LLM: Add qlora finetunning CPU example (#9275)

* add qlora finetunning example

* update readme

* update example

* remove merge.py and update readme

python/llm/example/CPU/QLoRA-FineTuning/README.md

@@ -0,0 +1,76 @@
+# Finetuning LLAMA Using QLoRA (experimental support)
+
+This example demonstrates how to finetune a llama2-7b model using Big-LLM 4bit optimizations on [Intel CPUs](../README.md).
+
+
+## Example: Finetune llama2-7b using QLoRA
+
+This example is ported from [bnb-4bit-training](https://colab.research.google.com/drive/1VoYNfYDKcKRQRor98Zbf2-9VQTtGJ24k). 
+
+### 1. Install
+
+```bash
+conda create -n llm python=3.9
+conda activate llm
+pip install --pre --upgrade bigdl-llm[all]
+pip install transformers==4.34.0
+pip install peft==0.5.0
+pip install datasets
+```
+
+### 2. Finetune model
+
+```
+python ./qlora_finetuning_cpu.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --dataset DATASET
+```
+
+#### Sample Output
+```log
+{'loss': 2.5668, 'learning_rate': 0.0002, 'epoch': 0.03}
+{'loss': 1.6988, 'learning_rate': 0.00017777777777777779, 'epoch': 0.06}
+{'loss': 1.3073, 'learning_rate': 0.00015555555555555556, 'epoch': 0.1}
+{'loss': 1.3495, 'learning_rate': 0.00013333333333333334, 'epoch': 0.13}
+{'loss': 1.1746, 'learning_rate': 0.00011111111111111112, 'epoch': 0.16}
+{'loss': 1.0794, 'learning_rate': 8.888888888888889e-05, 'epoch': 0.19}
+{'loss': 1.2214, 'learning_rate': 6.666666666666667e-05, 'epoch': 0.22}
+{'loss': 1.1698, 'learning_rate': 4.4444444444444447e-05, 'epoch': 0.26}
+{'loss': 1.2044, 'learning_rate': 2.2222222222222223e-05, 'epoch': 0.29}
+{'loss': 1.1516, 'learning_rate': 0.0, 'epoch': 0.32}
+{'train_runtime': 474.3254, 'train_samples_per_second': 1.687, 'train_steps_per_second': 0.422, 'train_loss': 1.3923714351654053, 'epoch': 0.32}
+100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 200/200 [07:54<00:00,  2.37s/it]
+TrainOutput(global_step=200, training_loss=1.3923714351654053, metrics={'train_runtime': 474.3254, 'train_samples_per_second': 1.687, 'train_steps_per_second': 0.422, 'train_loss': 1.3923714351654053, 'epoch': 0.32})
+```
+
+### 3. Merge the adapter into the original model
+Using the [export_merged_model.py](https://github.com/intel-analytics/BigDL/blob/main/python/llm/example/GPU/QLoRA-FineTuning/export_merged_model.py) to merge.
+```
+python ./export_merged_model.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --adapter_path ./outputs/checkpoint-200 --output_path ./outputs/checkpoint-200-merged
+```
+
+Then you can use `./outputs/checkpoint-200-merged` as a normal huggingface transformer model to do inference.
+
+### 4. Use BigDL-LLM to verify the fine-tuning effect
+Train more steps and try input sentence like `['quote'] -> [?]` to verify. For example, using `“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->: ` to inference.
+BigDL-LLM llama2 example [link](https://github.com/intel-analytics/BigDL/tree/main/python/llm/example/CPU/HF-Transformers-AutoModels/Model/llama2). Update the `LLAMA2_PROMPT_FORMAT = "{prompt}"`.
+```bash
+python ./generate.py --repo-id-or-model-path REPO_ID_OR_MODEL_PATH --prompt "“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->:"  --n-predict 20
+```
+
+#### Sample Output
+Base_model output
+```log
+Inference time: 1.7017452716827393 s
+-------------------- Prompt --------------------
+“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->:
+-------------------- Output --------------------
+“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->: 💻 Fine-tuning a language model on a powerful device like an Intel CPU
+```
+Merged_model output
+```log
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Inference time: 2.864234209060669 s
+-------------------- Prompt --------------------
+“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->:
+-------------------- Output --------------------
+“QLoRA fine-tuning using BigDL-LLM 4bit optimizations on Intel CPU is Efficient and convenient” ->: ['bigdl'] ['deep-learning'] ['distributed-computing'] ['intel'] ['optimization'] ['training'] ['training-speed']
+```

python/llm/example/CPU/QLoRA-FineTuning/qlora_finetuning_cpu.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 os
+
+import transformers
+from transformers import LlamaTokenizer
+
+from peft import LoraConfig
+from bigdl.llm.transformers.qlora import get_peft_model, prepare_model_for_kbit_training
+from bigdl.llm.transformers import AutoModelForCausalLM
+from datasets import load_dataset
+import argparse
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Predict Tokens using `generate()` API for Llama2 model')
+    parser.add_argument('--repo-id-or-model-path', type=str, default="meta-llama/Llama-2-7b-hf",
+                        help='The huggingface repo id for the Llama2 (e.g. `meta-llama/Llama-2-7b-hf` and `meta-llama/Llama-2-13b-chat-hf`) to be downloaded'
+                             ', or the path to the huggingface checkpoint folder')
+    parser.add_argument('--dataset', type=str, default="Abirate/english_quotes")
+
+    args = parser.parse_args()
+    model_path = args.repo_id_or_model_path
+    dataset_path = args.dataset
+    tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+    data = load_dataset(dataset_path)
+    def merge(row):
+        row['prediction'] = row['quote'] + ' ->: ' + str(row['tags'])
+        return row
+    data['train'] = data['train'].map(merge)
+    data = data.map(lambda samples: tokenizer(samples["prediction"]), batched=True)
+    model = AutoModelForCausalLM.from_pretrained(model_path,
+                                                 load_in_low_bit="sym_int4",
+                                                 optimize_model=False,
+                                                 torch_dtype=torch.float16,
+                                                 modules_to_not_convert=["lm_head"], )
+    model = model.to('cpu')
+    model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=False)
+    model.enable_input_require_grads()
+    config = LoraConfig(
+        r=8,
+        lora_alpha=32,
+        target_modules=["q_proj", "k_proj", "v_proj"],
+        lora_dropout=0.05,
+        bias="none",
+        task_type="CAUSAL_LM"
+    )
+    model = get_peft_model(model, config)
+    tokenizer.pad_token_id = 0
+    tokenizer.padding_side = "left"
+    trainer = transformers.Trainer(
+        model=model,
+        train_dataset=data["train"],
+        args=transformers.TrainingArguments(
+            per_device_train_batch_size=4,
+            gradient_accumulation_steps=1,
+            warmup_steps=20,
+            max_steps=200,
+            learning_rate=2e-4,
+            save_steps=100,
+            bf16=True,
+            logging_steps=20,
+            output_dir="outputs",
+            optim="adamw_hf",  # paged_adamw_8bit is not supported yet
+            # gradient_checkpointing=True, # can further reduce memory but slower
+        ),
+        data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),
+    )
+    model.config.use_cache = False  # silence the warnings. Please re-enable for inference!
+    result = trainer.train()
+    print(result)

python/llm/example/CPU/README.md

@@ -7,6 +7,7 @@ This folder contains examples of running BigDL-LLM on Intel CPU:
 - [Native-Models](Native-Models): converting & running LLM in `llama`/`chatglm`/`bloom`/`gptneox`/`starcoder` model family using native (cpp) implementation
 - [LangChain](LangChain): running LangChain applications on BigDL-LLM
 - [Applications](Applications): running Transformers applications on BigDl-LLM
+- [QLoRA-FineTuning](QLoRA-FineTuning): running QLoRA finetuning using BigDL-LLM on intel CPUs
 
 ## System Support
 **Hardware**: