root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 19:50:17 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

Added perplexity metrics for llama 3.1 with different quantization settings

Browse Source
This commit is contained in:
fedric95 2024-08-08 10:55:33 +02:00 committed by GitHub
parent ebd541a570
commit 924c832461
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 23 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
examples/perplexity/README.md
View File

@ -169,6 +169,29 @@ Results were calculated with LLaMA 3 8b BF16 as `--kl-divergence-base` and LLaMA
| RMS Δp | 0.150 ± 0.001 % | | RMS Δp | 0.150 ± 0.001 % |
| Same top p | 99.739 ± 0.013 % | | Same top p | 99.739 ± 0.013 % |
## LLaMA 3.1 BF16 Scoreboard
| Revision | b3472 |
|:---------|:-------------------|
| Backend | CUDA |
| CPU | AMD Epyc 7R13 |
| GPU | 1x NVIDIA L4 |
| Quantization | imatrix | PPL |
|--------------|---------|---------------------|
| bf16 | None | 6.4006 +/- 0.03938 |
| fp16 | None | 6.4016 +/- 0.03939 |
| q8_0 | None | 6.4070 +/- 0.03941 |
| q6_K | None | 6.4231 +/- 0.03957 |
| q5_K_M | None | 6.4623 +/- 0.03987 |
| q5_K_S | None | 6.5161 +/- 0.04028 |
| q4_K_M | None | 6.5837 +/- 0.04068 |
| q4_K_S | None | 6.6751 +/- 0.04125 |
| q3_K_L | None | 6.9458 +/- 0.04329 |
| q3_K_M | None | 7.0488 +/- 0.04384 |
| q3_K_S | None | 7.8823 +/- 0.04920 |
| q2_K | None | 9.7262 +/- 0.06393 |
## Old Numbers ## Old Numbers
<details> <details>