LLaMA-Factory
支持通過 WebUI 零代碼微調大語言模型
環境準備
|
系統
|
GPU型號
|
python 版本
|
|
rocky9.2
|
NVIDIA GeForce RTX 4090
|
python3.9
|
訓練模型
github 下載地址（框架）
將 框架下載下來之後，需要配置如下：
# 安裝一個虛擬環境
conda create -n py39 python=3.9 -y
進入到虛擬環境中
conda activate py39
需要安裝的模塊如下：
# 或者手動創建包含您列出的所有包
cat > requirements.txt << 'EOF'
accelerate==1.7.0
aiofiles==23.2.1
aiohttp==3.13.0
aioprometheus==23.12.0
airportsdata==20250909
annotated-types==0.7.0
antlr4-python3-runtime==4.9.3
anyio==4.11.0
astor==0.8.1
async-timeout==5.0.1
attrs==25.4.0
audioread==3.0.1
autoawq==0.2.9
av==15.1.0
blake3==1.0.7
cachetools==6.2.1
cbor2==5.7.0
certifi==2025.10.5
cffi==2.0.0
charset-normalizer==3.4.4
click==8.1.8
cloudpickle==3.1.1
compressed-tensors==0.9.3
contourpy==1.3.0
cupy-cuda12x==13.6.0
cycler==0.12.1
datasets==3.6.0
decorator==5.2.1
Deprecated==1.2.18
depyf==0.18.0
dill==0.3.8
diskcache==5.6.3
distro==1.9.0
dnspython==2.7.0
docstring_parser==0.17.0
einops==0.8.1
email-validator==2.3.0
et_xmlfile==2.0.0
eval_type_backport==0.2.2
exceptiongroup==1.3.0
fastapi==0.119.0
fastapi-cli==0.0.13
fastapi-cloud-cli==0.3.1
fastrlock==0.8.3
ffmpy==0.6.3
filelock==3.19.1
fire==0.7.1
fonttools==4.60.1
frozendict==2.4.6
frozenlist==1.8.0
fsspec==2025.3.0
gguf==0.17.1
googleapis-common-protos==1.70.0
gradio==4.44.1
gradio_client==1.3.0
grpcio==1.75.1
h11==0.16.0
hf-xet==1.1.10
httpcore==1.0.9
httptools==0.7.1
httpx==0.28.1
huggingface-hub==0.35.3
idna==3.11
importlib_metadata==8.0.0
importlib_resources==6.5.2
interegular==0.3.3
Jinja2==3.1.6
jiter==0.11.0
joblib==1.5.2
jsonschema==4.25.1
jsonschema-specifications==2025.9.1
kiwisolver==1.4.7
lark==1.2.2
lazy_loader==0.4
librosa==0.11.0
llamafactory==0.9.3
llguidance==0.7.30
llvmlite==0.43.0
lm-format-enforcer==0.10.12
markdown-it-py==3.0.0
MarkupSafe==2.1.5
matplotlib==3.9.4
mdurl==0.1.2
mistral_common==1.8.5
mpmath==1.3.0
msgpack==1.1.2
msgspec==0.19.0
multidict==6.7.0
multiprocess==0.70.16
nest-asyncio==1.6.0
networkx==3.2.1
ninja==1.13.0
numba==0.60.0
numpy==1.26.4
nvidia-cublas-cu12==12.4.5.8
nvidia-cuda-cupti-cu12==12.4.127
nvidia-cuda-nvrtc-cu12==12.4.127
nvidia-cuda-runtime-cu12==12.4.127
nvidia-cudnn-cu12==9.1.0.70
nvidia-cufft-cu12==11.2.1.3
nvidia-cufile-cu12==1.13.1.3
nvidia-curand-cu12==10.3.5.147
nvidia-cusolver-cu12==11.6.1.9
nvidia-cusparse-cu12==12.3.1.170
nvidia-cusparselt-cu12==0.6.2
nvidia-nccl-cu12==2.21.5
nvidia-nvjitlink-cu12==12.4.127
nvidia-nvtx-cu12==12.4.127
omegaconf==2.3.0
openai==2.3.0
openai-harmony==0.0.4
opencv-python-headless==4.11.0.86
openpyxl==3.1.5
opentelemetry-api==1.26.0
opentelemetry-exporter-otlp==1.26.0
opentelemetry-exporter-otlp-proto-common==1.26.0
opentelemetry-exporter-otlp-proto-grpc==1.26.0
opentelemetry-exporter-otlp-proto-http==1.26.0
opentelemetry-proto==1.26.0
opentelemetry-sdk==1.26.0
opentelemetry-semantic-conventions==0.47b0
opentelemetry-semantic-conventions-ai==0.4.13
orjson==3.11.3
outlines==0.1.11
outlines_core==0.1.26
packaging==25.0
pandas==2.3.3
partial-json-parser==0.2.1.1.post6
peft==0.15.2
pillow==10.4.0
platformdirs==4.4.0
pooch==1.8.2
prometheus_client==0.23.1
prometheus-fastapi-instrumentator==7.1.0
propcache==0.4.1
protobuf==4.25.8
psutil==7.1.0
py-cpuinfo==9.0.0
pyarrow==21.0.0
pybase64==1.4.2
pycountry==24.6.1
pycparser==2.23
pydantic==2.10.6
pydantic_core==2.27.2
pydantic-extra-types==2.10.6
pydub==0.25.1
Pygments==2.19.2
pynvml==11.5.0
pyparsing==3.2.5
python-dateutil==2.9.0.post0
python-dotenv==1.1.1
python-json-logger==4.0.0
python-multipart==0.0.20
pytz==2025.2
PyYAML==6.0.3
pyzmq==27.1.0
quantile-python==1.1
ray==2.50.0
referencing==0.36.2
regex==2025.9.18
requests==2.32.5
rich==14.2.0
rich-toolkit==0.15.1
rignore==0.7.0
rpds-py==0.27.1
ruff==0.14.0
safetensors==0.6.2
scikit-learn==1.6.1
scipy==1.13.1
semantic-version==2.10.0
sentencepiece==0.2.1
sentry-sdk==2.41.0
setproctitle==1.3.7
setuptools==65.5.1
shellingham==1.5.4
shtab==1.7.2
six==1.17.0
sniffio==1.3.1
soundfile==0.13.1
soxr==1.0.0
sse-starlette==3.0.2
starlette==0.48.0
sympy==1.13.1
termcolor==3.1.0
threadpoolctl==3.6.0
tiktoken==0.12.0
tokenizers==0.21.1
tomlkit==0.12.0
torch==2.6.0
torchaudio==2.6.0
torchvision==0.21.0
tqdm==4.67.1
transformers==4.52.4
triton==3.2.0
trl==0.9.6
typer==0.19.2
typing_extensions==4.15.0
typing-inspection==0.4.2
tyro==0.8.14
tzdata==2025.2
urllib3==2.5.0
uvicorn==0.37.0
uvloop==0.21.0
vllm==0.8.5.post1
watchfiles==1.1.0
websockets==12.0
wheel==0.38.4
wrapt==1.17.3
xformers==0.0.29.post2
xgrammar==0.1.18
xlrd==2.0.2
xxhash==3.6.0
yarl==1.22.0
zipp==3.23.0
zstandard==0.25.0
EOF
# 使用清華源加速
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple/
# 如果遇到依賴衝突，可以嘗試
pip install -r requirements.txt --no-deps -i https://pypi.tuna.tsinghua.edu.cn/simple/
如果還有依賴報錯
# 徹底清理並讓pip重新解決所有依賴
pip uninstall opencv-python-headless datasets fsspec numpy -y
# 重新安裝，讓pip自動解決依賴版本
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple \
llamafactory \
opencv-python-headless
然後運行程序，使用第一塊顯卡
export CUDA_VISIBLE_DEVICES=0 && llamafactory-cli webui
運行起來應該是一個 7860 端口
http://192.168.3.166:7860/指定模型
注意，模型不能是量化版，因為後續需要將訓練好的模型，合併一下。（合併不支持量化版模型）
如果自定義數據集，需要修改以下兩個地方
(llamafactory-py39) [root@gpu-node-166 data]# cd LLaMA-Factory/data
vim bin.json # 新增模版文件
[
{
"instruction": "請對以下進行分類，分類選項包括：建材、裝飾、其他",
"input": "標題：關於通知\n正文：根據工作安排，現將相關事項通知如下。",
"output": "建材"
},
{
"instruction": "請對以下進行分類，分類選項包括：建材、裝飾、其他",
"input": "標題：關於通知\n正文：根據工作安排，現將相關事項通知如下，具體詳情請參閲附件。",
"output": "裝飾"
}
]
(llamafactory-py39) [root@gpu-node-166 data]# cat dataset_info.json
"adgen_local": { ## 這個名稱就是自定義數據集名稱
"file_name": "bid.json", ## 這個自定義模版文件路徑
"columns": {
"prompt": "instruction",
"query": "input",
"response": "output"
}
}
}修改好後，需要重啓LLaMA-Factory 程序。
然後開始訓練
等數據跑完之後，會在當前目錄下生成一個路徑文件，裏面存放的就是訓練好的模型
(llamafactory-py39) [root@gpu-node-166 data]# ls saves/Qwen1.5-7B-Chat/lora/
train_2025-10-17-16-12-17
(llamafactory-py39) [root@gpu-node-166 data]# ls saves/Qwen1.5-7B-Chat/lora/train_2025-10-17-16-12-17/
adapter_config.json checkpoint-100 checkpoint-300 checkpoint-800 README.md trainer_log.jsonl train_results.json
adapter_model.safetensors checkpoint-1000 checkpoint-400 checkpoint-900 running_log.txt trainer_state.json video_preprocessor_config.json
added_tokens.json checkpoint-1100 checkpoint-500 llamaboard_config.yaml special_tokens_map.json training_args.bin vocab.json
all_results.json checkpoint-1191 checkpoint-600 merges.txt tokenizer_config.json training_args.yaml
chat_template.jinja checkpoint-200 checkpoint-700 preprocessor_config.json tokenizer.json training_loss.png合併模型
# 編輯一個合併程序
(llamafactory-py39) [root@gpu-node-166 data]# vim llama3_q_lora.yaml
### Note: DO NOT use quantized model or quantization_bit when merging lora adapters
### model
# 註釋： 這個是本地存放原始模型路徑
model_name_or_path: /data/qwen1.5/Qwen1.5-7B-Chat
# 這是訓練後的模型存放路徑
adapter_name_or_path: saves/Qwen1.5-7B-Chat/lora/train_2025-10-17-16-12-17/
# 這個模型屬於 千問
template: qwen
finetuning_type: lora
### export
# 合併後的模型存放路徑
export_dir: models/qwen2.5-7b-chat-lora-merged
export_size: 2
export_device: cpu
export_legacy_format: false
#export_legacy_format: true
# 運行合併程序
(llamafactory-py39) [root@gpu-node-166 data]# llamafactory-cli export llama3_q_lora.yaml
# 合併後，程序存放在
(llamafactory-py39) [root@gpu-node-166 data]# ls models/qwen2.5-7b-chat-lora-merged
added_tokens.json merges.txt model-00004-of-00008.safetensors model-00008-of-00008.safetensors tokenizer_config.json
chat_template.jinja model-00001-of-00008.safetensors model-00005-of-00008.safetensors Modelfile tokenizer.json
config.json model-00002-of-00008.safetensors model-00006-of-00008.safetensors model.safetensors.index.json vocab.json
generation_config.json model-00003-of-00008.safetensors model-00007-of-00008.safetensors special_tokens_map.json然後使用合併後的模型跑數據。
vim batch_classify-2.py
import pandas as pd
import time
import logging
import os
from tqdm import tqdm
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
# --------------------------
# 基礎配置
# --------------------------
INPUT_EXCEL = "/data/qwen1.5/LLaMA-Factory/data/測試數據.xlsx"
OUTPUT_EXCEL = "/data/qwen1.5/LLaMA-Factory/data/分類結果.xlsx"
SHEET_NAME = "Sheet1"
DATA_COLUMN = "文本"
# --------------------------
# 模型配置
# --------------------------
MODEL_PATH = "/data/qwen1.5/lx/LLaMA-Factory/data/models/qwen2.5-7b-chat-lora-merged"
# --------------------------
# 分類規則
# --------------------------
BIDDING_CLASSIFICATION_RULES = """
你是專業的分類器，必須嚴格遵守以下規則：
1. 分類範圍：僅從以下3個固定類別中選擇，絕對不能輸出其他類別；
2. 類別列表（按優先級排序）：
- 建材（最高優先級）
- 裝飾（最低優先級）
- 其他（僅當完全不符合以上2類時使用）
3. 輸出要求：只返回1個類別名稱，不附加任何解釋、標點或多餘文字。
"""
# --------------------------
# 系統配置
# --------------------------
DELAY_BETWEEN_REQUESTS = 0.1
MAX_RESPONSE_TOKENS = 10
MAX_INPUT_LENGTH = 1500
# 日誌配置
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s',
handlers=[logging.StreamHandler()]
)
# --------------------------
# 模型加載
# --------------------------
logging.info("正在加載模型...")
try:
tokenizer = AutoTokenizer.from_pretrained(
MODEL_PATH,
trust_remote_code=True,
local_files_only=True
)
model = AutoModelForCausalLM.from_pretrained(
MODEL_PATH,
device_map="auto",
torch_dtype=torch.float16,
trust_remote_code=True,
low_cpu_mem_usage=True,
local_files_only=True
)
logging.info("模型加載成功！")
except Exception as e:
logging.error(f"模型加載失敗: {e}")
exit(1)
# --------------------------
# 分類核心函數（修復版）
# --------------------------
def classify_text(text):
text = str(text).strip()
if len(text) > MAX_INPUT_LENGTH:
text = text[:MAX_INPUT_LENGTH] + "..."
logging.debug(f"文本過長已截斷：{text[:50]}...")
messages = [
{"role": "system", "content": BIDDING_CLASSIFICATION_RULES},
{"role": "user", "content": f"請對以下分類：{text}"}
]
try:
# 修復：正確處理 apply_chat_template 的返回值
formatted_prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
# 確保 formatted_prompt 是字符串
if isinstance(formatted_prompt, list):
formatted_prompt = formatted_prompt[0] if formatted_prompt else ""
elif not isinstance(formatted_prompt, str):
formatted_prompt = str(formatted_prompt)
inputs = tokenizer(
formatted_prompt,
return_tensors="pt",
max_length=2048,
truncation=True
).to(model.device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=MAX_RESPONSE_TOKENS,
temperature=0.0,
do_sample=False,
pad_token_id=tokenizer.eos_token_id,
eos_token_id=tokenizer.eos_token_id
)
response = tokenizer.decode(
outputs[0][len(inputs["input_ids"][0]):],
skip_special_tokens=True
).strip()
valid_classes = ["建材","裝飾", "其他"]
# 精確匹配類別
for cls in valid_classes:
if cls in response:
return cls
return "其他"
except Exception as e:
logging.error(f"處理失敗：{str(e)}")
return "其他"
# --------------------------
# 主流程
# --------------------------
def main():
if not os.path.exists(INPUT_EXCEL):
logging.error(f"找不到測試數據，請檢查路徑：{INPUT_EXCEL}")
return
try:
logging.info("開始讀取Excel文件...")
# 嘗試多種引擎
try:
df = pd.read_excel(INPUT_EXCEL, sheet_name=SHEET_NAME, engine='openpyxl')
except:
try:
df = pd.read_excel(INPUT_EXCEL, sheet_name=SHEET_NAME, engine='xlrd')
except:
# 如果都不行，嘗試自動選擇
df = pd.read_excel(INPUT_EXCEL, sheet_name=SHEET_NAME)
logging.info(f"成功讀取Excel，共{len(df)}行數據")
except Exception as e:
logging.error(f"讀取Excel失敗：{str(e)}")
return
if DATA_COLUMN not in df.columns:
logging.error(f"Excel中沒有'{DATA_COLUMN}'列，可用列：{df.columns.tolist()}")
return
df[DATA_COLUMN] = df[DATA_COLUMN].fillna("(空值)").astype(str)
texts = df[DATA_COLUMN].tolist()
results = []
# 測試模型
logging.info("測試模型分類能力...")
test_texts = [
"本項目因故廢除",
"XX公司XX公告",
"XX時間記錄表"
]
for test in test_texts:
result = classify_text(test)
logging.info(f"測試 '{test}' -> '{result}'")
logging.info(f"開始正式分類，共{len(texts)}條數據...")
# 創建臨時目錄
temp_dir = "/data/qwen1.5/LLaMA-Factory/data/tmp"
if not os.path.exists(temp_dir):
os.makedirs(temp_dir)
temp_excel_path = f"{temp_dir}/中間結果_彙總.xlsx"
# 批量處理
for i, text in enumerate(tqdm(texts, desc="分類進度")):
result = classify_text(text)
results.append(result)
# 每20條保存一次中間結果
if (i + 1) % 20 == 0:
temp_df = df.iloc[:i+1].copy()
temp_df['分類結果'] = results[:i+1]
try:
temp_df.to_excel(temp_excel_path, index=False, engine='openpyxl')
except:
temp_df.to_excel(temp_excel_path, index=False)
logging.info(f"已更新中間結果（累計處理{i+1}條）")
time.sleep(DELAY_BETWEEN_REQUESTS)
# 保存最終結果
df['分類結果'] = results
try:
df.to_excel(OUTPUT_EXCEL, index=False, engine='openpyxl')
logging.info(f"完成！結果保存至：{OUTPUT_EXCEL}")
# 統計分類結果
from collections import Counter
counter = Counter(results)
logging.info(f"分類統計: {dict(counter)}")
except Exception as e:
backup = "/data/qwen1.5/lx/LLaMA-Factory/data/分類結果_備份.csv"
df.to_csv(backup, index=False, encoding='utf-8-sig')
logging.error(f"保存Excel失敗，已備份至CSV：{backup}，錯誤：{str(e)}")
if __name__ == "__main__":
main()(qwen-inference) [root@gpu-node-166 data]# python batch_classify-2.py
2025-10-17 17:48:02,878 - INFO - 正在加載模型...
`torch_dtype` is deprecated! Use `dtype` instead!
2025-10-17 17:48:03,927 - INFO - We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
Loading checkpoint shards: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 8/8 [00:02<00:00, 3.61it/s]
2025-10-17 17:48:06,273 - WARNING - Some parameters are on the meta device because they were offloaded to the cpu.
2025-10-17 17:48:06,274 - INFO - 模型加載成功！
2025-10-17 17:48:06,274 - INFO - 開始讀取Excel文件...
2025-10-17 17:48:06,400 - INFO - 成功讀取Excel，共1588行數據
2025-10-17 17:48:06,401 - INFO - 測試模型分類能力...
The following generation flags are not valid and may be ignored: ['temperature', 'top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
2025-10-17 17:48:10,905 - INFO - 測試 '本項目因故廢除' -> 'XX公告'
2025-10-17 17:48:13,945 - INFO - 測試 'XX公司XX公告' -> 'XX公告'
2025-10-17 17:48:17,996 - INFO - 測試 'XX時間記錄表' -> 'XX記錄'
2025-10-17 17:48:17,996 - INFO - 開始正式分類，共1588條數據...
分類進度: 0%| |
分類進度: 0%| | 1/1588
分類進度: 0%|▏ |
分類進度: 1%|█▏ | 14/1588 [00:42<1:27:47, 3.35s/it]
分類進度: 1%|█▎ | 15/1588 [00:46<1:26:38, 3..30s/it]
分類進度: 1%|█▋ | 19/1588 [00:57<1:19:27, 3.04s/it]
2025-10-17 17:49:18,801 - INFO - 已更新中間結果（累計處理20條）
分類進度: 2%|███▍ | 39/1588 [01:57<1:16:59, 2.98s/it]
本文章為轉載內容，我們尊重原作者對文章享有的著作權。如有內容錯誤或侵權問題，歡迎原作者聯繫我們進行內容更正或刪除文章。