MarsupialAI
commited on
Commit
•
9a9cf2e
1
Parent(s):
6376837
Update README.md
Browse files
README.md
CHANGED
@@ -1,32 +1,3 @@
|
|
1 |
-
# Initial Testing 2024-04-25
|
2 |
-
|
3 |
-
Some folks are claiming there's something funky going on with GGUF quanting for Llama 3 models. I don't disagree.
|
4 |
-
|
5 |
-
Some of those people are speculating that it has something to do with converting the raw weights from bf16 to fp16 instead
|
6 |
-
of converting to fp32 as an intermediate step. I think that's bollocks. There is no logical or mathmatical justification for
|
7 |
-
how that could possibly matter.
|
8 |
-
|
9 |
-
So to test this crazy theory, I downloaded Undi95/Meta-Llama-3-8B-Instruct-hf and converted it to GGUF three ways:
|
10 |
-
- fp16 specifically with `--outtype f16`
|
11 |
-
- fp32 specifically with `--outtype f32`
|
12 |
-
- "Auto" with no outtype specified
|
13 |
-
|
14 |
-
I then quantized each of these conversions to Q4_K_M and ran perplexity tests on everything using my abbreviated wiki.short.raw
|
15 |
-
text file. The results:
|
16 |
-
|
17 |
-
````
|
18 |
-
FP16 specified: size 14.9GB PPL @ fp16 9.5158 +/- 0.15418 PPL @ Q4km 9.6414 +/- 0.15494
|
19 |
-
FP32 specified: size 29.9GB PPL @ fp32 9.5158 +/- 0.15418 PPL @ Q4km 9.6278 +/- 0.15466
|
20 |
-
None specified: size 29.9GB PPL @ ???? 9.5158 +/- 0.15418 PPL @ Q4km 9.6278 +/- 0.15466
|
21 |
-
````
|
22 |
-
|
23 |
-
|
24 |
-
As you can see, converting to fp32 has no meaningful effect on PPL compared to converting to fp16. PPL is identical at full weight,
|
25 |
-
and the miniscule loss shown at Q4km is will within the margin of error. There will no doubt be some people who will claim
|
26 |
-
"PpL iSn'T gOoD eNoUgH!!1!". For those people, I have uploaded all GGUFs used in this test. Feel free to use those files to do
|
27 |
-
more extensive testing on your own time. I consider the matter resolved until somebody can conclusively demonstrate otherwise.
|
28 |
-
|
29 |
-
|
30 |
# Continued Experiments 2024-05-11
|
31 |
|
32 |
As an imatrix enjoyer, it has been bugging me whether the precision of the quant used to generate the imatrix actually
|
@@ -70,4 +41,36 @@ the fp32 GGUF and the fp32-generated imatrix was best, it was by such a miniscul
|
|
70 |
difference between that (11.9314) and the Q4km made from the fp16 GGUF with the Q4_0-generaged imatrix (11.9355) could be detected
|
71 |
under normal usage. The only counterintuitive result here is that the Q4_0-imat quants outperformed the Q8_0-imat quants. I cannot
|
72 |
think of a reason why this should be the case. But as it seemingly *is* the case, I will be using Q4_0 as my intermediate step for
|
73 |
-
generating imatrices in the future when the full fp16 model is too big for my measly 72GB of VRAM.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
# Continued Experiments 2024-05-11
|
2 |
|
3 |
As an imatrix enjoyer, it has been bugging me whether the precision of the quant used to generate the imatrix actually
|
|
|
41 |
difference between that (11.9314) and the Q4km made from the fp16 GGUF with the Q4_0-generaged imatrix (11.9355) could be detected
|
42 |
under normal usage. The only counterintuitive result here is that the Q4_0-imat quants outperformed the Q8_0-imat quants. I cannot
|
43 |
think of a reason why this should be the case. But as it seemingly *is* the case, I will be using Q4_0 as my intermediate step for
|
44 |
+
generating imatrices in the future when the full fp16 model is too big for my measly 72GB of VRAM.
|
45 |
+
|
46 |
+
|
47 |
+
|
48 |
+
|
49 |
+
# Initial Testing 2024-04-25
|
50 |
+
|
51 |
+
Some folks are claiming there's something funky going on with GGUF quanting for Llama 3 models. I don't disagree.
|
52 |
+
|
53 |
+
Some of those people are speculating that it has something to do with converting the raw weights from bf16 to fp16 instead
|
54 |
+
of converting to fp32 as an intermediate step. I think that's bollocks. There is no logical or mathmatical justification for
|
55 |
+
how that could possibly matter.
|
56 |
+
|
57 |
+
So to test this crazy theory, I downloaded Undi95/Meta-Llama-3-8B-Instruct-hf and converted it to GGUF three ways:
|
58 |
+
- fp16 specifically with `--outtype f16`
|
59 |
+
- fp32 specifically with `--outtype f32`
|
60 |
+
- "Auto" with no outtype specified
|
61 |
+
|
62 |
+
I then quantized each of these conversions to Q4_K_M and ran perplexity tests on everything using my abbreviated wiki.short.raw
|
63 |
+
text file. The results:
|
64 |
+
|
65 |
+
````
|
66 |
+
FP16 specified: size 14.9GB PPL @ fp16 9.5158 +/- 0.15418 PPL @ Q4km 9.6414 +/- 0.15494
|
67 |
+
FP32 specified: size 29.9GB PPL @ fp32 9.5158 +/- 0.15418 PPL @ Q4km 9.6278 +/- 0.15466
|
68 |
+
None specified: size 29.9GB PPL @ ???? 9.5158 +/- 0.15418 PPL @ Q4km 9.6278 +/- 0.15466
|
69 |
+
````
|
70 |
+
|
71 |
+
|
72 |
+
As you can see, converting to fp32 has no meaningful effect on PPL compared to converting to fp16. PPL is identical at full weight,
|
73 |
+
and the miniscule loss shown at Q4km is will within the margin of error. There will no doubt be some people who will claim
|
74 |
+
"PpL iSn'T gOoD eNoUgH!!1!". For those people, I have uploaded all GGUFs used in this test. Feel free to use those files to do
|
75 |
+
more extensive testing on your own time. I consider the matter resolved until somebody can conclusively demonstrate otherwise.
|
76 |
+
|