大模型实践--QLoRA代码实践(基于Unsloth)

import os
import torch
import unsloth
from datasets import load_dataset
from trl import SFTConfig, SFTTrainer

# 配置预训练模型下载地址
os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com/'
os.environ['TRITON_CACHE_DIR'] = 'C:\\triton_cache'

# Configure model and tokenizer
model_name = "Qwen/Qwen3-0.6B"
max_length = 1024  # Supports automatic RoPE Scaling, so choose any number
model, tokenizer = unsloth.FastLanguageModel.from_pretrained(
    model_name=model_name,
    max_seq_length=max_length,
    dtype=torch.bfloat16,  # For auto-detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
    load_in_4bit=True,  # Use 4bit quantization to reduce memory usage. Can be False
)

tokenizer = unsloth.chat_templates.get_chat_template(
    tokenizer,
    chat_template="qwen3",  # change this to the right chat_template name
)

# r: rank dimension for LoRA update matrices (smaller = more compression)
rank_dimension = 4
# lora_alpha: scaling factor for LoRA layers (higher = stronger adaptation)
lora_alpha = 8
# lora_dropout: dropout probability for LoRA layers (helps prevent overfitting)
lora_dropout = 0.05

# Do model patching and add fast LoRA weights
model = unsloth.FastLanguageModel.get_peft_model(
    model,
    r=rank_dimension,
    target_modules=[
        "q_proj",
        "k_proj",
        "v_proj",
        "o_proj",
        "gate_proj",
        "up_proj",
        "down_proj",
    ],
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,  # Dropout = 0 is currently optimized
    bias="none",  # Bias = "none" is currently optimized
    use_gradient_checkpointing=True,
    random_state=3407,
)

# 数据集
# Load dataset
dataset_dict = load_dataset("json",
                            data_files={"train": "hf/model/Qwen3-0.6B/peft_sample/data/keywords_data_test.jsonl",
                                        "test": "hf/model/Qwen3-0.6B/peft_sample/data/keywords_data_train.jsonl"})
def map_func(message):
    conversation = message['conversation']
    messages = []
    for item in conversation:
        messages.append({"role": "user", "content": item['human']})
        messages.append({"role": "assistant", "content": item['assistant']})
    return {"messages": messages}
# 将数据转化为对话式的 ShareGPT 格式，并移除不需要的字段
dataset_dict = dataset_dict.map(map_func, batched=False,
                                 remove_columns=['conversation_id', 'category', 'conversation', 'dataset'])

def formatting_prompts_func(messages):
    conversation = messages["messages"]
    texts = [tokenizer.apply_chat_template(convo, tokenize=False, add_generation_prompt=False) for convo in
             conversation]
    return {"text": texts}
# 应用Chat Template
dataset_dict = dataset_dict.map(formatting_prompts_func, batched=True, remove_columns=['messages'])

training_args = SFTConfig(
    output_dir="hf/model/Qwen3-0.6B/unsloth_sample/model1",
    max_steps=1000,
    per_device_train_batch_size=4,
    learning_rate=5e-5,
    save_steps=50,
    save_total_limit=2,
    eval_strategy="steps",
    eval_steps=50,
    load_best_model_at_end=True,
    logging_dir="hf/model/Qwen3-0.6B/peft_sample/logs/",
    logging_strategy='steps',
    logging_steps=50,
    bf16=True,
    warmup_steps=50,
    assistant_only_loss=False,
)

# Initialize trainer
trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset_dict["train"],
    eval_dataset=dataset_dict["test"].train_test_split(test_size=0.005, shuffle=True)['test'],
    processing_class=tokenizer,
)

# Start training
trainer.train()

# 保存模型
trainer.save_model("hf/model/Qwen3-0.6B/unsloth_sample/model1/final_model1/")

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen3-0.6B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    dtype="auto",
    device_map="auto"
)
# prepare the model input
prompt = "识别出文本中的关键词：\n题目：基于产权理论的智慧城市信息共享机制研究\n摘要：[目的/意义]新型智慧城市的建设是实现国家治理体系和治理能力现代化的关键手段,然而目前普遍存在信息孤岛问题严重影响新型智慧城市落地实现,如何合理高效解决信息资源共享问题,已经成为急需解决的问题。[方法/过程]通过分析智慧城市信息共享存在的问题原因以及解决思路,以信息产权及信息确权分析为基础,对智慧城市信息共享的利益冲突与平衡进行分析,最后提出一种基于产权理论的智慧城市信息共享机制框架。[结果/结论]新型智慧城市应充分考虑信息价值及信息产权问题,形成信息共享的正向激励机制,使智慧城市信息资源得以高效利用,实现政府各部门相互协作,适应复杂多变的城市环境需求,形成智慧政府能力。"
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)

import torch
from transformers import AutoModelForCausalLM,AutoTokenizer
import os
from peft import PeftModel

# 配置预训练模型下载地址
os.environ['HF_ENDPOINT']='https://hf-mirror.com/'

# Configure model and tokenizer
model_name = "Qwen/Qwen3-0.6B"

# 1. Load the base model
base_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")

# 2. Load the Unsloth model with adapter
peft_model = PeftModel.from_pretrained(
    base_model, "hf/model/Qwen3-0.6B/unsloth_sample/model1/final_model1/",device_map="auto")

# 3. Merge adapter weights with base model
merged_model = peft_model.merge_and_unload()

# 4. prepare the model input
prompt = "识别出文本中的关键词：\n题目：基于产权理论的智慧城市信息共享机制研究\n摘要：[目的/意义]新型智慧城市的建设是实现国家治理体系和治理能力现代化的关键手段,然而目前普遍存在信息孤岛问题严重影响新型智慧城市落地实现,如何合理高效解决信息资源共享问题,已经成为急需解决的问题。[方法/过程]通过分析智慧城市信息共享存在的问题原因以及解决思路,以信息产权及信息确权分析为基础,对智慧城市信息共享的利益冲突与平衡进行分析,最后提出一种基于产权理论的智慧城市信息共享机制框架。[结果/结论]新型智慧城市应充分考虑信息价值及信息产权问题,形成信息共享的正向激励机制,使智慧城市信息资源得以高效利用,实现政府各部门相互协作,适应复杂多变的城市环境需求,形成智慧政府能力。"
messages = [
    {"role": "user", "content": prompt}
]

# 5. load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(merged_model.device)

# 6. conduct text completion
generated_ids = merged_model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

# 7. parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)