Update Llava GPU Example (intel-analytics#12311)

* update-llava-example * add warmup * small fix on llava example * remove space& extra print prompt * renew example * small fix --------- Co-authored-by: Jinhe Tang <[email protected]>
ch1y0q · Nov 1, 2024 · cd5e22c · cd5e22c
1 parent f53bb4e
commit cd5e22c
Show file tree

Hide file tree

Showing 2 changed files with 83 additions and 359 deletions.
diff --git a/python/llm/example/GPU/PyTorch-Models/Model/llava/README.md b/python/llm/example/GPU/PyTorch-Models/Model/llava/README.md
@@ -1,11 +1,11 @@
 # LLaVA
-In this directory, you will find examples on how you could use IPEX-LLM `optimize_model` API on LLaVA models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [liuhaotian/llava-v1.5-7b](https://huggingface.co/liuhaotian/llava-v1.5-7b) as a reference LLaVA model.
+In this directory, you will find examples on how you could use IPEX-LLM `optimize_model` API to accelerate LLaVA models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [llava-hf/llava-1.5-7b-hf](https://huggingface.co/llava-hf/llava-1.5-7b-hf) as a reference LLaVA 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: Multi-turn chat centered around an image using `generate()` API
-In the example [generate.py](./generate.py), we show a basic use case for a LLaVA model to start a multi-turn chat centered around an image using `generate()` API, with IPEX-LLM INT4 optimizations on Intel GPUs.
+## Example: Predict Tokens using `generate()` API
+In the example [generate.py](./generate.py), we show a basic use case for a LLaVA model to predict the next N tokens using `generate()` API, with IPEX-LLM INT4 optimizations on Intel GPUs.
 ### 1. Install
 #### 1.1 Installation on Linux
 We suggest using conda to manage environment:
@@ -15,12 +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 einops # install dependencies required by llava
-
-git clone https://github.com/haotian-liu/LLaVA.git # clone the llava libary
-cp generate.py ./LLaVA/ # copy our example to the LLaVA folder
-cd LLaVA # change the working directory to the LLaVA folder
-git checkout tags/v1.2.0 -b 1.2.0 # Get the branch which is compatible with transformers 4.36
+pip install transformers==4.43.0
 ```
 
 #### 1.2 Installation on Windows
@@ -32,12 +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 einops # install dependencies required by llava
-
-git clone https://github.com/haotian-liu/LLaVA.git # clone the llava libary
-copy generate.py .\LLaVA\ # copy our example to the LLaVA folder
-cd LLaVA # change the working directory to the LLaVA folder
-git checkout tags/v1.2.0 -b 1.2.0 # Get the branch which is compatible with transformers 4.36
+pip install transformers==4.43.0
 ```
 
 ### 2. Configures OneAPI environment variables for Linux
@@ -116,42 +106,30 @@ 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
 
-```bash
-python ./generate.py --image-path-or-url 'https://llava-vl.github.io/static/images/monalisa.jpg'
+```
+python ./generate.py
 ```
 
-In the example, several arguments can be passed to satisfy your requirements:
-
-- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the LLaVA model (e.g. `liuhaotian/llava-v1.5-7b` to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'liuhaotian/llava-v1.5-7b'`.
-- `--image-path-or-url IMAGE_PATH_OR_URL`: argument defining the input image that the chat will focus on. It is required.
-- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `512`.
-
-If you encounter some network error (which means your machine is unable to access huggingface.co) when running this example, refer to [Trouble Shooting](#4-trouble-shooting) section.
-
+Arguments info:
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the LLaVA model (e.g. `llava-hf/llava-1.5-7b-hf`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'llava-hf/llava-1.5-7b-hf'`.
+- `--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 `'Describe image in detail'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
 
 #### Sample Output
-#### [liuhaotian/llava-v1.5-7b](https://huggingface.co/liuhaotian/llava-v1.5-7b)
+#### [llava-hf/llava-1.5-7b-hf](https://huggingface.co/llava-hf/llava-1.5-7b-hf)
 
 ```log
-USER: Do you know who drew this painting?
-ASSISTANT: Yes, the painting is a portrait of a woman by Leonardo da Vinci. It's a famous artwork known as the "Mona Lisa."
-USER: Can you describe this painting?
-ASSISTANT: The painting features a well-detailed portrait of a woman, painted in oil on a canvas. The woman appears to be a young woman staring straight ahead in a direct gaze towards the viewer. The woman's facial features are rendered sharply in the brush strokes, giving her a lifelike, yet enigmatic expression.
-The background of the image mainly showcases the woman's face, with some hills visible in the lower part of the painting. The artist employs a wide range of shades, evoking a sense of depth and realism in the subject matter. The technique used in this portrait sets it apart from other artworks during the Renaissance period, making it a notable piece in art history.
+Inference time: xxxx s
+-------------------- Input Image --------------------
+http://farm6.staticflickr.com/5268/5602445367_3504763978_z.jpg
+-------------------- Prompt --------------------
+Describe image in detail
+-------------------- Output --------------------
+<s> USER: <image>
+Describe image in detail ASSISTANT: The image features a young girl holding a white teddy bear in her hands. She is smiling and appears to be enjoying the moment. The girl is
 ```
 
-The sample input image is:
-
-<a href="https://llava-vl.github.io/static/images/monalisa.jpg"><img width=400px src="https://llava-vl.github.io/static/images/monalisa.jpg" ></a>
-
-### 5 Trouble shooting
-
-#### 5.1 SSLError
-If you encounter the following output, it means your machine has some trouble accessing huggingface.co.
-```log
-requests.exceptions.SSLError: (MaxRetryError("HTTPSConnectionPool(host='huggingface.co', port=443): Max retries exceeded with url: /openai/clip-vit-large-patch14-336/resolve/main/config.json (Caused by SSLError(SSLZeroReturnError(6, 'TLS/SSL connection has been closed (EOF) (_ssl.c:1129)')))"),
-```
+The sample input image is (which is fetched from [COCO dataset](https://cocodataset.org/#explore?id=264959)):
 
-You can resolve this problem with the following steps:
-1. Download https://huggingface.co/openai/clip-vit-large-patch14-336 on some machine that can access huggingface.co, and put it in huggingface's local cache (default to be `~/.cache/huggingface/hub`) on the machine that you are going to run this example.
-2. Set the environment variable (`export TRANSFORMERS_OFFLINE=1`) before you run the example.
+<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>