File size: 275 Bytes
dd3d085 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 |
---
license: mit
language:
- en
---
My own (ZeroWw) quantizations.
output and embed tensors quantized to f16.
all other tensors quantized to q5_k or q6_k.
Result:
both f16.q6 and f16.q5 are smaller than q8_0 standard quantization
and they perform as well as the pure f16.
|