1
# Qwen 2 7B 架构参数
2
qwen2_7b_config = {
3
    "vocab_size": 151936,           # 大词表，覆盖多语言
4
    "hidden_size": 3584,            # 隐藏维度
5
    "intermediate_size": 18944,     # FFN 中间维度（约 5.3x）
6
    "num_hidden_layers": 28,        # Transformer 层数
7
    "num_attention_heads": 28,      # 查询头数
8
    "num_key_value_heads": 4,       # GQA 的 KV 头数
9
    "max_position_embeddings": 131072, # 128K 上下文
10
    "rope_theta": 1000000.0,        # RoPE 基频
11
    "activation": "SwiGLU",         # 激活函数
12
    "norm": "RMSNorm",              # 归一化方式
13
    "tie_embeddings": False,        # 输入输出 embedding 不共享
14
}

1
# Qwen 2.5 训练数据组成（推测）
2
training_data = {
3
    "中文网页与书籍": "约 25%",
4
    "英文网页与书籍": "约 30%",
5
    "代码": "约 20%（GitHub + 编程教材）",
6
    "数学与科学": "约 10%（arXiv + 数学教材）",
7
    "其他语言": "约 10%",
8
    "多模态对齐数据": "约 5%",
9
}
10

11
total_training_tokens = "~18T tokens（Qwen 2.5 推测）"

1
# InternLM 2 7B 架构参数
2
internlm2_7b_config = {
3
    "vocab_size": 103168,            # 中英文优化词表
4
    "hidden_size": 4096,             # 隐藏维度
5
    "intermediate_size": 11008,      # FFN 中间维度
6
    "num_hidden_layers": 32,         # Transformer 层数
7
    "num_attention_heads": 32,       # 查询头数
8
    "num_key_value_heads": 8,        # GQA 的 KV 头数
9
    "max_position_embeddings": 262144, # 256K 位置编码
10
    "rope_theta": 500000.0,          # 高基频 RoPE
11
    "activation": "SwiGLU",          # 激活函数
12
}

1
# InternLM 2 长上下文推理配置
2
inference_config = {
3
    "max_length": 200000,
4
    "flash_attention": True,       # 必须开启
5
    "kv_cache_dtype": "fp8",       # 压缩 KV Cache
6
    "chunk_prefill": True,         # 分块预填充
7
    "rope_scaling": {
8
        "type": "dynamic",         # 动态 RoPE 缩放
9
        "factor": 2.0,
10
    }
11
}

1
# InternLM 2.5 工具调用示例
2
tools = [
3
    {
4
        "name": "search_web",
5
        "description": "搜索互联网获取最新信息",
6
        "parameters": {
7
            "query": {"type": "string", "description": "搜索关键词"}
8
        }
9
    },
10
    {
11
        "name": "run_python",
12
        "description": "执行 Python 代码",
13
        "parameters": {
14
            "code": {"type": "string", "description": "Python 代码"}
15
        }
16
    }
17
]
18

19
# InternLM 2.5 会自动判断何时调用工具
20
response = model.chat("今天上海天气怎么样？", tools=tools)
21
# 输出: <tool_call: search_web({"query": "上海今天天气"})>

1
# 使用 transformers 加载 Qwen 2.5
2
from transformers import AutoModelForCausalLM, AutoTokenizer
3

4
model_name = "Qwen/Qwen2.5-7B-Instruct"
5
tokenizer = AutoTokenizer.from_pretrained(model_name)
6
model = AutoModelForCausalLM.from_pretrained(
7
    model_name,
8
    torch_dtype="auto",
9
    device_map="auto",
10
    attn_implementation="flash_attention_2",  # 推荐 Flash Attention 2
11
)
12

13
messages = [
14
    {"role": "system", "content": "你是一个有帮助的助手。"},
15
    {"role": "user", "content": "用 Python 实现快速排序"},
16
]
17
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
18
inputs = tokenizer(text, return_tensors="pt").to(model.device)
19
outputs = model.generate(**inputs, max_new_tokens=512)
20
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

1
# 使用 LMDeploy 部署 InternLM 2.5
2
# pip install lmdeploy
3

4
from lmdeploy import pipeline, PytorchEngineConfig
5

6
# 配置推理引擎
7
engine_config = PytorchEngineConfig(
8
    model_name="internlm/internlm2_5-7b-chat",
9
    tp=1,                    # Tensor Parallel
10
    session_len=32768,       # 会话长度
11
    cache_max_entry_count=0.8, # KV Cache 占比
12
)
13

14
pipe = pipeline("internlm/internlm2_5-7b-chat", backend_config=engine_config)
15
response = pipe(["请解释什么是强化学习"])
16
print(response[0].text)