// #include <bits/allocator.h>
// #pragma GCC optimize("Ofast,unroll-loops")
// #pragma GCC target("avx2")
#include "peru.h"
#include <bits/stdc++.h>
using namespace std;
const int N=2.5e6+10;
long long f[N], a[N];
const int mod=1e9+7;
int32_t solve(int32_t n, int32_t k, int32_t *aa){
priority_queue<pair<long long, int>, vector<pair<long long, int>>, greater<pair<long long, int>>> pq;
vector<int> st;
vector<deque<pair<int, long long>>> dq;
for (int i=1; i<=n; ++i) a[i]=aa[i-1];
long long ans=0;
int qid=0;
for (int i=1; i<=n; ++i){
deque<pair<int, long long>> cur;
int id=st.size();
while (id && a[st[id-1]]<=a[i]) --id;
int cnt=0;
for (int j=id; j<(int)st.size(); ++j){
++cnt;
int x=a[st[j]];
while (dq[j].size()){
while (cur.size() && cur.back().second>=dq[j].front().second+a[i]-x){
cur.pop_back();
}
cur.emplace_back(dq[j].front().first, dq[j].front().second+a[i]-x);
dq[j].pop_front();
}
}
while (cnt--) dq.pop_back(), st.pop_back();
pair<int, long long> pp={i, a[i]+f[i-1]};
while (cur.size() && cur.back().second>=pp.second){
cur.pop_back();
}
cur.push_back(pp);
qid=min(qid, (int)dq.size());
dq.push_back(cur);
st.push_back(i);
pq.emplace(cur.front().second, (int)dq.size()-1);
while (qid<(int)dq.size()){
if (dq[qid].size() && dq[qid].front().first<i-k+1){
while (dq[qid].size() && dq[qid].front().first<i-k+1) dq[qid].pop_front();
}
if (dq[qid].size()){
break;
}
++qid;
}
while (1){
auto t=pq.top();
if (t.second>=(int)dq.size() || dq[t.second].empty()) pq.pop();
else if (dq[t.second].front().second!=t.first){
pq.pop();
pq.emplace(dq[t.second].front().second, t.second);
}else break;
}
f[i]=pq.top().first;
// if (i<n) seg.update_pos(1, 1, n, i+1, f[i]);
ans=(ans*23+f[i])%mod;
}
return ans;
}
