Add chatglm3 long input example (#10739)

* Add long context input example for chatglm3

* Small fix

* Small fix

* Small fix

python/llm/example/GPU/Long-Context/Chatglm3-32K/README.md

@@ -0,0 +1,126 @@
+# Chatglm3-32k
+In this directory, you will find examples on how you could apply IPEX-LLM INT4/FP8 optimizations on Chatglm3-32K models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [THUDM/chatglm3-6b-32k](https://huggingface.co/THUDM/chatglm3-6b-32k) as reference Chatglm3-32K models.
+
+## 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 `generate()` API
+In the example [generate.py](./generate.py), we show a basic use case for a Chatglm3 model to predict the next N tokens using `generate()` API, with IPEX-LLM INT4/FP8 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/
+```
+
+#### 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/
+```
+
+### 2. Configures OneAPI environment variables
+#### 2.1 Configurations for Linux
+```bash
+source /opt/intel/oneapi/setvars.sh
+```
+#### 2.2 Configurations for Windows
+```cmd
+call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
+```
+> Note: Please make sure you are using **CMD** (**Anaconda Prompt** if using conda) to run the command as PowerShell is not supported.
+### 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
+```
+
+</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 ENABLE_SDP_FUSION=1
+```
+> Note: Please note that `libtcmalloc.so` can be installed by `conda install -c conda-forge -y gperftools=2.10`.
+</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™ A300-Series or Pro A60</summary>
+
+```cmd
+set SYCL_CACHE_PERSISTENT=1
+```
+
+</details>
+
+<details>
+
+<summary>For other Intel dGPU Series</summary>
+
+There is no need to set further environment variables.
+
+</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
+#### 4.1 Using simple prompt
+```
+python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT --low-bit LOW_BIT
+```
+
+Arguments info:
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Chatglm3 model (e.g. `THUDM/chatglm3-6b-32k`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm3-6b-32k'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'What is AI?'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+- `--low-bit LOW_BIT`: argument defining which low bit optimization to use. Options are sym_int4 or fp8. It is default to be `sym_int4`.
+
+#### 4.2 Using long context input prompt
+You can set the `prompt` argument to be a `.txt` file path containing the long context prompt text. An example command using the 8k input size prompt we provide is given below:
+```
+python ./generate.py --repo-id-or-model-path togethercomputer/chatglm3-6b-32k --prompt 8k.txt
+```
+> Note: If you need to run longer input or use less memory, please set `IPEX_LLM_LOW_MEM=1`, which will enable memory optimization and may slightly affect the latency performance.
+#### Sample Output
+#### [THUDM/chatglm3-6b-32k](https://huggingface.co/THUDM/chatglm3-6b-32k)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+<|user|>
+What is AI?
+<|assistant|>
+-------------------- Output --------------------
+[gMASK]sop <|user|>
+What is AI?
+<|assistant|>
+ AI stands for Artificial Intelligence. It refers to the ability of computers and other machines to perform tasks that typically require human intelligence, such as recognizing patterns, making
+```

python/llm/example/GPU/Long-Context/Chatglm3-32K/generate.py

@@ -0,0 +1,86 @@
+#
+# 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 torch
+import time
+import argparse
+import numpy as np
+
+from ipex_llm.transformers import AutoModel
+from transformers import AutoTokenizer
+
+# you could tune the prompt based on your own model,
+# here the prompt tuning refers to https://github.com/THUDM/ChatGLM3/blob/main/PROMPT.md
+CHATGLM_V3_PROMPT_FORMAT = "<|user|>\n{prompt}\n<|assistant|>"
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM3 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm3-6b-32k",
+                        help='The huggingface repo id for the ChatGLM3 model to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--prompt', type=str, default="What is AI?",
+                        help='Prompt to infer')
+    parser.add_argument('--n-predict', type=int, default=32,
+                        help='Max tokens to predict')
+    parser.add_argument('--low-bit', type=str, default="sym_int4",
+                    help='Load model in low bit')
+
+    args = parser.parse_args()
+    model_path = args.repo_id_or_model_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=args.low_bit,
+                                      optimize_model=True,
+                                      trust_remote_code=True,
+                                      use_cache=True)
+    model = model.half().to('xpu')
+
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path,
+                                              trust_remote_code=True)
+
+    # Generate predicted tokens
+    with torch.inference_mode():
+        if not args.prompt.endswith('.txt'):
+            prompt = CHATGLM_V3_PROMPT_FORMAT.format(prompt=args.prompt)
+        else:
+            with open(args.prompt, 'r') as f:
+                prompt = f.read()
+        input_ids = tokenizer.encode(prompt, return_tensors="pt").to('xpu')
+        # ipex_llm model needs a warmup, then inference time can be accurate
+        output = model.generate(input_ids,
+                                max_new_tokens=args.n_predict)
+
+        # start inference
+        st = time.time()
+        # if your selected model is capable of utilizing previous key/value attentions
+        # to enhance decoding speed, but has `"use_cache": false` in its model config,
+        # it is important to set `use_cache=True` explicitly in the `generate` function
+        # to obtain optimal performance with IPEX-LLM INT4 optimizations
+        output = model.generate(input_ids,
+                                max_new_tokens=args.n_predict)
+        torch.xpu.synchronize()
+        end = time.time()
+        output_str = tokenizer.decode(output[0], skip_special_tokens=True)
+        print(f'Inference time: {end-st} s')
+        print('-'*20, 'Prompt', '-'*20)
+        print(prompt)
+        print('-'*20, 'Output', '-'*20)
+        print(output_str)

python/llm/example/GPU/Long-Context/LLaMA2-32K/README.md

@@ -1,11 +1,11 @@
-# Llama2
+# Llama2-32k
-In this directory, you will find examples on how you could apply IPEX-LLM INT4 optimizations on Llama2-32K models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [togethercomputer/Llama-2-7B-32K-Instruct](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct) as reference Llama2-32K models.
+In this directory, you will find examples on how you could apply IPEX-LLM INT4/FP8 optimizations on Llama2-32K models on [Intel GPUs](../../../README.md). For illustration purposes, we utilize the [togethercomputer/Llama-2-7B-32K-Instruct](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct) as reference Llama2-32K models.
 
 ## 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 `generate()` API
-In the example [generate.py](./generate.py), we show a basic use case for a Llama2 model to predict the next N tokens using `generate()` API, with IPEX-LLM INT4 optimizations on Intel GPUs.
+In the example [generate.py](./generate.py), we show a basic use case for a Llama2 model to predict the next N tokens using `generate()` API, with IPEX-LLM INT4/FP8 optimizations on Intel GPUs.
 ### 1. Install
 #### 1.1 Installation on Linux
 We suggest using conda to manage environment:
@@ -95,20 +95,21 @@ There is no need to set further environment variables.
 ### 4. Running examples
 #### 4.1 Using simple prompt
 ```
-python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
+python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT --low-bit LOW_BIT
 ```
 
 Arguments info:
 - `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the Llama2 model (e.g. `togethercomputer/Llama-2-7B-32K-Instruct`) to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'togethercomputer/Llama-2-7B-32K-Instruct'`.
 - `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'What is AI?'`.
 - `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+- `--low-bit LOW_BIT`: argument defining which low bit optimization to use. Options are sym_int4 or fp8. It is default to be `sym_int4`.
 
-#### 4.2 Using 8k input size prompt
+#### 4.2 Using long context input prompt
-You can set the `prompt` argument to be a `.txt` file path containing the 8k size prompt text. An example command using the 8k input size prompt we provide is given below:
+You can set the `prompt` argument to be a `.txt` file path containing the long context prompt text. An example command using the 8k input size prompt we provide is given below:
 ```
 python ./generate.py --repo-id-or-model-path togethercomputer/Llama-2-7B-32K-Instruct --prompt 8k.txt
 ```
-> Note: If you need to use less memory, please set `IPEX_LLM_LOW_MEM=1`, which will enable memory optimization and may slightly affect the latency performance.
+> Note: If you need to run longer input or use less memory, please set `IPEX_LLM_LOW_MEM=1`, which will enable memory optimization and may slightly affect the latency performance.
 #### Sample Output
 #### [togethercomputer/Llama-2-7B-32K-Instruct](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct)
 ```log

python/llm/example/GPU/Long-Context/LLaMA2-32K/generate.py

@@ -48,6 +48,8 @@ if __name__ == '__main__':
                         help='Prompt to infer')
     parser.add_argument('--n-predict', type=int, default=32,
                         help='Max tokens to predict')
+    parser.add_argument('--low-bit', type=str, default="sym_int4",
+                        help='Load model in low bit')
 
     args = parser.parse_args()
     model_path = args.repo_id_or_model_path
@@ -57,7 +59,7 @@ if __name__ == '__main__':
     # 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 = AutoModelForCausalLM.from_pretrained(model_path,
-                                                 load_in_4bit=True,
+                                                 load_in_low_bit=args.low_bit,
                                                  optimize_model=True,
                                                  trust_remote_code=True,
                                                  use_cache=True)