Add MiniCPM-V-2 GPU example (#11699)
* Add MiniCPM-V-2 GPU example * add example in README.md * add example in README.md
This commit is contained in:
Jin, Qiao
GitHub
parent
8fb36b9f4a
commit
808d9a7bae
3 changed files with 313 additions and 0 deletions
|
|
@ -309,6 +309,8 @@ Over 50 models have been optimized/verified on `ipex-llm`, including *LLaMA/LLaM
|
|||
| CodeGeeX2 | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/codegeex2) | [link](python/llm/example/GPU/HuggingFace/LLM/codegeex2) |
|
||||
| MiniCPM | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/minicpm) | [link](python/llm/example/GPU/HuggingFace/LLM/minicpm) |
|
||||
| MiniCPM-V | | [link](python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V) |
|
||||
| MiniCPM-V-2 | | [link](python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2) |
|
||||
| MiniCPM-Llama3-V-2_5 | | [link](python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-Llama3-V-2_5) |
|
||||
|
||||
## Get Support
|
||||
- Please report a bug or raise a feature request by opening a [Github Issue](https://github.com/intel-analytics/ipex-llm/issues)
|
||||
|
|
|
|||
|
|
@ -0,0 +1,135 @@
|
|||
# MiniCPM-V-2
|
||||
In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on MiniCPM-V-2 models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [openbmb/MiniCPM-V-2](https://huggingface.co/openbmb/MiniCPM-V-2) as a reference MiniCPM-V-2 model.
|
||||
|
||||
## 0. Requirements
|
||||
To run these examples with IPEX-LLM on Intel GPUs, we have some recommended requirements for your machine, please refer to [here](../../../README.md#requirements) for more information.
|
||||
|
||||
## Example: Predict Tokens using `chat()` API
|
||||
In the example [generate.py](./generate.py), we show a basic use case for a MiniCPM-V-2 model to predict the next N tokens using `chat()` API, with IPEX-LLM INT4 optimizations on Intel GPUs.
|
||||
### 1. Install
|
||||
#### 1.1 Installation on Linux
|
||||
We suggest using conda to manage environment:
|
||||
```bash
|
||||
conda create -n llm python=3.11
|
||||
conda activate llm
|
||||
# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
|
||||
pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
|
||||
|
||||
pip install timm peft
|
||||
```
|
||||
|
||||
#### 1.2 Installation on Windows
|
||||
We suggest using conda to manage environment:
|
||||
```bash
|
||||
conda create -n llm python=3.11 libuv
|
||||
conda activate llm
|
||||
|
||||
# below command will install intel_extension_for_pytorch==2.1.10+xpu as default
|
||||
pip install --pre --upgrade ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
|
||||
|
||||
pip install timm peft
|
||||
```
|
||||
|
||||
### 2. Configures OneAPI environment variables for Linux
|
||||
|
||||
> [!NOTE]
|
||||
> Skip this step if you are running on Windows.
|
||||
|
||||
This is a required step on Linux for APT or offline installed oneAPI. Skip this step for PIP-installed oneAPI.
|
||||
|
||||
```bash
|
||||
source /opt/intel/oneapi/setvars.sh
|
||||
```
|
||||
|
||||
### 3. Runtime Configurations
|
||||
For optimal performance, it is recommended to set several environment variables. Please check out the suggestions based on your device.
|
||||
#### 3.1 Configurations for Linux
|
||||
<details>
|
||||
|
||||
<summary>For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series</summary>
|
||||
|
||||
```bash
|
||||
export USE_XETLA=OFF
|
||||
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
|
||||
export SYCL_CACHE_PERSISTENT=1
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
<details>
|
||||
|
||||
<summary>For Intel Data Center GPU Max Series</summary>
|
||||
|
||||
```bash
|
||||
export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so
|
||||
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
|
||||
export SYCL_CACHE_PERSISTENT=1
|
||||
export ENABLE_SDP_FUSION=1
|
||||
```
|
||||
> Note: Please note that `libtcmalloc.so` can be installed by `conda install -c conda-forge -y gperftools=2.10`.
|
||||
</details>
|
||||
|
||||
<details>
|
||||
|
||||
<summary>For Intel iGPU</summary>
|
||||
|
||||
```bash
|
||||
export SYCL_CACHE_PERSISTENT=1
|
||||
export BIGDL_LLM_XMX_DISABLED=1
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
#### 3.2 Configurations for Windows
|
||||
<details>
|
||||
|
||||
<summary>For Intel iGPU</summary>
|
||||
|
||||
```cmd
|
||||
set SYCL_CACHE_PERSISTENT=1
|
||||
set BIGDL_LLM_XMX_DISABLED=1
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
<details>
|
||||
|
||||
<summary>For Intel Arc™ A-Series Graphics</summary>
|
||||
|
||||
```cmd
|
||||
set SYCL_CACHE_PERSISTENT=1
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
> [!NOTE]
|
||||
> For the first time that each model runs on Intel iGPU/Intel Arc™ A300-Series or Pro A60, it may take several minutes to compile.
|
||||
### 4. Running examples
|
||||
|
||||
```
|
||||
python ./generate.py --prompt 'What is in the image?'
|
||||
```
|
||||
|
||||
Arguments info:
|
||||
- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the MiniCPM-V-2 (e.g. `openbmb/MiniCPM-V-2`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'openbmb/MiniCPM-V-2'`.
|
||||
- `--image-url-or-path IMAGE_URL_OR_PATH`: argument defining the image to be infered. It is default to be `'http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg'`.
|
||||
- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'What is in the image?'`.
|
||||
- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
|
||||
|
||||
#### Sample Output
|
||||
|
||||
#### [openbmb/MiniCPM-V-2](https://huggingface.co/openbmb/MiniCPM-V-2)
|
||||
|
||||
```log
|
||||
Inference time: xxxx s
|
||||
-------------------- Input --------------------
|
||||
http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg
|
||||
-------------------- Prompt --------------------
|
||||
What is in the image?
|
||||
-------------------- Output --------------------
|
||||
In the image, there is a young child holding a teddy bear. The teddy bear appears to be dressed in a pink tutu. The child is also wearing a red and white striped dress. The background of the image includes a stone wall and some red flowers.
|
||||
```
|
||||
|
||||
The sample input image is (which is fetched from [COCO dataset](https://cocodataset.org/#explore?id=264959)):
|
||||
|
||||
<a href="http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg"><img width=400px src="http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg" ></a>
|
||||
|
|
@ -0,0 +1,176 @@
|
|||
#
|
||||
# Copyright 2016 The BigDL Authors.
|
||||
#
|
||||
# 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.
|
||||
#
|
||||
|
||||
|
||||
from typing import List, Tuple, Optional, Union
|
||||
import math
|
||||
import timm
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
|
||||
# patched: `timm` has limited support for XPU backend, so we need to use CPU as a workaround
|
||||
def resample_abs_pos_embed(
|
||||
posemb: torch.Tensor,
|
||||
new_size: List[int],
|
||||
old_size: Optional[List[int]] = None,
|
||||
num_prefix_tokens: int = 1,
|
||||
interpolation: str = 'bicubic',
|
||||
antialias: bool = True,
|
||||
verbose: bool = False,
|
||||
):
|
||||
# sort out sizes, assume square if old size not provided
|
||||
num_pos_tokens = posemb.shape[1]
|
||||
num_new_tokens = new_size[0] * new_size[1] + num_prefix_tokens
|
||||
if num_new_tokens == num_pos_tokens and new_size[0] == new_size[1]:
|
||||
return posemb
|
||||
|
||||
if old_size is None:
|
||||
hw = int(math.sqrt(num_pos_tokens - num_prefix_tokens))
|
||||
old_size = hw, hw
|
||||
|
||||
if num_prefix_tokens:
|
||||
posemb_prefix, posemb = posemb[:, :num_prefix_tokens], posemb[:, num_prefix_tokens:]
|
||||
else:
|
||||
posemb_prefix, posemb = None, posemb
|
||||
|
||||
# do the interpolation
|
||||
embed_dim = posemb.shape[-1]
|
||||
orig_dtype = posemb.dtype
|
||||
posemb = posemb.float() # interpolate needs float32
|
||||
posemb = posemb.reshape(1, old_size[0], old_size[1], -1).permute(0, 3, 1, 2)
|
||||
#posemb = F.interpolate(posemb, size=new_size, mode=interpolation, antialias=antialias)
|
||||
posemb = F.interpolate(posemb.to("cpu"), size=new_size, mode=interpolation, antialias=antialias).to(posemb.device)
|
||||
posemb = posemb.permute(0, 2, 3, 1).reshape(1, -1, embed_dim)
|
||||
posemb = posemb.to(orig_dtype)
|
||||
|
||||
# add back extra (class, etc) prefix tokens
|
||||
if posemb_prefix is not None:
|
||||
posemb = torch.cat([posemb_prefix, posemb], dim=1)
|
||||
|
||||
if not torch.jit.is_scripting() and verbose:
|
||||
_logger.info(f'Resized position embedding: {old_size} to {new_size}.')
|
||||
|
||||
return posemb
|
||||
|
||||
|
||||
def _pos_embed(self, x: torch.Tensor) -> torch.Tensor:
|
||||
if self.pos_embed is None:
|
||||
return x.view(x.shape[0], -1, x.shape[-1])
|
||||
|
||||
if self.dynamic_img_size:
|
||||
B, H, W, C = x.shape
|
||||
pos_embed = resample_abs_pos_embed(
|
||||
self.pos_embed,
|
||||
(H, W),
|
||||
num_prefix_tokens=0 if self.no_embed_class else self.num_prefix_tokens,
|
||||
)
|
||||
x = x.view(B, -1, C)
|
||||
else:
|
||||
pos_embed = self.pos_embed
|
||||
|
||||
to_cat = []
|
||||
if self.cls_token is not None:
|
||||
to_cat.append(self.cls_token.expand(x.shape[0], -1, -1))
|
||||
if self.reg_token is not None:
|
||||
to_cat.append(self.reg_token.expand(x.shape[0], -1, -1))
|
||||
|
||||
if self.no_embed_class:
|
||||
# deit-3, updated JAX (big vision)
|
||||
# position embedding does not overlap with class token, add then concat
|
||||
x = x + pos_embed
|
||||
if to_cat:
|
||||
x = torch.cat(to_cat + [x], dim=1)
|
||||
else:
|
||||
# original timm, JAX, and deit vit impl
|
||||
# pos_embed has entry for class token, concat then add
|
||||
if to_cat:
|
||||
x = torch.cat(to_cat + [x], dim=1)
|
||||
x = x + pos_embed
|
||||
|
||||
return self.pos_drop(x)
|
||||
|
||||
|
||||
setattr(timm.models.VisionTransformer, "_pos_embed", _pos_embed)
|
||||
|
||||
|
||||
import os
|
||||
import time
|
||||
import argparse
|
||||
import requests
|
||||
from PIL import Image
|
||||
from ipex_llm.transformers import AutoModel
|
||||
from transformers import AutoTokenizer
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for openbmb/MiniCPM-V-2 model')
|
||||
parser.add_argument('--repo-id-or-model-path', type=str, default="openbmb/MiniCPM-V-2",
|
||||
help='The huggingface repo id for the openbmb/MiniCPM-V-2 model to be downloaded'
|
||||
', or the path to the huggingface checkpoint folder')
|
||||
parser.add_argument('--image-url-or-path', type=str,
|
||||
default='http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg',
|
||||
help='The URL or path to the image to infer')
|
||||
parser.add_argument('--prompt', type=str, default="What is in the image?",
|
||||
help='Prompt to infer')
|
||||
parser.add_argument('--n-predict', type=int, default=32,
|
||||
help='Max tokens to predict')
|
||||
|
||||
args = parser.parse_args()
|
||||
model_path = args.repo_id_or_model_path
|
||||
image_path = args.image_url_or_path
|
||||
|
||||
# Load model in 4 bit,
|
||||
# which convert the relevant layers in the model into INT4 format
|
||||
# When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the from_pretrained function.
|
||||
# This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
|
||||
model = AutoModel.from_pretrained(model_path,
|
||||
load_in_low_bit="asym_int4",
|
||||
optimize_model=True,
|
||||
trust_remote_code=True,
|
||||
modules_to_not_convert=["vpm", "resampler", "lm_head"],
|
||||
use_cache=True)
|
||||
model = model.float().to(device='xpu')
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_path,
|
||||
trust_remote_code=True)
|
||||
model.eval()
|
||||
|
||||
query = args.prompt
|
||||
if os.path.exists(image_path):
|
||||
image = Image.open(image_path).convert('RGB')
|
||||
else:
|
||||
image = Image.open(requests.get(image_path, stream=True).raw).convert('RGB')
|
||||
|
||||
# Generate predicted tokens
|
||||
# here the prompt tuning refers to https://huggingface.co/openbmb/MiniCPM-V-2/blob/main/README.md
|
||||
msgs = [{'role': 'user', 'content': args.prompt}]
|
||||
st = time.time()
|
||||
res, context, _ = model.chat(
|
||||
image=image,
|
||||
msgs=msgs,
|
||||
context=None,
|
||||
tokenizer=tokenizer,
|
||||
sampling=False,
|
||||
temperature=0.7
|
||||
)
|
||||
end = time.time()
|
||||
print(f'Inference time: {end-st} s')
|
||||
print('-'*20, 'Input', '-'*20)
|
||||
print(image_path)
|
||||
print('-'*20, 'Prompt', '-'*20)
|
||||
print(args.prompt)
|
||||
output_str = res
|
||||
print('-'*20, 'Output', '-'*20)
|
||||
print(output_str)
|
||||
Loading…
Reference in a new issue