Update MiniCPM-V-2_6 Example (#11958)

* Update example scripts regarding warmup, stream generate, moudles to not convert, etc.

* Update readme accordingly

* Fix based on comments

* Small fix

* Remove n_predict

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
 
+- 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 --------------------
+-------------------- Chat 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.
+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 Image --------------------
--------------------- Input --------------------
 http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg
--------------------- Prompt --------------------
+-------------------- Input Prompt --------------------
 图片里有什么？
--------------------- Output --------------------
+-------------------- Stream Chat 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。
+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)):
 

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)

python/llm/example/GPU/HuggingFace/Multimodal/MiniCPM-V-2_6/generate.py

@@ -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)