#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#define MAX_TOKENS 4096
#define KV_CACHE_SIZE 32768

typedef struct {
    int device_type;
    void* compute_handle;
    size_t kv_cache_pos;
    float kv_cache[KV_CACHE_SIZE];
} InferenceContext;

int ds4_init(InferenceContext* ctx, int device) {
    ctx->device_type = device;
    ctx->kv_cache_pos = 0;
    switch(device) {
        case DEVICE_METAL:
            ctx->compute_handle = metal_init();
            break;
        case DEVICE_CUDA:
            ctx->compute_handle = cuda_init();
            break;
        case DEVICE_ROCM:
            ctx->compute_handle = rocm_init();
            break;
        default:
            return DS4_ERR_DEVICE;
    }
    return DS4_OK;
}

int ds4_forward(InferenceContext* ctx, int* input_tokens, int len, int* output_tokens) {
    int generated = 0;
    for (int i = 0; i < len; i++) {
        attention_compute(ctx, input_tokens[i]);
        update_kv_cache(ctx, input_tokens[i]);
    }
    while (generated < MAX_TOKENS) {
        int next_token = logits_sample(ctx);
        if (next_token == EOS_TOKEN) break;
        output_tokens[generated] = next_token;
        attention_compute(ctx, next_token);
        update_kv_cache(ctx, next_token);
        generated++;
    }
    return generated;
}

int main(int argc, char** argv) {
    InferenceContext ctx;
    if (ds4_init(&ctx, DEVICE_METAL) != DS4_OK) {
        fprintf(stderr, "init failed
");
        return 1;
    }
    int input[] = {1, 234, 5678, 90};
    int output[MAX_TOKENS];
    int n = ds4_forward(&ctx, input, 4, output);
    printf("generated %d tokens
", n);
    return 0;
}

# 克隆仓库并编译
git clone https://github.com/antirez/ds4.git
cd ds4
make -j$(nproc)

# 下载量化模型（INT4版本约4.2GB）
./ds4-download --model deepseek-v4-flash-int4

# 启动交互式推理
./ds4 --model deepseek-v4-flash-int4 --device metal --temp 0.7

# 以服务模式启动（兼容OpenAI API格式）
./ds4 --model deepseek-v4-flash-int4 --serve --port 8080

# 安装airllm
pip install airllm

# Python中使用airllm加载70B模型

from airllm import AutoModel

model = AutoModel.from_pretrained(
    "deepseek-ai/deepseek-v4-flash",
    compression_mode="int4",
    gpu_memory_limit_gb=4,
    disk_cache_path="/data/airllm_cache"
)

input_text = "解释量子纠缠现象"
tokens = model.tokenize(input_text)
output = model.generate(tokens, max_new_tokens=512)
print(model.detokenize(output))

from openclaw import Assistant, SkillRegistry

registry = SkillRegistry()
assistant = Assistant(
    model="deepseek-v4-flash",
    skill_registry=registry,
    auto_extend=True
)

# 请求一个当前未安装的技能
response = assistant.handle(
    user_id="user_001",
    message="帮我查询北京到上海的高铁票价并列成表格"
)

# 系统日志显示自动扩展过程
# [openclaw] 技能未找到: train_ticket_query
# [openclaw] 正在自动生成技能代码...
# [openclaw] 技能 train_ticket_query 已生成并注册
# [openclaw] 正在执行技能...

print(response.text)

# 验证新技能已持久化
skills = registry.list_skills()
for skill in skills:
    print("已注册技能:", skill.name, "版本:", skill.version)

from tencent_agent_memory import MemoryCenter, Agent

memory = MemoryCenter(
    storage_backend="postgresql",
    connection="postgresql://user:pass@localhost/agent_memory",
    embedding_model="bge-large-zh-v2"
)

agent_a = Agent("数据分析师", memory)
agent_b = Agent("报告撰写人", memory)

# Agent A 存储分析结论
agent_a.remember(
    content="2026年Q2 AI芯片市场同比增长47%",
    metadata={"source": "市场调研", "confidence": 0.95}
)

# Agent B 检索Agent A的知识
results = agent_b.recall(
    query="AI芯片市场增长数据",
    top_k=3
)
for item in results:
    print("记忆:", item.content, "置信度:", item.score)

from agent_reach import InternetAgent

agent = InternetAgent(
    model="deepseek-v4-flash",
    headless=True,
    safe_mode=True
)

# 让Agent自主搜索并总结信息
task = "查找2026年8月最新的开源大模型发布情况，整理成列表"
result = agent.execute(task)

for item in result.structured_data:
    print("模型:", item["name"])
    print("发布方:", item["organization"])
    print("参数量:", item["parameters"])
    print("---")

# 安装reverse-skill
pip install reverse-skill

# 运行提示注入测试套件
reverse-skill run --target http://localhost:8080/v1 \
    --suite prompt-injection \
    --vectors all \
    --report html

from reverse_skill import PromptInjectionSuite, Target

target = Target(
    endpoint="http://localhost:8080/v1/chat",
    model="deepseek-v4-flash",
    api_key="sk-test"
)

suite = PromptInjectionSuite()
results = suite.run(target, vectors="all")

print("测试向量总数:", results.total)
print("成功攻击数:", results.successes)
print("防护成功率:", str(round(results.defense_rate * 100, 1)) + "%")

for item in results.failed_attacks:
    print("成功攻击向量:", item.vector_name)
    print("攻击载荷:", item.payload[:80])