[Model] Support E5-V (#9576)

This commit is contained in:
Cyrus Leung
2024-10-23 11:35:29 +08:00
committed by GitHub
parent 29061ed9df
commit 831540cf04
12 changed files with 528 additions and 86 deletions
+36 -24
View File
@@ -43,10 +43,12 @@ _TEST_DIR = os.path.dirname(__file__)
_TEST_PROMPTS = [os.path.join(_TEST_DIR, "prompts", "example.txt")]
_LONG_PROMPTS = [os.path.join(_TEST_DIR, "prompts", "summary.txt")]
PromptImageInput = Union[List[Image.Image], List[List[Image.Image]]]
PromptAudioInput = Union[List[Tuple[np.ndarray, int]],
List[List[Tuple[np.ndarray, int]]]]
PromptVideoInput = Union[List[np.ndarray], List[List[np.ndarray]]]
_M = TypeVar("_M")
_PromptMultiModalInput = Union[List[_M], List[List[_M]]]
PromptImageInput = _PromptMultiModalInput[Image.Image]
PromptAudioInput = _PromptMultiModalInput[Tuple[np.ndarray, int]]
PromptVideoInput = _PromptMultiModalInput[np.ndarray]
def _read_prompts(filename: str) -> List[str]:
@@ -318,12 +320,12 @@ class HfRunner:
"text": prompt,
"return_tensors": "pt",
}
if images is not None and images[i] is not None:
processor_kwargs["images"] = images[i]
if videos is not None and videos[i] is not None:
processor_kwargs["videos"] = videos[i]
if audios is not None and audios[i] is not None:
audio, sr = audios[i]
if images is not None and (image := images[i]) is not None:
processor_kwargs["images"] = image
if videos is not None and (video := videos[i]) is not None:
processor_kwargs["videos"] = video
if audios is not None and (audio_tuple := audios[i]) is not None:
audio, sr = audio_tuple
processor_kwargs["audio"] = audio
processor_kwargs["sampling_rate"] = sr
@@ -338,7 +340,7 @@ class HfRunner:
self,
prompts: List[str],
images: Optional[PromptImageInput] = None,
videos: Optional[List[np.ndarray]] = None,
videos: Optional[PromptVideoInput] = None,
audios: Optional[PromptAudioInput] = None,
**kwargs: Any,
) -> List[Tuple[List[List[int]], List[str]]]:
@@ -368,7 +370,7 @@ class HfRunner:
prompts: List[str],
max_tokens: int,
images: Optional[PromptImageInput] = None,
videos: Optional[List[np.ndarray]] = None,
videos: Optional[PromptVideoInput] = None,
audios: Optional[PromptAudioInput] = None,
**kwargs: Any,
) -> List[Tuple[List[int], str]]:
@@ -409,7 +411,7 @@ class HfRunner:
prompts: List[str],
max_tokens: int,
images: Optional[PromptImageInput] = None,
videos: Optional[List[np.ndarray]] = None,
videos: Optional[PromptVideoInput] = None,
audios: Optional[PromptAudioInput] = None,
**kwargs: Any,
) -> List[List[torch.Tensor]]:
@@ -488,7 +490,7 @@ class HfRunner:
num_logprobs: int,
images: Optional[PromptImageInput] = None,
audios: Optional[PromptAudioInput] = None,
videos: Optional[List[np.ndarray]] = None,
videos: Optional[PromptVideoInput] = None,
**kwargs: Any,
) -> List[TokensTextLogprobs]:
all_inputs = self.get_inputs(prompts,
@@ -657,15 +659,18 @@ class VllmRunner:
inputs = [TextPrompt(prompt=prompt) for prompt in prompts]
if images is not None:
for i, image in enumerate(images):
inputs[i]["multi_modal_data"] = {"image": image}
if image is not None:
inputs[i]["multi_modal_data"] = {"image": image}
if videos is not None:
for i, video in enumerate(videos):
inputs[i]["multi_modal_data"] = {"video": video}
if video is not None:
inputs[i]["multi_modal_data"] = {"video": video}
if audios is not None:
for i, audio in enumerate(audios):
inputs[i]["multi_modal_data"] = {"audio": audio}
if audio is not None:
inputs[i]["multi_modal_data"] = {"audio": audio}
return inputs
@@ -837,13 +842,20 @@ class VllmRunner:
returned_outputs.append((token_ids, texts))
return returned_outputs
def encode(self, prompts: List[str]) -> List[List[float]]:
req_outputs = self.model.encode(prompts)
outputs = []
for req_output in req_outputs:
embedding = req_output.outputs.embedding
outputs.append(embedding)
return outputs
def encode(
self,
prompts: List[str],
images: Optional[PromptImageInput] = None,
videos: Optional[PromptVideoInput] = None,
audios: Optional[PromptAudioInput] = None,
) -> List[List[float]]:
inputs = self.get_inputs(prompts,
images=images,
videos=videos,
audios=audios)
req_outputs = self.model.encode(inputs)
return [req_output.outputs.embedding for req_output in req_outputs]
def __enter__(self):
return self
+2 -1
View File
@@ -16,7 +16,8 @@ def check_embeddings_close(
for prompt_idx, (embeddings_0, embeddings_1) in enumerate(
zip(embeddings_0_lst, embeddings_1_lst)):
assert len(embeddings_0) == len(embeddings_1)
assert len(embeddings_0) == len(embeddings_1), (
f"Length mismatch: {len(embeddings_0)} vs. {len(embeddings_1)}")
sim = F.cosine_similarity(torch.tensor(embeddings_0),
torch.tensor(embeddings_1),
@@ -0,0 +1,135 @@
from typing import List, Type
import pytest
import torch.nn.functional as F
from transformers import AutoModelForVision2Seq
from ....conftest import IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner
from ....utils import large_gpu_test
from ..utils import check_embeddings_close
llama3_template = '<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n \n' # noqa: E501
HF_TEXT_PROMPTS = [
# T -> X
llama3_template.format(
"The label of the object is stop sign\nSummary above sentence in one word: " # noqa: E501
),
# T -> X
llama3_template.format(
"cherry blossom\nSummary above sentence in one word: "),
]
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
# I -> X
"stop_sign":
llama3_template.format("<image>\nSummary above image in one word: "),
# I -> X
"cherry_blossom":
llama3_template.format("<image>\nSummary above image in one word: "),
})
MODELS = ["royokong/e5-v"]
def _run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
input_texts: List[str],
input_images: PromptImageInput,
model: str,
*,
dtype: str,
) -> None:
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(model,
task="embedding",
dtype=dtype,
max_model_len=4096,
enforce_eager=True) as vllm_model:
vllm_outputs = vllm_model.encode(input_texts, images=input_images)
with hf_runner(model, dtype=dtype,
auto_cls=AutoModelForVision2Seq) as hf_model:
# Patch the issue where image_token_id
# exceeds the maximum allowed vocab size
hf_model.model.resize_token_embeddings(
hf_model.model.language_model.vocab_size + 1)
all_inputs = hf_model.get_inputs(input_texts, images=input_images)
all_outputs = []
for inputs in all_inputs:
# Based on: https://huggingface.co/royokong/e5-v
outputs = hf_model.model(
**hf_model.wrap_device(inputs,
device=hf_model.model.device.type),
return_dict=True,
output_hidden_states=True,
)
pooled_output = F.normalize(outputs.hidden_states[-1][0, -1, :],
dim=-1)
all_outputs.append(pooled_output.tolist())
hf_outputs = all_outputs
check_embeddings_close(
embeddings_0_lst=hf_outputs,
embeddings_1_lst=vllm_outputs,
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
def test_models_text(
hf_runner,
vllm_runner,
image_assets,
model: str,
dtype: str,
) -> None:
input_texts_images = [(text, None) for text in HF_TEXT_PROMPTS]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]
_run_test(
hf_runner,
vllm_runner,
input_texts,
input_images, # type: ignore
model,
dtype=dtype,
)
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
def test_models_image(
hf_runner,
vllm_runner,
image_assets,
model: str,
dtype: str,
) -> None:
input_texts_images = [
(text, asset.pil_image)
for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]
_run_test(
hf_runner,
vllm_runner,
input_texts,
input_images,
model,
dtype=dtype,
)
@@ -1,42 +1,53 @@
from typing import List, Type
import pytest
import torch.nn.functional as F
from ....conftest import IMAGE_ASSETS
from ....conftest import IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner
from ....utils import large_gpu_test
from ..utils import check_embeddings_close
HF_TEXT_PROMPTS = [
# T -> X
"Find me an everyday image that matches the given caption: The label of the object is stop sign", # noqa: E501
# T -> X
"Retrieve an image of this caption: cherry blossom",
]
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
# T + I -> X
"stop_sign":
"<|image_1|> Select the portion of the image that isolates the object of the given label: The label of the object is stop sign", # noqa: E501
# I -> X
"cherry_blossom":
"<|image_1|> Represent the given image with the following question: What is in the image", # noqa: E501
"<|image_1|> Represent the given image for classification", # noqa: E501
})
MODELS = ["TIGER-Lab/VLM2Vec-Full"]
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
def test_models(
hf_runner,
vllm_runner,
example_prompts,
def _run_test(
hf_runner: Type[HfRunner],
vllm_runner: Type[VllmRunner],
input_texts: List[str],
input_images: PromptImageInput,
model: str,
*,
dtype: str,
) -> None:
# NOTE: take care of the order. run vLLM first, and then run HF.
# vLLM needs a fresh new process without cuda initialization.
# if we run HF first, the cuda initialization will be done and it
# will hurt multiprocessing backend with fork method (the default method).
with vllm_runner(model,
task="embedding",
max_model_len=4096,
max_num_seqs=2,
dtype=dtype,
with vllm_runner(model, task="embedding", dtype=dtype,
enforce_eager=True) as vllm_model:
vllm_outputs = vllm_model.encode(example_prompts)
vllm_outputs = vllm_model.encode(input_texts, images=input_images)
with hf_runner(model, dtype=dtype) as hf_model:
all_inputs = hf_model.get_inputs(example_prompts)
# use eager mode for hf runner, since phi3_v didn't work with flash_attn
hf_model_kwargs = {"_attn_implementation": "eager"}
with hf_runner(model, dtype=dtype,
model_kwargs=hf_model_kwargs) as hf_model:
all_inputs = hf_model.get_inputs(input_texts, images=input_images)
all_outputs = []
for inputs in all_inputs:
@@ -61,3 +72,53 @@ def test_models(
name_0="hf",
name_1="vllm",
)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
def test_models_text(
hf_runner,
vllm_runner,
image_assets,
model: str,
dtype: str,
) -> None:
input_texts_images = [(text, None) for text in HF_TEXT_PROMPTS]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]
_run_test(
hf_runner,
vllm_runner,
input_texts,
input_images, # type: ignore
model,
dtype=dtype,
)
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
def test_models_image(
hf_runner,
vllm_runner,
image_assets,
model: str,
dtype: str,
) -> None:
input_texts_images = [
(text, asset.pil_image)
for text, asset in zip(HF_IMAGE_PROMPTS, image_assets)
]
input_texts = [text for text, _ in input_texts_images]
input_images = [image for _, image in input_texts_images]
_run_test(
hf_runner,
vllm_runner,
input_texts,
input_images,
model,
dtype=dtype,
)