GLM 系列模型选择指南：智谱的模型矩阵

标准 Decoder-only:
  Token → Token → Token → Token（全因果注意力）

Prefix-LM (GLM):
  [Prefix 部分: 双向注意力] → [生成部分: 因果注意力]
  用户输入的 prompt 可以双向交互理解，再自回归生成答案

# GLM-4V 架构示意
class GLM4V(nn.Module):
    def __init__(self, language_model, vision_encoder):
        self.lm = language_model          # GLM-4 语言基座
        self.vision_encoder = vision_encoder  # ViT-H/14 视觉编码器
        self.adapter = nn.Sequential(
            nn.Linear(vision_dim, lm_dim),
            nn.GELU(),
            nn.Linear(lm_dim, lm_dim)
        )

    def forward(self, images, text):
        # 1. 视觉特征提取
        visual_features = self.vision_encoder(images)
        # 2. 对齐到语言空间
        visual_tokens = self.adapter(visual_features)
        # 3. 与文本 token 拼接
        input_embeds = torch.cat([visual_tokens, text_embeds], dim=1)
        # 4. 语言模型统一处理
        return self.lm(inputs_embeds=input_embeds)

# CodeGeeX-4 FIM 补全示例
from openai import OpenAI

client = OpenAI(
    api_key="your-codegeex-api-key",
    base_url="https://open.bigmodel.cn/api/paas/v4"
)

# 前缀+后缀的补全模式
response = client.completions.create(
    model="codegeex-4",
    prompt="<|prefix|>def binary_search(arr, target):\n    left, right = 0, len(arr) - 1\n<|suffix|>\n    return -1<|middle|>",
    max_tokens=200,
    temperature=0.2
)
print(response.choices[0].text)
# 输出: while left <= right:\n    mid = (left + right) // 2\n    ...

from openai import OpenAI

client = OpenAI(
    api_key="your-api-key",
    base_url="https://open.bigmodel.cn/api/paas/v4"
)

# GLM-4-Plus 对话
response = client.chat.completions.create(
    model="glm-4-plus",
    messages=[
        {"role": "system", "content": "你是一个数据分析助手"},
        {"role": "user", "content": "分析2024年中国新能源汽车市场趋势"}
    ],
    temperature=0.7,
    max_tokens=4096,
    tools=[{
        "type": "web_search",
        "web_search": {"enable": True, "search_result": True}
    }]
)

# GLM-4V 图像理解
response = client.chat.completions.create(
    model="glm-4v-plus",
    messages=[{
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}},
            {"type": "text", "text": "解析这张数据图表的内容"}
        ]
    }]
)

# GLM-Zero 推理
response = client.chat.completions.create(
    model="glm-zero",
    messages=[{"role": "user", "content": "一个水池有进水管和出水管...求水池注满时间"}]
)

# 使用 vLLM 部署 GLM-4-9B-Chat
python -m vllm.entrypoints.openai.api_server \
  --model THUDM/glm-4-9b-chat \
  --tensor-parallel-size 1 \
  --gpu-memory-utilization 0.9 \
  --max-model-len 131072 \
  --trust-remote-code \
  --port 8000