num_image_patches" in output)
self.assertEqual(len(output["num_image_patches"]), 3)
@require_torch
def test_process_interleaved_images_videos(self):
processor = self.get_processor()
messages = [
[
{
"role": "user",
"content": [
{
"type": "image",
"url": url_to_local_path(
"https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
),
},
{
"type": "image",
"url": url_to_local_path(
"https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg"
),
},
{"type": "text", "text": "What are the differences between these two images?"},
],
},
],
[
{
"role": "user",
"content": [
{
"type": "image",
"url": url_to_local_path("https://llava-vl.github.io/static/images/view.jpg"),
},
{"type": "text", "text": "Write a haiku for this image"},
],
}
],
]
inputs_batched = processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
padding=True,
)
# Process non batched inputs to check if the pixel_values and input_ids are reconstructed in the correct order when batched together
images_patches_index = 0
for i, message in enumerate(messages):
inputs = processor.apply_chat_template(
message,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
padding=True,
)
# We slice with [-inputs["input_ids"].shape[1] :] as the input_ids are left padded
torch.testing.assert_close(
inputs["input_ids"][0], inputs_batched["input_ids"][i][-inputs["input_ids"].shape[1] :]
)
torch.testing.assert_close(
inputs["pixel_values"],
inputs_batched["pixel_values"][
images_patches_index : images_patches_index + inputs["pixel_values"].shape[0]
],
)
images_patches_index += inputs["pixel_values"].shape[0]