Only the keys from the original quantization map (found in base64 in
original model) were added quantized in the final safetensors file, the others are directly coming from the source model.
The quantization was done using native module loading from Wan2GP
from models.wan.modules.t5 import T5EncoderModel and
optimum-quanto.
No ready-made code as it was made manually in REPL.
