Deep-Miqu-103B / mergekit_config.yml
jukofyork's picture
Update mergekit_config.yml
cd4aba4 verified
const_tag: &MODEL1 jukofyork/Dawn-Miqu-70B
const_tag: &MODEL2 jukofyork/Dark-Miqu-70B
const_tag: &QK_ATTENUATION_FACTOR 0.8408964153 # sqrt(sqrt(1/2))
const_tag: &MLP_DOWN_SCALE_FACTOR 0.7071067812 # sqrt(1/2)
scale-filter-env: &scale_filter_env
parameters:
scale:
- filter: q_proj
value: *QK_ATTENUATION_FACTOR
- filter: k_proj
value: *QK_ATTENUATION_FACTOR
- filter: down_proj
value: *MLP_DOWN_SCALE_FACTOR
- value: 1.0
slices:
- sources:
- model: *MODEL1
layer_range: [0, 20]
- sources:
- model: *MODEL1
layer_range: [20, 40]
<<: *scale_filter_env
- sources:
- model: *MODEL2
layer_range: [20, 40]
<<: *scale_filter_env
- sources:
- model: *MODEL1
layer_range: [40, 60]
<<: *scale_filter_env
- sources:
- model: *MODEL2
layer_range: [40, 60]
<<: *scale_filter_env
- sources:
- model: *MODEL1
layer_range: [60, 80]
merge_method: passthrough
dtype: float16