add llm-finetune example: baichuan2

example/LLM/llama/README.md

@@ -24,7 +24,6 @@ Current supports datasets:
 - For finetune:
   - openassistant
 
-
 ## Build Starwhale Runtime
 
 ```bash

example/llm-finetune/README.md

@@ -0,0 +1,40 @@
+LLM Finetune
+======
+
+LLM finetune is a state-of-art task for large language model.
+
+In these examples, we will use Starwhale to finetune a set of LLM base models, evaluate and release models. The demos are in the [starwhale/llm-finetuning](https://cloud.starwhale.cn/projects/401/overview) project of Starwhale Cloud.
+
+What we learn
+------
+
+- use the `@starwhale.finetune` decorator to define a finetune handler for Starwhale Model to finish the LLM finetune.
+- use the `@starwhale.evaluation.predict` to define a model evaluation for LLM.
+- use the `@starwhale.handler` to define a web handler for LLM online evaluation.
+- use one Starwhale Runtime to run all models.
+- build Starwhale Dataset by the one-line command from the Huggingface, no code.
+
+Models
+------
+
+- [Baichuan2](https://github.com/baichuan-inc/Baichuan2): Baichuan 2 is the new generation of open-source large language models launched by Baichuan Intelligent Technology. It was trained on a high-quality corpus with 2.6 trillion tokens.
+- [ChatGLM3](https://github.com/THUDM/ChatGLM3): ChatGLM3 is a new generation of pre-trained dialogue models jointly released by Zhipu AI and Tsinghua KEG. ChatGLM3-6B is the open-source model in the ChatGLM3 series.
+
+Datasets
+------
+
+- [Belle multiturn chat](https://huggingface.co/datasets/BelleGroup/multiturn_chat_0.8M): The dataset includes approx. 0.8M Chinese multiturn dialogs between human and assistant from BELLE Group.
+
+    ```bash
+    # build the origin dataset from huggingface
+    swcli dataset build -hf BelleGroup/multiturn_chat_0.8M --name belle-multiturn-chat
+
+    # build the random 10k items by baichuan2
+    swcli dataset build --json https://raw.githubusercontent.com/baichuan-inc/Baichuan2/main/fine-tune/data/belle_chat_ramdon_10k.json --name belle_chat_random_10k
+    ```
+
+- [COIG](https://huggingface.co/datasets/BAAI/COIG): The Chinese Open Instruction Generalist (COIG) project is a harmless, helpful, and diverse set of Chinese instruction corpora.
+
+    ```bash
+    swcli dataset build -hf BAAI/COIG --name coig
+    ```

example/llm-finetune/models/baichuan2/.gitignore

@@ -0,0 +1,2 @@
+pretrain/
+.cache/

example/llm-finetune/models/baichuan2/.swignore

@@ -0,0 +1 @@
+.cache/

example/llm-finetune/models/baichuan2/README.md

@@ -0,0 +1,57 @@
+Baichuan2 Finetune with Starwhale
+======
+
+- 🍬 Parameters: 7b
+- 🔆 Github: https://github.com/baichuan-inc/Baichuan2
+- 🥦 Author: Baichuan Inc.
+- 📝 License: baichuan
+- 🐱 Starwhale Example: https://github.com/star-whale/starwhale/tree/main/example/llm-finetune/models/baichuan2
+- 🌽 Introduction: Baichuan 2 is the new generation of large-scale open-source language models launched by Baichuan Intelligence inc..It is trained on a high-quality corpus with 2.6 trillion tokens and has achieved the best performance in authoritative Chinese and English benchmarks of the same size.Baichuan2-7b-chat is chat model of Baichuan 2, which contains 7 billion parameters.
+
+In this example, we will use Baichuan2-7b-chat as the base model to finetune and evaluate.
+
+- Evaluate baichuan2-7b-chat model.
+- Provide baichuan2-7b-chat multi-turn chat online evaluation.
+- Fine-tune baichuan2-7b-chat model with belle-multiturn-chat dataset.
+- Evaluate the fine-tuned model.
+- Provide the fine-tuned model multi-turn chat online evaluation.
+- Fine-tune fine-tuned baichuan2-7n-chat model.
+
+Because of 4bit quantization technical, the single T4/A10/A100 gpu card is ok for evaluation and finetune.
+
+Build Starwhale Model
+------
+
+```bash
+python3 build.py
+```
+
+Run Online Evaluation in the Standalone instance
+------
+
+```bash
+# for source code
+swcli model run -w . -m evaluation --handler evaluation:chatbot
+
+# for model package with runtime
+swcli model run --uri baichuan2-7b-chat --handler evaluation:chatbot --runtime llm-finetune
+```
+
+Run Starwhale Model for evaluation in the Standalone instance
+------
+
+```bash
+swcli dataset cp https://cloud.starwhale.cn/projects/401/datasets/161/versions/223/ .
+swcli -vvv model run -w . -m evaluation --handler evaluation:copilot_predict --dataset z-bench-common --dataset-head 3
+```
+
+Finetune base model
+------
+
+```bash
+# build finetune dataset from baichuan2
+swcli dataset build --json https://raw.githubusercontent.com/baichuan-inc/Baichuan2/main/fine-tune/data/belle_chat_ramdon_10k.json --name belle_chat_random_10k
+
+swcli -vvv model run -w . -m finetune --dataset belle_chat_random_10k --handler finetune:lora_finetune
+```
+

example/llm-finetune/models/baichuan2/build.py

@@ -0,0 +1,32 @@
+from huggingface_hub import snapshot_download
+
+import starwhale
+
+try:
+    from .utils import BASE_MODEL_DIR
+    from .finetune import lora_finetune
+    from .evaluation import chatbot, copilot_predict
+except ImportError:
+    from utils import BASE_MODEL_DIR
+    from finetune import lora_finetune
+    from evaluation import chatbot, copilot_predict
+
+starwhale.init_logger(3)
+
+
+def build_starwhale_model() -> None:
+    BASE_MODEL_DIR.mkdir(parents=True, exist_ok=True)
+
+    snapshot_download(
+        repo_id="baichuan-inc/Baichuan2-7B-Chat",
+        local_dir=BASE_MODEL_DIR,
+    )
+
+    starwhale.model.build(
+        name="baichuan2-7b-chat",
+        modules=[copilot_predict, chatbot, lora_finetune],
+    )
+
+
+if __name__ == "__main__":
+    build_starwhale_model()

example/llm-finetune/models/baichuan2/evaluation.py

@@ -0,0 +1,138 @@
+from __future__ import annotations
+
+import os
+import typing as t
+
+import torch
+import gradio
+from peft import PeftModel
+from transformers import AutoTokenizer, BitsAndBytesConfig, AutoModelForCausalLM
+from transformers.generation.utils import GenerationConfig
+
+from starwhale import handler, evaluation
+
+try:
+    from .utils import BASE_MODEL_DIR, ADAPTER_MODEL_DIR
+except ImportError:
+    from utils import BASE_MODEL_DIR, ADAPTER_MODEL_DIR
+
+_g_model = None
+_g_tokenizer = None
+
+
+def _load_model_and_tokenizer() -> t.Tuple:
+    global _g_model, _g_tokenizer
+
+    if _g_model is None:
+        print(f"load model from {BASE_MODEL_DIR} ...")
+        _g_model = AutoModelForCausalLM.from_pretrained(
+            BASE_MODEL_DIR,
+            device_map="auto",
+            torch_dtype=torch.float16,
+            trust_remote_code=True,
+            load_in_4bit=True,  # for lower gpu memory usage
+            quantization_config=BitsAndBytesConfig(
+                load_in_4bit=True,
+                llm_int8_threshold=6.0,
+                llm_int8_has_fp16_weight=False,
+                bnb_4bit_compute_dtype=torch.float16,
+                bnb_4bit_use_double_quant=True,
+                bnb_4bit_quant_type="nf4",
+            ),
+        )
+        _g_model.generation_config = GenerationConfig.from_pretrained(BASE_MODEL_DIR)
+
+        if (ADAPTER_MODEL_DIR / "adapter_config.json").exists():
+            print(f"load adapter from {ADAPTER_MODEL_DIR} ...")
+            _g_model = PeftModel.from_pretrained(
+                _g_model, str(ADAPTER_MODEL_DIR), is_trainable=False
+            )
+
+    if _g_tokenizer is None:
+        print(f"load tokenizer from {BASE_MODEL_DIR} ...")
+        _g_tokenizer = AutoTokenizer.from_pretrained(
+            BASE_MODEL_DIR, use_fast=False, trust_remote_code=True
+        )
+
+    return _g_model, _g_tokenizer
+
+
+@evaluation.predict(
+    resources={"nvidia.com/gpu": 1},
+    replicas=1,
+    log_mode="plain",
+)
+def copilot_predict(data: dict) -> str:
+    model, tokenizer = _load_model_and_tokenizer()
+    # support z-bench-common dataset: https://cloud.starwhale.cn/projects/401/datasets/161/versions/223/files
+    messages = [{"role": "user", "content": data["prompt"]}]
+
+    config_dict = model.generation_config.to_dict()
+    # TODO: use arguments
+    config_dict.update(
+        max_new_tokens=int(os.environ.get("MAX_MODEL_LENGTH", 512)),
+        do_sample=True,
+        temperature=float(os.environ.get("TEMPERATURE", 0.7)),
+        top_p=float(os.environ.get("TOP_P", 0.9)),
+        top_k=int(os.environ.get("TOP_K", 30)),
+        repetition_penalty=float(os.environ.get("REPETITION_PENALTY", 1.3)),
+    )
+    return model.chat(
+        tokenizer,
+        messages=messages,
+        generation_config=GenerationConfig.from_dict(config_dict),
+    )
+
+
+@handler(expose=17860)
+def chatbot() -> None:
+    with gradio.Blocks() as server:
+        chatbot = gradio.Chatbot(height=800)
+        msg = gradio.Textbox(label="chat", show_label=True)
+        _max_gen_len = gradio.Slider(
+            0, 1024, value=256, step=1.0, label="Max Gen Len", interactive=True
+        )
+        _top_p = gradio.Slider(
+            0, 1, value=0.7, step=0.01, label="Top P", interactive=True
+        )
+        _temperature = gradio.Slider(
+            0, 1, value=0.95, step=0.01, label="Temperature", interactive=True
+        )
+        gradio.ClearButton([msg, chatbot])
+
+        def response(
+            from_user: str,
+            chat_history: t.List,
+            max_gen_len: int,
+            top_p: float,
+            temperature: float,
+        ) -> t.Tuple[str, t.List]:
+            dialog = []
+            for _user, _assistant in chat_history:
+                dialog.append({"role": "user", "content": _user})
+                if _assistant:
+                    dialog.append({"role": "assistant", "content": _assistant})
+            dialog.append({"role": "user", "content": from_user})
+
+            model, tokenizer = _load_model_and_tokenizer()
+            from_assistant = model.chat(
+                tokenizer,
+                messages=dialog,
+                generation_config=GenerationConfig(
+                    max_new_tokens=max_gen_len,
+                    do_sample=True,
+                    temperature=temperature,
+                    top_p=top_p,
+                ),
+            )
+
+            chat_history.append((from_user, from_assistant))
+            return "", chat_history
+
+        msg.submit(
+            response,
+            [msg, chatbot, _max_gen_len, _top_p, _temperature],
+            [msg, chatbot],
+        )
+
+    server.launch(server_name="0.0.0.0", server_port=17860, share=True)