llm-fine-tuning

大模型微调基础设施

高效训练和微调开源大模型——从单 GPU 上的 LoRA，到多节点集群上的分布式全量微调。

何时使用此技能

在以下情况下使用此技能：

• 使用领域专用数据（法律、医疗、代码、支持服务）微调大模型
• 使用 QLoRA 在消费级 GPU 上微调 70B 模型
• 使用 DeepSpeed 或 FSDP 设置分布式训练
• 导出微调后的适配器，以便在生产环境中提供服务
• 实施 RLHF、DPO 或指令微调工作流

前提条件

• 配备 24GB 以上 VRAM 的 NVIDIA GPU（RTX 4090 / A100 / H100）
• CUDA 12.1+，且 nvidia-smi 可正常运行
• 安装了 pip 的 Python 3.10+
• Hugging Face 账户，以及用于访问受限模型的 HF_TOKEN
• 500GB 以上磁盘空间，用于存放模型权重和训练数据

快速入门：QLoRA 微调

pip install transformers datasets trl peft bitsandbytes accelerate

python - <<'EOF'
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer, SFTConfig
import torch

model_id = "meta-llama/Llama-3.1-8B-Instruct"

# 4-bit quantization (QLoRA)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(
    model_id, quantization_config=bnb_config, device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# LoRA configuration
peft_config = LoraConfig(
    r=16,                    # rank
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
)

dataset = load_dataset("your-org/your-dataset", split="train")

trainer = SFTTrainer(
    model=model,
    args=SFTConfig(
        output_dir="./output",
        num_train_epochs=3,
        per_device_train_batch_size=2,
        gradient_accumulation_steps=8,
        learning_rate=2e-4,
        bf16=True,
        logging_steps=10,
        save_strategy="epoch",
        report_to="wandb",
    ),
    train_dataset=dataset,
    peft_config=peft_config,
    processing_class=tokenizer,
)
trainer.train()
trainer.save_model("./fine-tuned-model")
EOF

Axolotl（生产级微调框架）

# config.yaml — Axolotl QLoRA config for Llama 3.1
base_model: meta-llama/Llama-3.1-8B-Instruct
model_type: LlamaForCausalLM
tokenizer_type: PreTrainedTokenizerFast

load_in_4bit: true
adapter: qlora
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj

datasets:
  - path: your-org/your-dataset
    type: alpaca              # or sharegpt, chat_template, etc.

dataset_prepared_path: ./prepared-data
val_set_size: 0.05
output_dir: ./output

sequence_len: 4096
sample_packing: true         # pack multiple short samples for efficiency

micro_batch_size: 2
gradient_accumulation_steps: 8
num_epochs: 3
learning_rate: 2e-4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
warmup_ratio: 0.05

bf16: true
flash_attention: true

logging_steps: 10
eval_steps: 100
save_steps: 200
wandb_project: my-fine-tune

# Run with Axolotl
pip install axolotl[flash-attn,deepspeed]
accelerate launch -m axolotl.cli.train config.yaml

使用 DeepSpeed 进行分布式训练

// deepspeed_zero3.json — ZeRO Stage 3 (split optimizer + gradients + params)
{
  "zero_optimization": {
    "stage": 3,
    "offload_optimizer": {"device": "cpu", "pin_memory": true},
    "offload_param": {"device": "cpu", "pin_memory": true},
    "overlap_comm": true,
    "contiguous_gradients": true,
    "sub_group_size": 1e9,
    "reduce_bucket_size": "auto",
    "stage3_prefetch_bucket_size": "auto",
    "stage3_param_persistence_threshold": "auto",
    "stage3_max_live_parameters": 1e9,
    "stage3_max_reuse_distance": 1e9,
    "gather_16bit_weights_on_model_save": true
  },
  "bf16": {"enabled": true},
  "gradient_clipping": 1.0,
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto"
}

# Launch 4-GPU DeepSpeed training
deepspeed --num_gpus=4 train.py \
  --deepspeed deepspeed_zero3.json \
  --model_name meta-llama/Llama-3.1-70B-Instruct \
  --output_dir ./output

DPO / RLHF 对齐

from trl import DPOTrainer, DPOConfig
from datasets import load_dataset

# Dataset format: {"prompt": ..., "chosen": ..., "rejected": ...}
dataset = load_dataset("your-org/preference-data")

trainer = DPOTrainer(
    model=model,
    ref_model=None,           # None = implicit reference with peft
    args=DPOConfig(
        output_dir="./dpo-output",
        beta=0.1,             # KL divergence weight
        num_train_epochs=1,
        per_device_train_batch_size=1,
        gradient_accumulation_steps=16,
        learning_rate=5e-7,
        bf16=True,
    ),
    train_dataset=dataset["train"],
    peft_config=peft_config,
    processing_class=tokenizer,
)
trainer.train()

合并 LoRA 适配器以供部署

from peft import PeftModel
from transformers import AutoModelForCausalLM

# Load base model in full precision
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="cpu",
)

# Load and merge LoRA adapter
model = PeftModel.from_pretrained(base_model, "./fine-tuned-model")
merged_model = model.merge_and_unload()

# Save merged model (ready for vLLM serving)
merged_model.save_pretrained("./merged-model", safe_serialization=True)
tokenizer.save_pretrained("./merged-model")

# Push to Hugging Face Hub
merged_model.push_to_hub("your-org/your-fine-tuned-model")

Kubernetes 训练任务

apiVersion: batch/v1
kind: Job
metadata:
  name: llm-fine-tune
spec:
  template:
    spec:
      restartPolicy: OnFailure
      nodeSelector:
        nvidia.com/gpu.product: A100-SXM4-80GB
      containers:
      - name: trainer
        image: nvcr.io/nvidia/pytorch:24.05-py3
        command: ["accelerate", "launch", "-m", "axolotl.cli.train", "/config/config.yaml"]
        resources:
          limits:
            nvidia.com/gpu: "4"
            memory: "320Gi"
          requests:
            nvidia.com/gpu: "4"
        volumeMounts:
        - name: config
          mountPath: /config
        - name: model-cache
          mountPath: /root/.cache/huggingface
        - name: output
          mountPath: /output
        env:
        - name: HUGGING_FACE_HUB_TOKEN
          valueFrom:
            secretKeyRef:
              name: hf-token
              key: token
        - name: WANDB_API_KEY
          valueFrom:
            secretKeyRef:
              name: wandb-token
              key: key
      volumes:
      - name: config
        configMap:
          name: axolotl-config
      - name: model-cache
        persistentVolumeClaim:
          claimName: model-cache-pvc
      - name: output
        persistentVolumeClaim:
          claimName: training-output-pvc

常见问题

问题	原因	解决方法
CUDA out of memory	批次过大	减小 micro_batch_size；增大 gradient_accumulation_steps
训练损失为 NaN	学习率过高	将 LR 降至 1e-4 或 5e-5；添加预热
训练缓慢	未启用 Flash Attention	安装 flash-attn；启用 flash_attention: true
微调效果不佳	数据格式错误	验证数据集格式；检查与 sample_packing 的兼容性
合并适配器时出错	量化方式混用	在 CPU 上以 bf16 合并，而非使用 4-bit

最佳实践

• 使用 Flash Attention 2——速度可提升 2–4 倍，且占用更少内存。
• 通过 W&B 或 MLflow 监控训练损失和评估损失；过拟合时，增加 dropout 或减少数据量。
• 使用留出的评估集（5–10%）进行验证；使用 MMLU 或自定义评估作为质量门槛。
• 先从 LoRA r=16 开始，再逐步提高——秩越高，参数越多，但收益会递减。
• 在 Axolotl 中使用 sample_packing，最大限度提高 GPU 对短序列的利用率。

相关技能

• vllm-server - 提供微调后模型的服务
• gpu-server-management - GPU 设置
• llm-inference-scaling - 大规模部署
• ai-pipeline-orchestration - 训练工作流