shawwn / llama-dl

High-speed download of LLaMA, Facebook's 65B parameter GPT model

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

why are the 13B, 30B and 65B models size dont match the original files?

Coderx7 opened this issue · comments

Hi, I noticed the 13B model is 12GB only. did you somehow mistakenly used the 7b model in 13b model?
I also noticed the 30B and 65B models are 15GB only.
the weights should look like this apparently :

model original size quantized size (4-bit)
7B 13 GB 3.9 GB
13B 24 GB 7.8 GB
30B 60 GB 19.5 GB
65B 120 GB 38.5 GB

OK, this is answered here : #4
basically starting from 13B, the download is multi-parts.

"%file: tokenizer_checklist.chk"
eeec4125e9c7560836b4873b6f8e3025  tokenizer.model

"%file: 7B_checklist.chk"
6efc8dab194ab59e49cd24be5574d85e  consolidated.00.pth
7596560e011154b90eb51a1b15739763  params.json

"%file: 13B_checklist.chk"
0804c42ca65584f50234a86d71e6916a  consolidated.00.pth
016017be6040da87604f77703b92f2bc  consolidated.01.pth
0171286599e0f8858893e49fb74b71f8  params.json

"%file: 30B_checklist.chk"
f856e9d99c30855d6ead4d00cc3a5573  consolidated.00.pth
d9dbfbea61309dc1e087f5081e98331a  consolidated.01.pth
2b2bed47912ceb828c0a37aac4b99073  consolidated.02.pth
ea0405cdb5bc638fee12de614f729ebc  consolidated.03.pth
4babdbd05b8923226a9e9622492054b6  params.json

"%file: 65B_checklist.chk"
9deae67e2e7b5ccfb2c738f390c00854  consolidated.00.pth
0c4b00c30460c3818bd184ee949079ee  consolidated.01.pth
847194df776dd38f8ae9ddcede8829a1  consolidated.02.pth
3b6c8adcb5654fd36abab3206b46a0f1  consolidated.03.pth
68d61d1242597ad92616ec31b8cb6b4c  consolidated.04.pth
7f71259eaee2b906aa405d8edf39925f  consolidated.05.pth
0574e26b6891ab2cb0df7340d773fe9b  consolidated.06.pth
e5d9790df955270b836aec79462ead22  consolidated.07.pth
0306cf93a1c2c3891166fe1fd782c2b7  params.json