hyunsik

vLLM Model Conversion (Pooling Model Adapter) 분석

기준 커밋: 83319b44c

개요

vLLM은 --convert <type> 옵션을 통해 기존 Text Generation 모델(예: *ForCausalLM)을 Pooling 모델(Embedding, Classification, Reward)로 변환하는 기능을 제공합니다. 이는 Adapter 패턴을 사용하여 원본 모델 코드 수정 없이 동적으로 클래스를 확장합니다.

---

hyunsik

{{- bos_token }}
{%- if custom_tools is defined %}
    {%- set tools = custom_tools %}
{%- endif %}
{%- if not tools_in_user_message is defined %}
    {%- set tools_in_user_message = true %}
{%- endif %}
{%- if not date_string is defined %}
    {%- set date_string = "26 Jul 2024" %}
{%- endif %}

hyunsik

int sel_le_long_long_col_col(int vec_num, int [] res_sel_vec, 
                             long [] col1, long [] col2, int [] sel_vec) {  
  int ret = 0;
  if (sel_vec == null) {
    for (int i = 0; i < vec_num; i++) {
      if (col1[i] < col2[i]) {
        res_sel_vec[ret++] = i
      }
    }
  } else {

hyunsik

map_add_long_int_col_col(int vec_num, long [] result, long [] col1, 
                         int [] col2, int [] sel_vec) {  
  if (sel_vec == null) { 
     for (int i = 0; i = 0; i < vec_num; i++) {
     	result[i] = col1[i] + col2[i];
     }
  } else {
    int sel_idx;
    for (int i = 0; I = 0; i < vec_num; i++) {
     	sel_idx = sel_vec[i];

hyunsik

for (i = 0; i < 128; i+=4) {
  a[i] = b[i] + c[i];
  a[i+1] = b[i+1] + c[i+1];
  a[i+2] = b[i+2] + c[i+2];
  a[i+3] = b[i+3] + c[i+3];
}

hyunsik

for (i = 0; i < 128; i++) {
  a[i] = b[i] + c[i];
}

hyunsik

int uselect_bt_void_int_bat_int_const(oid *output, int *input, int value, int size) { 
  oid i; 
  int j = 0; 
  for (i = 0; i < size; i++) {
    if (input[i] > value) {
      output[j++] = i; 
    }
  }
  return j; 
}

hyunsik

select
    l_returnflag,  l_linestatus, sum(l_quantity) as sum_qty,
    sum(l_extendedprice) as sum_base_price,
    sum(l_extendedprice*(1-l_discount)) as sum_disc_price,
    sum(l_extendedprice*(1-l_discount)*(1+l_tax)) as sum_charge,
  avg(l_quantity) as avg_qty,
    avg(l_extendedprice) as avg_price, avg(l_discount) as avg_disc,
  count(*) as count_order
from
    lineitem

hyunsik

Latency Comparison Numbers
--------------------------
execute typical instruction                  1   ns   1/1,000,000,000 sec                   
L1 cache reference                           0.5 ns
Branch mispredict                            5   ns
L2 cache reference                           7   ns   14x L1 cache
Mutex lock/unlock                           25   ns
Main memory reference                      100   ns   20x L2 cache, 200x L1 cache
Compress 1K bytes with Zippy             3,000   ns        
Send 1K bytes over 1 Gbps network       10,000   ns       

hyunsik

public abstract class ColumnVector {
   public boolean[] isNull;
   public boolean noNulls;
   public boolean isRepeating;
   ...
}

public class LongColumnVector extends ColumnVector {
  public long[] vector;
  public static final long NULL_VALUE = 1;