运行multimodal_understanding.py报错，只改了模型从魔搭社区下载那一部分

[zhrli]

Exception has occurred: ValueError
Unable to create tensor, you should probably activate padding with 'padding=True' to have batched tensors with the same length.
RuntimeError: Could not infer dtype of numpy.float32

During handling of the above exception, another exception occurred:

File "/home/lizhaorui/.cache/huggingface/modules/transformers_modules/Emu3-VisionTokenizer/image_processing_emu3visionvq.py", line 349, in preprocess
return BatchFeature(data=data, tensor_type=return_tensors)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/lizhaorui/DL/Emu3/emu3/mllm/processing_emu3.py", line 274, in tokenize_image
image_inputs = self.image_processor(image, return_tensors="pt")["pixel_values"]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/lizhaorui/DL/Emu3/emu3/mllm/processing_emu3.py", line 159, in call
image_tokens = self.tokenize_image(image, padding_image=padding_image)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/lizhaorui/DL/Emu3/multimodal_understanding.py", line 35, in
inputs = processor(
^^^^^^^^^^
ValueError: Unable to create tensor, you should probably activate padding with 'padding=True' to have batched tensors with the same length.

图片的例子是项目里的例子

[ryanzhangfan]

拉一下最新的代码和模型呢？我check了github，modelscope里所有最新代码，processing_emu3.py line 159都不是一行有效代码。

[zhrli]

拉一下最新的代码和模型呢？我check了github，modelscope里所有最新代码，processing_emu3.py line 159都不是一行有效代码。

确实能跑起来了，爆内存了。 项目能不能考虑分块儿分卡执行，单卡执行需要的资源太多了。Emu3确实是我们行业里可以依赖的唯一模型，感谢智源研究院。

[ryanzhangfan]

模型完全兼容transformers中的各种优化方法，可以直接使用transformers或者accelerate支持的自动化分卡（仅限多模态理解模型），代码可以参考Emu2 demo code，或者使用transformers自带的int4量化。如果只是kv cache爆了也可以尝试transformers库支持的offload kvcache的方式。

[zhrli]

拉一下最新的代码和模型呢？我check了github，modelscope里所有最新代码，processing_emu3.py line 159都不是一行有效代码。

Traceback (most recent call last):
File "/home/lizhaorui/anaconda3/envs/agent/lib/python3.12/site-packages/transformers/feature_extraction_utils.py", line 186, in convert_to_tensors
tensor = as_tensor(value)
^^^^^^^^^^^^^^^^
File "/home/lizhaorui/anaconda3/envs/agent/lib/python3.12/site-packages/transformers/feature_extraction_utils.py", line 142, in as_tensor
return torch.tensor(value)
^^^^^^^^^^^^^^^^^^^
RuntimeError: Could not infer dtype of numpy.float32

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "/home/lizhaorui/DL/Emu3/multimodal_understanding.py", line 34, in
inputs = processor(
^^^^^^^^^^
File "/home/lizhaorui/anaconda3/envs/agent/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/home/lizhaorui/DL/Emu3/emu3/mllm/processing_emu3.py", line 156, in call
image_tokens = self.tokenize_image(image, padding_image=padding_image)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/lizhaorui/DL/Emu3/emu3/mllm/processing_emu3.py", line 271, in tokenize_image
image_inputs = self.image_processor(image, return_tensors="pt")["pixel_values"]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/lizhaorui/anaconda3/envs/agent/lib/python3.12/site-packages/transformers/image_processing_utils.py", line 41, in call
return self.preprocess(images, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/lizhaorui/.cache/huggingface/modules/transformers_modules/Emu3-VisionTokenizer/image_processing_emu3visionvq.py", line 349, in preprocess
return BatchFeature(data=data, tensor_type=return_tensors)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/lizhaorui/anaconda3/envs/agent/lib/python3.12/site-packages/transformers/feature_extraction_utils.py", line 79, in init
self.convert_to_tensors(tensor_type=tensor_type)
File "/home/lizhaorui/anaconda3/envs/agent/lib/python3.12/site-packages/transformers/feature_extraction_utils.py", line 192, in convert_to_tensors
raise ValueError(
ValueError: Unable to create tensor, you should probably activate padding with 'padding=True' to have batched tensors with the same length.

在运行multimodal_understanding.py时候报错仍然存在，新拉了代码

[ryanzhangfan]

可以check下numpy和torch的版本，目测这是np.array转torch.tensor的时候报的错误。

[zhrli]

可以check下numpy和torch的版本，目测这是np.array转torch.tensor的时候报的错误。

Name: torch
Version: 2.2.1
Summary: Tensors and Dynamic neural networks in Python with strong GPU acceleration
Home-page: https://pytorch.org/
Author: PyTorch Team
Author-email: [email protected]
License: BSD-3
Location: /home/lizhaorui/anaconda3/envs/agent/lib/python3.12/site-packages
Requires: filelock, fsspec, jinja2, networkx, nvidia-cublas-cu12, nvidia-cuda-cupti-cu12, nvidia-cuda-nvrtc-cu12, nvidia-cuda-runtime-cu12, nvidia-cudnn-cu12, nvidia-cufft-cu12, nvidia-curand-cu12, nvidia-cusolver-cu12, nvidia-cusparse-cu12, nvidia-nccl-cu12, nvidia-nvtx-cu12, sympy, typing-extensions
Required-by: accelerate, bitsandbytes, flash_attn, torchaudio, torchvision

Name: numpy
Version: 1.26.4

[ryanzhangfan]

可以尝试换下numpy版本试试？看着是numpy转tensor报错，但是识别到的numpy.dtype也没啥问题。。我们的环境同样的版本torch 2.2.1, numpy 1.26.4, transformers 4.44.0是能正常运行的。如果换版本还是不行可以在/home/lizhaorui/.cache/huggingface/modules/transformers_modules/Emu3-VisionTokenizer/image_processing_emu3visionvq.py line 349前打印确认下pixel_values的dtype和shape.

[zhrli]

transformers自带的int4量化

print(pixel_values.shape)
(2, 3, 512, 512)

print(pixel_values.dtype)
float32

[ryanzhangfan]

确认下环境问题吧，仅从目前提供的信息看，看起来不太像是我们代码的问题，而是numpy.array转torch.tensor报错了。

[zhrli]

确认下环境问题吧，仅从目前提供的信息看，看起来不太像是我们代码的问题，而是numpy.array转torch.tensor报错了。
搞定了
pytorch 2.2.1 配合numpy1.25.2

[zhrli]

模型完全兼容transformers中的各种优化方法，可以直接使用transformers或者accelerate支持的自动化分卡（仅限多模态理解模型），代码可以参考Emu2 demo code，或者使用transformers自带的int4量化。如果只是kv cache爆了也可以尝试transformers库支持的offload kvcache的方式。

双卡4090跑成功了

-- coding: utf-8 --

from PIL import Image
from transformers import AutoTokenizer, AutoModel, AutoImageProcessor, AutoModelForCausalLM
from transformers import BitsAndBytesConfig # 导入量化配置
from transformers.generation.configuration_utils import GenerationConfig
import torch

from emu3.mllm.processing_emu3 import Emu3Processor

from modelscope import snapshot_download

model path

EMU_HUB = snapshot_download("BAAI/Emu3-Chat")
VQ_HUB = snapshot_download("BAAI/Emu3-VisionTokenizer")

Quantization configuration

quantization_config = BitsAndBytesConfig(
load_in_4bit=True, # 使用 int4 量化
bnb_4bit_quant_type='nf4', # 量化类型
bnb_4bit_compute_dtype=torch.bfloat16, # 计算精度
)

prepare model and processor

model = AutoModelForCausalLM.from_pretrained(
EMU_HUB,
quantization_config=quantization_config, # 使用量化配置
device_map="auto", # 自动分配到所有可用的GPU上
trust_remote_code=True,
)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(EMU_HUB, trust_remote_code=True, padding_side="left")
image_processor = AutoImageProcessor.from_pretrained(VQ_HUB, trust_remote_code=True)
image_tokenizer = AutoModel.from_pretrained(VQ_HUB, device_map="auto", trust_remote_code=True).eval() # 同样量化模型
processor = Emu3Processor(image_processor, image_tokenizer, tokenizer)

prepare input

text = ["Please describe the image", "Please describe the image"]
image = Image.open("assets/demo.png")
image = [image, image]

inputs = processor(
text=text,
image=image,
mode='U',
padding_image=True,
padding="longest",
return_tensors="pt",
)

prepare hyper parameters

GENERATION_CONFIG = GenerationConfig(pad_token_id=tokenizer.pad_token_id, bos_token_id=tokenizer.bos_token_id, eos_token_id=tokenizer.eos_token_id)

generate

outputs = model.generate(
inputs.input_ids.to("cuda:0"), # 这里应该使用 inputs.input_ids
generation_config=GENERATION_CONFIG,
max_new_tokens=1024,
attention_mask=inputs.attention_mask.to("cuda:0"),
)

outputs = outputs[:, inputs.input_ids.shape[-1]:]
answers = processor.batch_decode(outputs, skip_special_tokens=True)
for ans in answers:
print(ans)