Ayam-2x8B-GGUF / README.md
R136a1's picture
Create README.md
b9b98ca verified
|
raw
history blame
2.23 kB
metadata
license: cc-by-nc-4.0
language:
  - en

art by myself using AI ayam

Ayam 2x8B

GGUF
Quantized using lastest llama.cpp as per writing.
Quantized using imatrix calculated from FP16 while using BF16 to create the quants to preserve accuracy.

Technical test:

Quant PPL VRAM
FP16 3.7393 +/- 0.14377 22GB+
Q8_0 3.7393 +/- 0.14381 14.8GB
Q6_K 3.7283 +/- 0.14309 11.7GB
Q5_K_M 3.7490 +/- 0.14440 10.4GB
Q4_K_M 3.7263 +/- 0.14158 9.1GB
Q4_K_S 3.7276 +/- 0.14139 8.5GB
Q3_K_M 3.8198 +/- 0.14552 7.5GB

Perplexity test using llama.cpp/perplexity.
VRAM at full 8K context using Nvidia L4 GPU.
VRAM test using llama.cpp/server.

Another MoE, this time using L3.
Recipe: Sao's Stheno-v3.2 + L3 8B Instruct.

This model is intended for personal use but I think it's really good and worth sharing. Stheno-v3.2 is, as you probably know well, very good. In creative writing, RP and ERP it's far better than L3 Instruct and to be honest it's the best L3 finetunes I've tried so far so yeah I liked it very much. But while playing with it, I feel like the model is (a bit) dumber than L3 Instruct. It can't understand complex scenario well and confused in multi-char scenario, at least that what I was experiencing. So yeah, I tried to improve its intelligence while still preserving its creativity.

Why MoE not merge?
Well... 2 model working together is always better than merging it into one. And (surprisingly) the result is far exceeded my expectations.
And I think merging models sometimes can damage It's quality.

Testing condition (using SillyTavern):
Context and Instruct: Llama 3 Instruct.
Sampler:

Temperature : 1.15
MinP : 0.075
TopK : 50
Other is disabled.

Switch

BF16 - GGUF