Upload hos_optimizer/cli.py with huggingface_hub
Browse files- hos_optimizer/cli.py +488 -0
hos_optimizer/cli.py
ADDED
|
@@ -0,0 +1,488 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
# -*- coding: utf-8 -*-
|
| 3 |
+
"""
|
| 4 |
+
HOS Model Optimizer - 统一命令行接口
|
| 5 |
+
|
| 6 |
+
提供统一的命令行入口,整合所有功能模块:
|
| 7 |
+
- hos-quantize: 模型量化
|
| 8 |
+
- hos-infer: 推理服务
|
| 9 |
+
- hos-train: 微调训练
|
| 10 |
+
- hos-merge: 模型合并
|
| 11 |
+
- hos-deploy: 部署服务
|
| 12 |
+
- hos-config: 配置管理
|
| 13 |
+
- hos-evaluate: 模型评测
|
| 14 |
+
"""
|
| 15 |
+
|
| 16 |
+
import click
|
| 17 |
+
import sys
|
| 18 |
+
import os
|
| 19 |
+
|
| 20 |
+
# 添加项目根目录到路径
|
| 21 |
+
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
@click.group()
|
| 25 |
+
@click.version_option(version="1.0.0", prog_name="HOS Model Optimizer")
|
| 26 |
+
def cli():
|
| 27 |
+
"""HOS Model Optimizer - 小模型优化工具集
|
| 28 |
+
|
| 29 |
+
提供模型量化、推理、训练、合并和部署的完整工具链。
|
| 30 |
+
针对 8GB VRAM 场景进行了优化。
|
| 31 |
+
|
| 32 |
+
使用示例:
|
| 33 |
+
hos-quantize --method gguf --model ./model --output ./model.gguf
|
| 34 |
+
hos-infer --model ./model --prompt "你好"
|
| 35 |
+
hos-train --model ./model --dataset ./data.json
|
| 36 |
+
hos-deploy --model ./model --port 8000
|
| 37 |
+
"""
|
| 38 |
+
pass
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
# ============================================================
|
| 42 |
+
# hos-quantize: 模型量化
|
| 43 |
+
# ============================================================
|
| 44 |
+
|
| 45 |
+
@cli.command("quantize")
|
| 46 |
+
@click.option("--method", type=click.Choice(["gguf", "awq", "gptq", "evaluate", "convert"]),
|
| 47 |
+
required=True, help="量化方法")
|
| 48 |
+
@click.option("--model", type=str, required=True, help="输入模型路径")
|
| 49 |
+
@click.option("--output", type=str, default=None, help="输出路径")
|
| 50 |
+
@click.option("--bits", type=int, default=4, help="量化位数(默认: 4)")
|
| 51 |
+
@click.option("--quant-type", type=str, default="Q4_K_M", help="GGUF 量化类型")
|
| 52 |
+
@click.option("--group-size", type=int, default=128, help="量化分组大小")
|
| 53 |
+
@click.option("--llama-cpp-path", type=str, default=None, help="llama.cpp 安装路径")
|
| 54 |
+
@click.option("--from", "from_format", type=click.Choice(["gguf", "awq", "gptq", "hf"]),
|
| 55 |
+
help="源格式(用于格式转换)")
|
| 56 |
+
@click.option("--to", "to_format", type=click.Choice(["gguf", "awq", "gptq", "hf"]),
|
| 57 |
+
help="目标格式(用于格式转换)")
|
| 58 |
+
def quantize_cmd(method, model, output, bits, quant_type, group_size, llama_cpp_path,
|
| 59 |
+
from_format, to_format):
|
| 60 |
+
"""量化模型
|
| 61 |
+
|
| 62 |
+
支持多种量化方法:GGUF、AWQ、GPTQ,以及格式转换和质量评估。
|
| 63 |
+
|
| 64 |
+
示例:
|
| 65 |
+
# GGUF 量化
|
| 66 |
+
hos-quantize --method gguf --model ./model --output ./model.gguf
|
| 67 |
+
|
| 68 |
+
# AWQ 4-bit 量化
|
| 69 |
+
hos-quantize --method awq --bits 4 --model ./model
|
| 70 |
+
|
| 71 |
+
# 评估 PPL
|
| 72 |
+
hos-quantize --method evaluate --model ./quantized-model
|
| 73 |
+
"""
|
| 74 |
+
from hos_optimizer.quantize import quantize_gguf, quantize_awq, quantize_gptq, evaluate_perplexity, convert_format
|
| 75 |
+
|
| 76 |
+
if method == "gguf":
|
| 77 |
+
quantize_gguf(model, output or f"{model}.gguf", quant_type, llama_cpp_path)
|
| 78 |
+
elif method == "awq":
|
| 79 |
+
quantize_awq(model, output or f"{model}-awq", bits, group_size)
|
| 80 |
+
elif method == "gptq":
|
| 81 |
+
quantize_gptq(model, output or f"{model}-gptq", bits, group_size)
|
| 82 |
+
elif method == "evaluate":
|
| 83 |
+
evaluate_perplexity(model)
|
| 84 |
+
elif method == "convert":
|
| 85 |
+
if not from_format or not to_format:
|
| 86 |
+
click.echo("错误: 格式转换需要指定 --from 和 --to 参数", err=True)
|
| 87 |
+
sys.exit(1)
|
| 88 |
+
convert_format(model, output or f"{model}-converted", from_format, to_format)
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
# ============================================================
|
| 92 |
+
# hos-infer: 推理服务
|
| 93 |
+
# ============================================================
|
| 94 |
+
|
| 95 |
+
@cli.command("infer")
|
| 96 |
+
@click.option("--model", type=str, required=True, help="模型路径")
|
| 97 |
+
@click.option("--backend", type=click.Choice(["llama-cpp", "vllm", "sglang"]),
|
| 98 |
+
default=None, help="推理后端(默认自动检测)")
|
| 99 |
+
@click.option("--prompt", type=str, default=None, help="单次推理的输入提示")
|
| 100 |
+
@click.option("--serve", is_flag=True, help="启动 API 服务")
|
| 101 |
+
@click.option("--chat", is_flag=True, help="交互模式")
|
| 102 |
+
@click.option("--benchmark", is_flag=True, help="性能基准测试")
|
| 103 |
+
@click.option("--host", type=str, default="0.0.0.0", help="服务监听地址")
|
| 104 |
+
@click.option("--port", type=int, default=None, help="服务端口")
|
| 105 |
+
@click.option("--max-tokens", type=int, default=256, help="最大生成 token 数")
|
| 106 |
+
@click.option("--temperature", type=float, default=0.7, help="采样温度")
|
| 107 |
+
@click.option("--top-p", type=float, default=0.9, help="nucleus sampling 参数")
|
| 108 |
+
def infer_cmd(model, backend, prompt, serve, chat, benchmark, host, port,
|
| 109 |
+
max_tokens, temperature, top_p):
|
| 110 |
+
"""推理服务
|
| 111 |
+
|
| 112 |
+
支持三种推理后端:llama-cpp、vLLM、SGLang,自动选择最优后端。
|
| 113 |
+
|
| 114 |
+
示例:
|
| 115 |
+
# 单次推理
|
| 116 |
+
hos-infer --model ./model --prompt "你好"
|
| 117 |
+
|
| 118 |
+
# 启动 API 服务
|
| 119 |
+
hos-infer --model ./model --serve --port 8000
|
| 120 |
+
|
| 121 |
+
# 交互模式
|
| 122 |
+
hos-infer --model ./model --chat
|
| 123 |
+
"""
|
| 124 |
+
from hos_optimizer.inference import UnifiedInferenceEngine, run_chat, run_benchmark, run_single_inference
|
| 125 |
+
|
| 126 |
+
# 确定默认端口
|
| 127 |
+
default_ports = {"llama-cpp": 8080, "llama_cpp": 8080, "vllm": 8000, "sglang": 30000}
|
| 128 |
+
backend_key = (backend or "auto").lower().replace("-", "_")
|
| 129 |
+
actual_port = port or default_ports.get(backend_key, 8000)
|
| 130 |
+
|
| 131 |
+
try:
|
| 132 |
+
if serve:
|
| 133 |
+
# 启动服务模式
|
| 134 |
+
from hos_optimizer.inference import UnifiedInferenceEngine
|
| 135 |
+
engine = UnifiedInferenceEngine(model_path=model, backend=backend, auto_load=True)
|
| 136 |
+
engine._backend.serve(host=host, port=actual_port)
|
| 137 |
+
else:
|
| 138 |
+
# 创建推理引擎
|
| 139 |
+
engine = UnifiedInferenceEngine(
|
| 140 |
+
model_path=model,
|
| 141 |
+
backend=backend,
|
| 142 |
+
auto_load=True,
|
| 143 |
+
)
|
| 144 |
+
|
| 145 |
+
if chat:
|
| 146 |
+
run_chat(engine)
|
| 147 |
+
elif benchmark:
|
| 148 |
+
run_benchmark(engine)
|
| 149 |
+
elif prompt:
|
| 150 |
+
run_single_inference(
|
| 151 |
+
engine,
|
| 152 |
+
prompt,
|
| 153 |
+
max_tokens=max_tokens,
|
| 154 |
+
temperature=temperature,
|
| 155 |
+
top_p=top_p,
|
| 156 |
+
)
|
| 157 |
+
else:
|
| 158 |
+
click.echo("请指定 --serve, --chat, --benchmark 或 --prompt 参数")
|
| 159 |
+
sys.exit(1)
|
| 160 |
+
except Exception as e:
|
| 161 |
+
click.echo(f"推理错误: {e}", err=True)
|
| 162 |
+
sys.exit(1)
|
| 163 |
+
|
| 164 |
+
|
| 165 |
+
# ============================================================
|
| 166 |
+
# hos-train: 微调训练
|
| 167 |
+
# ============================================================
|
| 168 |
+
|
| 169 |
+
@cli.command("train")
|
| 170 |
+
@click.option("--model", type=str, required=True, help="基础模型路径")
|
| 171 |
+
@click.option("--dataset", type=str, required=True, help="数据集路径(JSON 文件)")
|
| 172 |
+
@click.option("--output", type=str, default="./output", help="输出目录")
|
| 173 |
+
@click.option("--method", type=click.Choice(["qlora", "lora"]), default="qlora",
|
| 174 |
+
help="训练方法")
|
| 175 |
+
@click.option("--format", "dataset_format", type=click.Choice(["alpaca", "sharegpt", "messages"]),
|
| 176 |
+
default="alpaca", help="数据格式")
|
| 177 |
+
@click.option("--max-seq-length", type=int, default=2048, help="最大序列长度")
|
| 178 |
+
@click.option("--lora-rank", type=int, default=16, help="LoRA rank")
|
| 179 |
+
@click.option("--lora-alpha", type=int, default=32, help="LoRA alpha")
|
| 180 |
+
@click.option("--epochs", type=int, default=3, help="训练轮数")
|
| 181 |
+
@click.option("--batch-size", type=int, default=2, help="批次大小")
|
| 182 |
+
@click.option("--grad-accum", type=int, default=8, help="梯度累积步数")
|
| 183 |
+
@click.option("--lr", type=float, default=2e-4, help="学习率")
|
| 184 |
+
@click.option("--no-unsloth", is_flag=True, help="禁用 Unsloth 加速")
|
| 185 |
+
@click.option("--merge", is_flag=True, help="训练后自动合并模型")
|
| 186 |
+
def train_cmd(model, dataset, output, method, dataset_format, max_seq_length,
|
| 187 |
+
lora_rank, lora_alpha, epochs, batch_size, grad_accum, lr, no_unsloth, merge):
|
| 188 |
+
"""微调训练
|
| 189 |
+
|
| 190 |
+
支持 QLoRA 和 LoRA 微调,针对 8GB VRAM 场景优化。
|
| 191 |
+
|
| 192 |
+
示例:
|
| 193 |
+
# QLoRA 训练
|
| 194 |
+
hos-train --model ./model --dataset ./data.json --method qlora
|
| 195 |
+
|
| 196 |
+
# LoRA 训练并自动合并
|
| 197 |
+
hos-train --model ./model --dataset ./data.json --method lora --merge
|
| 198 |
+
"""
|
| 199 |
+
from hos_optimizer.train import TrainingConfig, train, merge_model
|
| 200 |
+
|
| 201 |
+
# 创建训练配置
|
| 202 |
+
config = TrainingConfig(
|
| 203 |
+
model_name_or_path=model,
|
| 204 |
+
dataset_path=dataset,
|
| 205 |
+
dataset_format=dataset_format,
|
| 206 |
+
max_seq_length=max_seq_length,
|
| 207 |
+
finetuning_type=method,
|
| 208 |
+
use_4bit=(method == "qlora"),
|
| 209 |
+
lora_rank=lora_rank,
|
| 210 |
+
lora_alpha=lora_alpha,
|
| 211 |
+
output_dir=output,
|
| 212 |
+
num_train_epochs=epochs,
|
| 213 |
+
per_device_train_batch_size=batch_size,
|
| 214 |
+
gradient_accumulation_steps=grad_accum,
|
| 215 |
+
learning_rate=lr,
|
| 216 |
+
use_unsloth=not no_unsloth,
|
| 217 |
+
)
|
| 218 |
+
|
| 219 |
+
try:
|
| 220 |
+
# 执行训练
|
| 221 |
+
train(config)
|
| 222 |
+
|
| 223 |
+
# 可选:自动合并
|
| 224 |
+
if merge:
|
| 225 |
+
click.echo("开始合并模型...")
|
| 226 |
+
merge_model(
|
| 227 |
+
base_model_path=model,
|
| 228 |
+
adapter_path=output,
|
| 229 |
+
output_path=f"{output}_merged",
|
| 230 |
+
)
|
| 231 |
+
click.echo("模型合并完成!")
|
| 232 |
+
except Exception as e:
|
| 233 |
+
click.echo(f"训练错误: {e}", err=True)
|
| 234 |
+
sys.exit(1)
|
| 235 |
+
|
| 236 |
+
|
| 237 |
+
# ============================================================
|
| 238 |
+
# hos-merge: 合并模型
|
| 239 |
+
# ============================================================
|
| 240 |
+
|
| 241 |
+
@cli.command("merge")
|
| 242 |
+
@click.option("--base-model", type=str, required=True, help="基础模型路径")
|
| 243 |
+
@click.option("--adapter", type=str, required=True, help="LoRA adapter 路径")
|
| 244 |
+
@click.option("--output", type=str, required=True, help="输出路径")
|
| 245 |
+
def merge_cmd(base_model, adapter, output):
|
| 246 |
+
"""合并模型
|
| 247 |
+
|
| 248 |
+
将 LoRA adapter 合并到基础模型中。
|
| 249 |
+
|
| 250 |
+
示例:
|
| 251 |
+
hos-merge --base-model ./base --adapter ./adapter --output ./merged
|
| 252 |
+
"""
|
| 253 |
+
from hos_optimizer.train import merge_model
|
| 254 |
+
|
| 255 |
+
click.echo(f"合并模型: {base_model} + {adapter} -> {output}")
|
| 256 |
+
merge_model(
|
| 257 |
+
base_model_path=base_model,
|
| 258 |
+
adapter_path=adapter,
|
| 259 |
+
output_path=output,
|
| 260 |
+
)
|
| 261 |
+
click.echo("模型合并完成!")
|
| 262 |
+
|
| 263 |
+
|
| 264 |
+
# ============================================================
|
| 265 |
+
# hos-deploy: 部署服务
|
| 266 |
+
# ============================================================
|
| 267 |
+
|
| 268 |
+
@cli.command("deploy")
|
| 269 |
+
@click.option("--model", type=str, required=True, help="模型文件路径")
|
| 270 |
+
@click.option("--model-size", type=float, default=7.0, help="模型大小(十亿参数)")
|
| 271 |
+
@click.option("--use-case", type=click.Choice(["general", "high_concurrency", "multi_turn"]),
|
| 272 |
+
default="general", help="使用场景")
|
| 273 |
+
@click.option("--host", type=str, default="0.0.0.0", help="服务主机地址")
|
| 274 |
+
@click.option("--port", type=int, default=8000, help="服务端口")
|
| 275 |
+
@click.option("--no-auto-start", is_flag=True, help="不自动启动服务")
|
| 276 |
+
def deploy_cmd(model, model_size, use_case, host, port, no_auto_start):
|
| 277 |
+
"""部署服务
|
| 278 |
+
|
| 279 |
+
自动检测硬件并选择最优配置,一键部署 API 服务。
|
| 280 |
+
|
| 281 |
+
示例:
|
| 282 |
+
# 部署 7B 模型
|
| 283 |
+
hos-deploy --model ./model.gguf --model-size 7.0
|
| 284 |
+
|
| 285 |
+
# 高并发场景
|
| 286 |
+
hos-deploy --model ./model --use-case high_concurrency
|
| 287 |
+
"""
|
| 288 |
+
from hos_optimizer.deploy import HardwareDetector, ConfigSelector, ServiceLauncher
|
| 289 |
+
|
| 290 |
+
try:
|
| 291 |
+
# 检测硬件
|
| 292 |
+
hw_info = HardwareDetector.get_hardware_info()
|
| 293 |
+
click.echo(f"检测到硬件: {hw_info.gpu_name}, {hw_info.gpu_memory_gb:.1f}GB VRAM")
|
| 294 |
+
|
| 295 |
+
# 选择配置
|
| 296 |
+
config = ConfigSelector.select_config(hw_info, model_size, use_case)
|
| 297 |
+
click.echo(f"选择配置: {config.backend.value}, {config.quantization.value}")
|
| 298 |
+
click.echo(f"推荐场景: {config.recommended_for}")
|
| 299 |
+
|
| 300 |
+
# 启动服务
|
| 301 |
+
launcher = ServiceLauncher(config, model, host, port)
|
| 302 |
+
if not no_auto_start:
|
| 303 |
+
if launcher.start():
|
| 304 |
+
click.echo(f"✓ 服务已启动,监听: {host}:{port}")
|
| 305 |
+
else:
|
| 306 |
+
click.echo("错误: 服务启动失败", err=True)
|
| 307 |
+
sys.exit(1)
|
| 308 |
+
else:
|
| 309 |
+
click.echo("服务未自动启动(--no-auto-start)")
|
| 310 |
+
except Exception as e:
|
| 311 |
+
click.echo(f"部署错误: {e}", err=True)
|
| 312 |
+
sys.exit(1)
|
| 313 |
+
|
| 314 |
+
|
| 315 |
+
# ============================================================
|
| 316 |
+
# hos-config: 配置管理
|
| 317 |
+
# ============================================================
|
| 318 |
+
|
| 319 |
+
@cli.command("config")
|
| 320 |
+
@click.option("--generate", is_flag=True, help="生成最优配置")
|
| 321 |
+
@click.option("--scenario", type=str, help="场景名称")
|
| 322 |
+
@click.option("--model-path", type=str, default="", help="模型路径")
|
| 323 |
+
@click.option("--vram", type=float, default=8.0, help="可用显存(GB)")
|
| 324 |
+
@click.option("--validate", is_flag=True, help="验证配置文件")
|
| 325 |
+
@click.option("--config", type=str, help="配置文件路径")
|
| 326 |
+
@click.option("--list-templates", is_flag=True, help="列出所有模板")
|
| 327 |
+
@click.option("--export-template", type=str, help="导出模板")
|
| 328 |
+
@click.option("--output", "-o", type=str, help="输出文件路径")
|
| 329 |
+
def config_cmd(generate, scenario, model_path, vram, validate, config,
|
| 330 |
+
list_templates, export_template, output):
|
| 331 |
+
"""配置管理
|
| 332 |
+
|
| 333 |
+
生成 8GB VRAM 最优配置,验证配置文件,管理配置模板。
|
| 334 |
+
|
| 335 |
+
示例:
|
| 336 |
+
# 生成推理配置
|
| 337 |
+
hos-config --generate --scenario inference_7b --model-path ./model
|
| 338 |
+
|
| 339 |
+
# 验证配置
|
| 340 |
+
hos-config --validate --config my_config.yaml
|
| 341 |
+
|
| 342 |
+
# 列出模板
|
| 343 |
+
hos-config --list-templates
|
| 344 |
+
"""
|
| 345 |
+
from hos_optimizer.config import ConfigManager
|
| 346 |
+
|
| 347 |
+
manager = ConfigManager()
|
| 348 |
+
|
| 349 |
+
try:
|
| 350 |
+
if list_templates:
|
| 351 |
+
templates = manager.list_templates()
|
| 352 |
+
click.echo("可用的配置模板:")
|
| 353 |
+
for name in templates:
|
| 354 |
+
click.echo(f" - {name}")
|
| 355 |
+
elif export_template:
|
| 356 |
+
if not output:
|
| 357 |
+
click.echo("错误: 导出模板需要指定 --output 参数", err=True)
|
| 358 |
+
sys.exit(1)
|
| 359 |
+
manager.export_template(export_template, output)
|
| 360 |
+
click.echo(f"模板 '{export_template}' 已导出到: {output}")
|
| 361 |
+
elif validate:
|
| 362 |
+
if not config:
|
| 363 |
+
click.echo("错误: 验证配置需要指定 --config 参数", err=True)
|
| 364 |
+
sys.exit(1)
|
| 365 |
+
cfg = manager.load_config(config)
|
| 366 |
+
issues = manager.validate_config(cfg)
|
| 367 |
+
if not issues:
|
| 368 |
+
click.echo("配置验证通过 ✓")
|
| 369 |
+
else:
|
| 370 |
+
click.echo(f"发现 {len(issues)} 个问题:")
|
| 371 |
+
for issue in issues:
|
| 372 |
+
click.echo(f" {issue}")
|
| 373 |
+
sys.exit(1)
|
| 374 |
+
elif generate:
|
| 375 |
+
if not scenario:
|
| 376 |
+
click.echo("错误: 生成配置需要指定 --scenario 参数", err=True)
|
| 377 |
+
sys.exit(1)
|
| 378 |
+
cfg = manager.generate_optimal_config(
|
| 379 |
+
scenario=scenario,
|
| 380 |
+
model_path=model_path,
|
| 381 |
+
vram_gb=vram,
|
| 382 |
+
)
|
| 383 |
+
if output:
|
| 384 |
+
manager.save_config(cfg, output)
|
| 385 |
+
click.echo(f"配置已保存到: {output}")
|
| 386 |
+
else:
|
| 387 |
+
import yaml
|
| 388 |
+
click.echo(yaml.dump(cfg, default_flow_style=False, allow_unicode=True))
|
| 389 |
+
else:
|
| 390 |
+
click.echo("请指定 --generate, --validate, --list-templates 或 --export-template")
|
| 391 |
+
sys.exit(1)
|
| 392 |
+
except Exception as e:
|
| 393 |
+
click.echo(f"配置错误: {e}", err=True)
|
| 394 |
+
sys.exit(1)
|
| 395 |
+
|
| 396 |
+
|
| 397 |
+
# ============================================================
|
| 398 |
+
# hos-evaluate: 模型评测
|
| 399 |
+
# ============================================================
|
| 400 |
+
|
| 401 |
+
@cli.command("evaluate")
|
| 402 |
+
@click.option("--model", type=str, required=True, multiple=True, help="模型路径(支持多个模型对比评测)")
|
| 403 |
+
@click.option("--dataset", type=str, required=True, help="评测数据集路径(JSON/JSONL 格式)")
|
| 404 |
+
@click.option("--metrics", type=str, multiple=True, default=["bleu", "rouge"],
|
| 405 |
+
help="评测指标(默认: bleu rouge),支持: ppl bleu rouge exact_match f1")
|
| 406 |
+
@click.option("--task", type=click.Choice(["text_generation", "perplexity"]),
|
| 407 |
+
default="text_generation", help="任务类型(默认: text_generation)")
|
| 408 |
+
@click.option("--output", "-o", type=str, default=None, help="结果输出路径(支持 .json 和 .md 格式)")
|
| 409 |
+
@click.option("--format", "dataset_format", type=click.Choice(["alpaca", "sharegpt", "messages"]),
|
| 410 |
+
default=None, help="数据集格式(默认自动检测)")
|
| 411 |
+
@click.option("--max-samples", type=int, default=None, help="最大评测样本数(默认全部)")
|
| 412 |
+
@click.option("--max-seq-length", type=int, default=512, help="最大序列长度(默认 512)")
|
| 413 |
+
@click.option("--max-new-tokens", type=int, default=256, help="最大生成 token 数(默认 256)")
|
| 414 |
+
@click.option("--temperature", type=float, default=0.7, help="采样温度(默认 0.7)")
|
| 415 |
+
@click.option("--top-p", type=float, default=0.9, help="nucleus sampling 参数(默认 0.9)")
|
| 416 |
+
@click.option("--batch-size", type=int, default=1, help="推理批大小(8GB VRAM 建议 1,默认 1)")
|
| 417 |
+
@click.option("--output-format", type=click.Choice(["json", "markdown"]),
|
| 418 |
+
default="json", help="输出格式(默认: json)")
|
| 419 |
+
@click.option("--load-in-4bit", is_flag=True, default=False,
|
| 420 |
+
help="使用 4-bit 量化加载模型(节省显存,适合 8GB VRAM 场景)")
|
| 421 |
+
@click.option("--verbose", is_flag=True, default=False, help="启用详细日志输出")
|
| 422 |
+
def evaluate_cmd(model, dataset, metrics, task, output, dataset_format, max_samples,
|
| 423 |
+
max_seq_length, max_new_tokens, temperature, top_p, batch_size,
|
| 424 |
+
output_format, load_in_4bit, verbose):
|
| 425 |
+
"""模型评测
|
| 426 |
+
|
| 427 |
+
支持多种评测指标:PPL、BLEU、ROUGE、Exact Match、F1 等。
|
| 428 |
+
支持多模型对比评测,结果可导出为 JSON 或 Markdown 格式。
|
| 429 |
+
|
| 430 |
+
示例:
|
| 431 |
+
# 单模型评测
|
| 432 |
+
hos-evaluate --model ./model --dataset ./test.json --metrics bleu rouge
|
| 433 |
+
|
| 434 |
+
# 多模型对比
|
| 435 |
+
hos-evaluate --model ./model_a --model ./model_b --dataset ./test.json --metrics bleu f1
|
| 436 |
+
|
| 437 |
+
# 指定输出格式
|
| 438 |
+
hos-evaluate --model ./model --dataset ./test.json --output result.md
|
| 439 |
+
|
| 440 |
+
# 使用 4-bit 量化加载(8GB VRAM 优化)
|
| 441 |
+
hos-evaluate --model ./model --dataset ./test.json --load-in-4bit
|
| 442 |
+
"""
|
| 443 |
+
from hos_optimizer.evaluate import evaluate_model, compare_models, EvaluationConfig
|
| 444 |
+
|
| 445 |
+
# 自动推断输出格式
|
| 446 |
+
actual_output_format = output_format
|
| 447 |
+
if output:
|
| 448 |
+
if output.endswith(".md"):
|
| 449 |
+
actual_output_format = "markdown"
|
| 450 |
+
elif output.endswith(".json"):
|
| 451 |
+
actual_output_format = "json"
|
| 452 |
+
|
| 453 |
+
# 构建评测配置
|
| 454 |
+
config = EvaluationConfig(
|
| 455 |
+
model_path=model[0] if len(model) == 1 else "",
|
| 456 |
+
dataset_path=dataset,
|
| 457 |
+
dataset_format=dataset_format,
|
| 458 |
+
max_samples=max_samples,
|
| 459 |
+
max_seq_length=max_seq_length,
|
| 460 |
+
metrics=list(metrics),
|
| 461 |
+
task_type=task,
|
| 462 |
+
max_new_tokens=max_new_tokens,
|
| 463 |
+
temperature=temperature,
|
| 464 |
+
top_p=top_p,
|
| 465 |
+
batch_size=batch_size,
|
| 466 |
+
output_format=actual_output_format,
|
| 467 |
+
output_path=output,
|
| 468 |
+
load_in_4bit=load_in_4bit,
|
| 469 |
+
verbose=verbose,
|
| 470 |
+
)
|
| 471 |
+
|
| 472 |
+
try:
|
| 473 |
+
if len(model) > 1:
|
| 474 |
+
compare_models(list(model), config)
|
| 475 |
+
else:
|
| 476 |
+
evaluate_model(config)
|
| 477 |
+
except Exception as e:
|
| 478 |
+
click.echo(f"评测错误: {e}", err=True)
|
| 479 |
+
sys.exit(1)
|
| 480 |
+
|
| 481 |
+
|
| 482 |
+
def main():
|
| 483 |
+
"""主入口函数"""
|
| 484 |
+
cli()
|
| 485 |
+
|
| 486 |
+
|
| 487 |
+
if __name__ == "__main__":
|
| 488 |
+
main()
|