diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/README.md b/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/README.md index 64192eb8..3a47448f 100644 --- a/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/README.md +++ b/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/README.md @@ -5,7 +5,7 @@ In this directory, you will find examples on how you could apply IPEX-LLM INT4 o 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_6 model to predict the next N tokens using `chat()` API, with IPEX-LLM INT4 optimizations on Intel GPUs. +In the example [chat.py](./chat.py), we show a basic use case for a MiniCPM-V-2_6 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: @@ -15,7 +15,7 @@ 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 transformers==4.40.0 trl +pip install transformers==4.40.0 trl ``` #### 1.2 Installation on Windows @@ -27,7 +27,7 @@ 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 transformers==4.40.0 trl +pip install transformers==4.40.0 trl ``` ### 2. Configures OneAPI environment variables for Linux @@ -106,15 +106,23 @@ set SYCL_CACHE_PERSISTENT=1 > 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?' -``` +- chat without streaming mode: + ``` + python ./generate.py --prompt 'What is in the image?' + ``` +- chat in streaming mode: + ``` + python ./generate.py --prompt 'What is in the image?' --stream + ``` + +> [!TIP] +> For chatting in streaming mode, it is recommended to set the environment variable `PYTHONUNBUFFERED=1`. Arguments info: - `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the MiniCPM-V-2_6 (e.g. `openbmb/MiniCPM-V-2_6`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'openbmb/MiniCPM-V-2_6'`. - `--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`. +- `--stream`: flag to chat in streaming mode #### Sample Output @@ -122,21 +130,20 @@ Arguments info: ```log Inference time: xxxx s --------------------- Input -------------------- +-------------------- Input Image -------------------- http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg --------------------- Prompt -------------------- +-------------------- Input Prompt -------------------- What is in the image? --------------------- Output -------------------- -The image features a young child holding a white teddy bear with a pink tutu. The child is wearing a striped dress and is standing in front of a stone wall with some red flowers in the background. +-------------------- Chat Output -------------------- +The image features a young child holding a white teddy bear wearing a pink dress. The background shows some red flowers and stone walls, suggesting an outdoor setting. ``` ```log -Inference time: xxxx s --------------------- Input -------------------- +-------------------- Input Image -------------------- http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg --------------------- Prompt -------------------- +-------------------- Input Prompt -------------------- 图片里有什么？ --------------------- Output -------------------- -这幅图片展示了一个年幼的孩子，可能是一个蹒跚学步的幼儿，手里拿着一个毛绒玩具熊。孩子穿着一件条纹连衣裙，主要颜色是粉红色和白色。毛绒熊是白色的，戴着一条粉色的蝴蝶结围裙。背景中有红色的花朵，暗示着室外的环境，可能是一个花园或公园。 +-------------------- Stream Chat Output -------------------- +图片中有一个穿着粉红色连衣裙的小孩，手里拿着一只穿着粉色芭蕾裙的白色泰迪熊。背景中有红色花朵和石头墙，表明照片可能是在户外拍摄的。 ``` The sample input image is (which is fetched from [COCO dataset](https://cocodataset.org/#explore?id=264959)): diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/chat.py b/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/chat.py new file mode 100644 index 00000000..d30fbd0d --- /dev/null +++ b/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/chat.py @@ -0,0 +1,111 @@ +# +# 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. +# + + +import os +import time +import argparse +import requests +import torch +from PIL import Image +from ipex_llm.transformers import AutoModel +from transformers import AutoTokenizer + + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description='Predict Tokens using `chat()` API for openbmb/MiniCPM-V-2_6 model') + parser.add_argument('--repo-id-or-model-path', type=str, default="openbmb/MiniCPM-V-2_6", + help='The huggingface repo id for the openbmb/MiniCPM-V-2_6 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('--stream', action='store_true', + help='Whether to chat in streaming mode') + + 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="sym_int4", + optimize_model=True, + trust_remote_code=True, + use_cache=True, + modules_to_not_convert=["vpm", "resampler"]) + model = model.half().to('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_6/blob/main/README.md + msgs = [{'role': 'user', 'content': [image, args.prompt]}] + + # ipex_llm model needs a warmup, then inference time can be accurate + model.chat( + image=None, + msgs=msgs, + context=None, + tokenizer=tokenizer, + ) + + if args.stream: + res = model.chat( + image=None, + msgs=msgs, + context=None, + tokenizer=tokenizer, + stream=True + ) + + print('-'*20, 'Input Image', '-'*20) + print(image_path) + print('-'*20, 'Input Prompt', '-'*20) + print(args.prompt) + print('-'*20, 'Stream Chat Output', '-'*20) + for new_text in res: + print(new_text, flush=True, end='') + else: + st = time.time() + res = model.chat( + image=None, + msgs=msgs, + context=None, + tokenizer=tokenizer, + ) + torch.xpu.synchronize() + end = time.time() + + print(f'Inference time: {end-st} s') + print('-'*20, 'Input Image', '-'*20) + print(image_path) + print('-'*20, 'Input Prompt', '-'*20) + print(args.prompt) + print('-'*20, 'Chat Output', '-'*20) + print(res) diff --git a/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/generate.py b/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/generate.py deleted file mode 100644 index 85f42060..00000000 --- a/python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/generate.py +++ /dev/null @@ -1,175 +0,0 @@ -# -# 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_6 model') - parser.add_argument('--repo-id-or-model-path', type=str, default="openbmb/MiniCPM-V-2_6", - help='The huggingface repo id for the openbmb/MiniCPM-V-2_6 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="sym_int4", - optimize_model=True, - trust_remote_code=True, - use_cache=True) - model = model.half().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_6/blob/main/README.md - msgs = [{'role': 'user', 'content': [image, args.prompt]}] - st = time.time() - res = model.chat( - image=None, - 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)