MarsupialAI
commited on
Commit
•
a383725
1
Parent(s):
c3cd2ef
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Some folks are claiming there's something funky going on with GGUF quanting for Llama 3 models. I don't disagree.
|
2 |
+
|
3 |
+
Some of those people are speculating that it has something to do with converting the raw weights from bf16 to fp16 instead
|
4 |
+
of converting to fp32 as an intermediate step. I think that's bollocks. There is no logical or mathmatical justification for
|
5 |
+
how that could possibly matter.
|
6 |
+
|
7 |
+
So to test this crazy theory, I downloaded Undi95/Meta-Llama-3-8B-Instruct-hf and converted it to GGUF three ways:
|
8 |
+
- fp16 specifically with `--outtype f16`
|
9 |
+
- fp32 specifically with `--outtype f32`
|
10 |
+
- "Auto" with no outtype specified
|
11 |
+
|
12 |
+
I then quantized each of these conversions to Q4_K_M and ran perplexity tests on everything using my abbreviated wiki.short.raw
|
13 |
+
|
14 |
+
The results:
|
15 |
+
|
16 |
+
|
17 |
+
|
18 |
+
|
19 |
+
As you can see, converting to fp32 has no meaningful effect on PPL. There will no doubt be some people who will claim
|
20 |
+
"PpL iSn'T gOoD eNoUgH!!1!". For those people, I have uploaded all GGUFs used in this test. Feel free to do more extensive
|
21 |
+
testing on your own time. I consider the matter resolved until somebody can conclusively demonstrate otherwise.
|