can not reproduce $6000

[taishan1994]

Here's my machine configuration
AMD双路服务器:
CPU: 2 X AMD EPYC 9115/2.60GHz/64M/16C/32T/165W；
内存: 24 X 32GB/DDR5/5600MHz/REG（消费级），总共768G；
SSD: １X SSD/3.84T/SATA 6Gb/2.5寸/读取型；
硬盘: １ X ４TB/SATA/7200RPM/3.5寸/企业级；
阵列卡: 1 X LSI 9560-8I 4G/支持RAID 0,1,5,6,10,50,60；
网卡: 1 X 双口/千兆电口/RJ45接口/I350-T2；
板载2个10G万兆电口 Broadcom® BCM57416；
电源：1200W冗余电源1+1；

When I tested with a sample used by $6000, the TPS was only 3.0, while $6000 reported 5.4.

llama-cli -m model_path --temp 0.6 -no-cnv -c 16384 -p "<丨User丨>How many Rs are there in strawberry?<丨Assistant丨>"

[ejrydhfs]

Are you running llama.cpp on a virtual machine by any chance? I noticed you were using llama.cpp as root and with some tabs at the top of the image. Virtualization could be the reason why the TPS is lower.

[ejrydhfs]

Also is this 6000? https://openi.pcl.ac.cn/6000/aiforge

[taishan1994]

Also is this 6000? https://openi.pcl.ac.cn/6000/aiforge

The above screenshot was taken inside the container. Later, we will directly conduct tests on the host. On average, the decoding speed of Q8_0 is approximately 3.3 tokens per second. We are referring to this place (or this content) here. https://x.com/carrigmat/status/1884244369907278106

[ejrydhfs]

Please forgive me if you have done this already but did you do this?

And that's your system! Put it all together and throw Linux on it. Also, an important tip: Go into the BIOS and set the number of NUMA groups to 0. This will ensure that every layer of the model is interleaved across all RAM chips, doubling our throughput. Don't forget!

Sounds like this problem is rather common. If you still experience issues I would look at the cpu temperatures to see if they are overheating or something. Maybe there's still a sticker between the heatsink and a CPU.

[taishan1994]

I have aligned it. I need to set NUMA to NPS0 and set the memory to be interactive.