推理系统、KV Cache 与服务性能边界

import java.util.*;
public class Main {
    static void near(double expected,double actual) {
        if (!Double.isFinite(actual) || Math.abs(expected-actual)>1e-9*Math.max(1,Math.abs(expected)))
            throw new AssertionError(expected+" != "+actual);
    }
    static void check(boolean ok) { if(!ok) throw new AssertionError(); }
    static void bad(Runnable r) {
        try { r.run(); } catch(IllegalArgumentException e) { return; }
        throw new AssertionError("invalid input accepted");
    }
static final class Page {
        final int[] tokens;int used=0,refs=1;
        Page(int size){tokens=new int[size];}
    }
    static final class Cache {
        final int blockSize,capacity;int nextId=0;
        final Map<Integer,Page> pages=new HashMap<>();
        final Map<String,List<Integer>> chains=new HashMap<>();
        Cache(int blockSize,int capacity) {
            if(blockSize<1||capacity<1) throw new IllegalArgumentException("capacity");
            this.blockSize=blockSize;this.capacity=capacity;
        }
        void create(String name) {
            if(name==null||name.isBlank()||chains.containsKey(name)) throw new IllegalArgumentException("name");
            chains.put(name,new ArrayList<>());
        }
        List<Integer> get(String name) {
            List<Integer> chain=chains.get(name);if(chain==null) throw new IllegalArgumentException("unknown sequence");return chain;
        }
        int allocate(Page copy) {
            if(pages.size()>=capacity||nextId==Integer.MAX_VALUE) throw new IllegalArgumentException("capacity exhausted"); // 配额先检查，失败不修改引用关系。
            Page page=new Page(blockSize);
            if(copy!=null) {System.arraycopy(copy.tokens,0,page.tokens,0,copy.used);page.used=copy.used;}
            int id=nextId++;pages.put(id,page);return id;
        }
        void append(String name,int token) {
            List<Integer> chain=get(name);Page last=chain.isEmpty()?null:pages.get(chain.get(chain.size()-1));
            if(last==null||last.used==blockSize) {int id=allocate(null);chain.add(id);last=pages.get(id);}
            else if(last.refs>1) {
                int id=allocate(last);last.refs--;chain.set(chain.size()-1,id);last=pages.get(id); // 共享的未满尾页在写入前复制。
            }
            last.tokens[last.used++]=token;
        }
        void fork(String from,String to) {
            List<Integer> copy=new ArrayList<>(get(from));create(to);
            for(int id:copy) pages.get(id).refs++;chains.put(to,copy); // 分支共享页，页引用次数按序列映射计数。
        }
        int[] read(String name) {
            List<Integer> chain=get(name);int size=0;for(int id:chain)size+=pages.get(id).used;
            int[] out=new int[size];int at=0;
            for(int id:chain){Page p=pages.get(id);System.arraycopy(p.tokens,0,out,at,p.used);at+=p.used;}return out;
        }
        void release(String name) {
            List<Integer> chain=get(name);chains.remove(name);
            for(int id:chain) if(--pages.get(id).refs==0) pages.remove(id); // 最后一位引用者离开后才释放物理页。
        }
    }
    public static void main(String[] args) {
Cache c=new Cache(2,4);c.create("a");c.append("a",1);c.append("a",2);
        check(c.pages.size()==1&&Arrays.equals(c.read("a"),new int[]{1,2}));
        c.append("a",3);check(c.pages.size()==2&&Arrays.equals(c.read("a"),new int[]{1,2,3}));
        c.fork("a","b");c.append("b",4);check(c.pages.size()==3&&Arrays.equals(c.read("a"),new int[]{1,2,3})&&Arrays.equals(c.read("b"),new int[]{1,2,3,4}));
        c.release("b");check(c.pages.size()==2);c.release("a");check(c.pages.isEmpty());
        Cache full=new Cache(2,2);full.create("a");full.append("a",1);full.append("a",2);full.fork("a","b");full.append("b",3);
        check(full.pages.size()==2&&Arrays.equals(full.read("a"),new int[]{1,2}));
        Cache tight=new Cache(2,1);tight.create("a");tight.append("a",1);tight.fork("a","b");bad(()->tight.append("b",2));
        check(tight.pages.size()==1&&Arrays.equals(tight.read("b"),new int[]{1}));tight.release("b");tight.append("a",2);check(tight.read("a").length==2);
        bad(()->new Cache(0,1));
        bad(()->c.release("a"));
        System.out.println("PASS 8 cases");
    }
}