Add example: MiniCPM-V (#11570)

README.md

@@ -305,6 +305,7 @@ Over 50 models have been optimized/verified on `ipex-llm`, including *LLaMA/LLaM
 | Command-R/cohere | [link](python/llm/example/CPU/HF-Transformers-AutoModels/Model/cohere) | [link](python/llm/example/GPU/HuggingFace/LLM/cohere) |
 | 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) |
 
 ## Get Support
 - Please report a bug or raise a feature request by opening a [Github Issue](https://github.com/intel-analytics/ipex-llm/issues)

python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V/README.md

@@ -0,0 +1,135 @@
+# MiniCPM-V
+In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on MiniCPM-V models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [openbmb/MiniCPM-V](https://huggingface.co/openbmb/MiniCPM-V) as a reference MiniCPM-V 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 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
+```
+
+#### 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
+```
+
+### 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 (e.g. `openbmb/MiniCPM-V`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'openbmb/MiniCPM-V'`.
+- `--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](https://huggingface.co/openbmb/MiniCPM-V)
+
+```log
+Inference time: xxxx s
+-------------------- Input --------------------
+https://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg
+-------------------- Prompt --------------------
+What is in the image?
+-------------------- Output --------------------
+The image showcases a young child holding a small white teddy bear. The teddy bear has a pink ribbon around its neck, and the child seems to be showing it off with a smile. Behind the child, there's a stone wall with red flowers, adding a touch of color to the scene.
+```
+
+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>

python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V/generate.py

@@ -0,0 +1,175 @@
+#
+# 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 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="openbmb/MiniCPM-V",
+                        help='The huggingface repo id for the openbmb/MiniCPM-V 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_4bit=True,
+                                      optimize_model=False,
+                                      trust_remote_code=True,
+                                      modules_to_not_convert=["vpm", "resampler"],
+                                      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/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=True,
+     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)