llama-2-70b memory usage

zhm-algo commented 2 months ago

use real model, not fake model the test cmd is: "bash run_benchmark.sh -m llama-2-70b -d w8a8 -s 2 -bs 1 -in 2048 -out 2048 -kvd int8 -i 5"

use “numactl -H” to observe hbm memory usage will find during the inference, there is no hbm memory left, and also many ddr memory are used.

[root@spr28 benchmark]# numactl -H available: 4 nodes (0-3) node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 node 0 size: 128482 MB node 0 free: 80533 MB node 1 cpus: 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 node 1 size: 128982 MB node 1 free: 71701 MB node 2 cpus: node 2 size: 65536 MB node 2 free: 27 MB node 3 cpus: node 3 size: 65536 MB node 3 free: 15 MB node distances: node 0 1 2 3 0: 10 21 13 23 1: 21 10 23 13 2: 13 23 10 23 3: 23 13 23 10

pujiang2018 commented 2 months ago

@xiangzez could you pls to take a look?

xiangzez commented 1 month ago

Hi @zhm-algo, this issue is due to model file cache taking up hbm memory, causing some memory allocated on ddr. Here are two ways to use hbm memory only and avoid allocate on ddr memory:

Check Linux system parameter zone_reclaim_mode
```
cat /proc/sys/vm/zone_reclaim_mode
```
If the value is 0, change it to 2
```
echo 2 > /proc/sys/vm/zone_reclaim_mode
```
Or use numactl -m to bind hbm memory instead of numactl -p in run script.

intel / xFasterTransformer

llama-2-70b memory usage #403