[LLM] Add more transformers examples (ChatGLM) (#8521)

* Add example for chatglm v1 and other small fixes

* Small fix

* Small further fix

* Small fix

* Update based on comments & updates for client windows recommended settingts

* Small fix

* Small refactor

* Small fix

* Small fix

* Small fix to dolly v1

* Small fix

python/llm/example/transformers/transformers_int4/README.md

@@ -6,12 +6,9 @@ To run the examples, we recommend using Intel® Xeon® processors (server), or >
 
 For OS, BigDL-LLM supports Ubuntu 20.04 or later, CentOS 7 or later, and Windows 10/11.
 
-## Best Known Configuration
+## Best Known Configuration on Linux
-For better performance, it is recommended to set environment variables with the help of BigDL-Nano:
+For better performance, it is recommended to set environment variables on Linux with the help of BigDL-Nano:
 ```bash
 pip install bigdl-nano
+source bigdl-nano-init
 ```
-following with
-| Linux | Windows (powershell)|
-|:------|:-------|
-|`source bigdl-nano-init`|`bigdl-nano-init`|

python/llm/example/transformers/transformers_int4/chatglm/README.md

@@ -0,0 +1,72 @@
+# ChatGLM
+
+In this directory, you will find examples on how you could apply BigDL-LLM INT4 optimizations on ChatGLM models. For illustration purposes, we utilize the [THUDM/chatglm-6b](https://huggingface.co/THUDM/chatglm-6b) as a reference ChatGLM model.
+
+## 0. Requirements
+To run these examples with BigDL-LLM, we have some recommended requirements for your machine, please refer to [here](../README.md#recommended-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 ChatGLM model to predict the next N tokens using `generate()` API, with BigDL-LLM INT4 optimizations.
+### 1. Install
+We suggest using conda to manage environment:
+```bash
+conda create -n llm python=3.9
+conda activate llm
+
+pip install bigdl-llm[all] # install bigdl-llm with 'all' option
+```
+
+### 2. Run
+```
+python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
+```
+
+Arguments info:
+- `--repo-id-or-model-path REPO_ID_OR_MODEL_PATH`: argument defining the huggingface repo id for the ChatGLM model to be downloaded, or the path to the huggingface checkpoint folder. It is default to be `'THUDM/chatglm-6b'`.
+- `--prompt PROMPT`: argument defining the prompt to be infered (with integrated prompt format for chat). It is default to be `'AI是什么？'`.
+- `--n-predict N_PREDICT`: argument defining the max number of tokens to predict. It is default to be `32`.
+
+> **Note**: When loading the model in 4-bit, BigDL-LLM converts linear layers in the model into INT4 format. In theory, a *X*B model saved in 16-bit will requires approximately 2*X* GB of memory for loading, and ~0.5*X* GB memory for further inference.
+>
+> Please select the appropriate size of the ChatGLM model based on the capabilities of your machine.
+
+#### 2.1 Client
+On client Windows machine, it is recommended to run directly with full utilization of all cores:
+```powershell
+python ./generate.py 
+```
+
+#### 2.2 Server
+For optimal performance on server, it is recommended to set several environment variables (refer to [here](../README.md#best-known-configuration) for more information), and run the example with all the physical cores of a single socket.
+
+E.g. on Linux,
+```bash
+# set BigDL-Nano env variables
+source bigdl-nano-init
+
+# e.g. for a server with 48 cores per socket
+export OMP_NUM_THREADS=48
+numactl -C 0-47 -m 0 python ./generate.py
+```
+
+#### 2.3 Sample Output
+#### [THUDM/chatglm-6b](https://huggingface.co/THUDM/chatglm-6b)
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+问：AI是什么？
+答：
+-------------------- Output --------------------
+问:AI是什么?
+答: AI是人工智能(Artificial Intelligence)的缩写,指的是一种能够模拟人类智能的技术或系统。AI系统可以通过学习、推理、解决问题等方式,实现类似于
+```
+
+```log
+Inference time: xxxx s
+-------------------- Prompt --------------------
+问：What is AI?
+答：
+-------------------- Output --------------------
+问:What is AI?
+答: AI stands for "Artificial Intelligence." AI refers to the development of computer systems that can perform tasks that typically require human intelligence, such as recognizing speech, understanding natural
+```

python/llm/example/transformers/transformers_int4/chatglm/generate.py

@@ -0,0 +1,69 @@
+#
+# 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 bigdl.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://huggingface.co/THUDM/chatglm-6b/blob/294cb13118a1e08ad8449ca542624a5c6aecc401/modeling_chatglm.py#L1281
+CHATGLM_V1_PROMPT_FORMAT = "问：{prompt}\n答："
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for ChatGLM model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="THUDM/chatglm-6b",
+                        help='The huggingface repo id for the ChatGLM model to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--prompt', type=str, default="AI是什么？",
+                        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
+
+    # Load model in 4 bit,
+    # which convert the relevant layers in the model into INT4 format
+    model = AutoModel.from_pretrained(model_path,
+                                      load_in_4bit=True,
+                                      trust_remote_code=True)
+
+    # Load tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(model_path,
+                                              trust_remote_code=True)
+    
+    # Generate predicted tokens
+    with torch.inference_mode():
+        prompt = CHATGLM_V1_PROMPT_FORMAT.format(prompt=args.prompt)
+        input_ids = tokenizer.encode(prompt, return_tensors="pt")
+        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 BigDL-LLM INT4 optimizations
+        output = model.generate(input_ids,
+                                max_new_tokens=args.n_predict)
+        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/transformers/transformers_int4/dolly_v1/README.md

@@ -15,10 +15,7 @@ conda activate llm
 pip install bigdl-llm[all] # install bigdl-llm with 'all' option
 ```
 
-### 2. Config
+### 2. Run
-It is recommended to set several environment variables for better performance. Please refer to [here](../README.md#best-known-configuration) for more information.
-
-### 3. Run
 ```
 python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
 ```
@@ -32,27 +29,26 @@ Arguments info:
 >
 > Please select the appropriate size of the Dolly v1 model based on the capabilities of your machine.
 
-#### 3.1 Client
+#### 2.1 Client
-For better utilization of multiple cores on the client machine, it is recommended to use all the performance-cores along with their hyperthreads.
+On client Windows machine, it is recommended to run directly with full utilization of all cores:
-
-E.g. on Windows,
 ```powershell
-# for a client machine with 8 Performance-cores
-$env:OMP_NUM_THREADS=16
 python ./generate.py 
 ```
 
-#### 3.2 Server
+#### 2.2 Server
-On server, it is recommended to run the example with all the physical cores of a single socket.
+For optimal performance on server, it is recommended to set several environment variables (refer to [here](../README.md#best-known-configuration) for more information), and run the example with all the physical cores of a single socket.
 
 E.g. on Linux,
 ```bash
-# for a server with 48 cores per socket
+# set BigDL-Nano env variables
+source bigdl-nano-init
+
+# e.g. for a server with 48 cores per socket
 export OMP_NUM_THREADS=48
-numactl -C 0-47 -m 0 python -u ./generate.py
+numactl -C 0-47 -m 0 python ./generate.py
 ```
 
-#### 3.3 Sample Output
+#### 2.3 Sample Output
 #### [databricks/dolly-v1-6b](https://huggingface.co/databricks/dolly-v1-6b)
 ```log
 Inference time: xxxx s

python/llm/example/transformers/transformers_int4/dolly_v1/generate.py

@@ -23,7 +23,7 @@ from bigdl.llm.transformers import AutoModelForCausalLM
 from transformers import AutoTokenizer
 
 # you could tune the prompt based on your own model,
-# here the prompt format refers to https://huggingface.co/databricks/dolly-v1-6b#generate-text
+# here the prompt tuning refers to https://huggingface.co/databricks/dolly-v1-6b#generate-text
 DOLLY_V1_PROMPT_FORMAT = """Below is an instruction that describes a task. Write a response that appropriately completes the request.
 
 ### Instruction:
@@ -33,7 +33,7 @@ DOLLY_V1_PROMPT_FORMAT = """Below is an instruction that describes a task. Write
 """
 
 if __name__ == '__main__':
-    parser = argparse.ArgumentParser(description='Transformer INT4 example for Dolly v1 model')
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Dolly v1 model')
     parser.add_argument('--repo-id-or-model-path', type=str, default="databricks/dolly-v1-6b",
                         help='The huggingface repo id for the Dolly v1 model to be downloaded'
                              ', or the path to the huggingface checkpoint folder')
@@ -59,11 +59,12 @@ if __name__ == '__main__':
         input_ids = tokenizer.encode(prompt, return_tensors="pt")
         end_key_token_id=tokenizer.encode("### End")[0]
         st = time.time()
-        # if your selected model is capable of utilizing previous key/value attentions
+        # enabling `use_cache=True` allows the model to utilize the previous
-        # to enhance decoding speed, but has `"use_cache": false` in its model config,
+        # key/values attentions to speed up decoding;
-        # it is important to set `use_cache=True` explicitly in the `generate` function
+        # to obtain optimal performance with BigDL-LLM INT4 optimizations,
-        # to obtain optimal performance with BigDL-LLM INT4 optimizations
+        # it is important to set use_cache=True for Dolly v1 models
         output = model.generate(input_ids,
+                                use_cache=True,
                                 max_new_tokens=args.n_predict,
                                 pad_token_id=tokenizer.pad_token_id,
                                 eos_token_id=end_key_token_id)

python/llm/example/transformers/transformers_int4/mpt/README.md

@@ -16,10 +16,7 @@ pip install bigdl-llm[all] # install bigdl-llm with 'all' option
 pip install einops  # additional package required for mpt-7b-chat to conduct generation
 ```
 
-### 2. Config
+### 2. Run
-It is recommended to set several environment variables for better performance. Please refer to [here](../README.md#best-known-configuration) for more information.
-
-### 3. Run
 ```
 python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt PROMPT --n-predict N_PREDICT
 ```
@@ -33,27 +30,26 @@ Arguments info:
 >
 > Please select the appropriate size of the MPT model based on the capabilities of your machine.
 
-#### 3.1 Client
+#### 2.1 Client
-For better utilization of multiple cores on the client machine, it is recommended to use all the performance-cores along with their hyperthreads.
+On client Windows machine, it is recommended to run directly with full utilization of all cores:
-
-E.g. on Windows,
 ```powershell
-# for a client machine with 8 Performance-cores
-$env:OMP_NUM_THREADS=16
 python ./generate.py 
 ```
 
-#### 3.2 Server
+#### 2.2 Server
-On server, it is recommended to run the example with all the physical cores of a single socket.
+For optimal performance on server, it is recommended to set several environment variables (refer to [here](../README.md#best-known-configuration) for more information), and run the example with all the physical cores of a single socket.
 
 E.g. on Linux,
 ```bash
-# for a server with 48 cores per socket
+# set BigDL-Nano env variables
+source bigdl-nano-init
+
+# e.g. for a server with 48 cores per socket
 export OMP_NUM_THREADS=48
-numactl -C 0-47 -m 0 python -u ./generate.py
+numactl -C 0-47 -m 0 python ./generate.py
 ```
 
-#### 3.3 Sample Output
+#### 2.3 Sample Output
 #### [mosaicml/mpt-7b-chat](https://huggingface.co/mosaicml/mpt-7b-chat)
 ```log
 Inference time: xxxx s

python/llm/example/transformers/transformers_int4/mpt/generate.py

@@ -25,7 +25,7 @@ from transformers import AutoTokenizer
 MPT_PROMPT_FORMAT = "<human>{prompt} <bot>"
 
 if __name__ == '__main__':
-    parser = argparse.ArgumentParser(description='Transformer INT4 example for MPT model')
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for MPT model')
     parser.add_argument('--repo-id-or-model-path', type=str, default="mosaicml/mpt-7b-chat",
                         help='The huggingface repo id for the MPT to be downloaded'
                              ', or the path to the huggingface checkpoint folder')
@@ -40,8 +40,8 @@ if __name__ == '__main__':
     # Load model in 4 bit,
     # which convert the relevant layers in the model into INT4 format
     model = AutoModelForCausalLM.from_pretrained(model_path,
-                                                 trust_remote_code=True,
+                                                 load_in_4bit=True,
-                                                 load_in_4bit=True)
+                                                 trust_remote_code=True)
 
     # Load tokenizer
     tokenizer = AutoTokenizer.from_pretrained(model_path,