37 lines
1.0 KiB
INI
37 lines
1.0 KiB
INI
# Defines the default ftype (the quantization mix code,
|
|
# that you pass to quantize if you're not using custom mix).
|
|
# tensors that are not overriden below will be quantized
|
|
# according to this mix.
|
|
#
|
|
# Must be one of
|
|
# Q4_0, Q4_1, Q5_0, Q5_1, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M,
|
|
# IQ1_S, IQ1_M, Q2_K, Q2_K_S, IQ3_XXS, IQ3_S, IQ3_M, Q3_K,
|
|
# IQ3_XS, Q3_K_S, Q3_K_M, Q3_K_L, IQ4_NL, IQ4_XS, Q4_K,
|
|
# Q4_K_S, Q4_K_M, Q5_K, Q5_K_S, Q5_K_M, Q6_K, Q8_0, F16
|
|
|
|
ftype=Q6_K
|
|
|
|
# Defines overrides for tensors with names matching a given
|
|
# string. Filters are processed in order given, the first
|
|
# matching will be used.
|
|
#
|
|
# Wildcards are allowed:
|
|
# ? single character
|
|
# * multiple characters
|
|
#
|
|
# Type must be one of
|
|
# F16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, Q8_1, Q2_K, Q3_K,
|
|
# Q4_K, Q5_K, Q6_K, Q8_K, IQ2_XXS, IQ2_XS, IQ3_XXS,
|
|
# IQ1_S, IQ4_NL, IQ3_S, IQ2_S, IQ4_XS, IQ1_M
|
|
|
|
blk.10.ffn_up.weight=Q5_K
|
|
blk.1?.ffn_up.weight=Q4_K
|
|
blk.23.*=Q2_K
|
|
blk.24.*=Q2_K
|
|
blk.25.*=Q2_K
|
|
blk.2?.ffn_up.weight=Q4_K
|
|
*_gate*=Q4_K
|
|
*.attn*=IQ4_XS
|
|
*_down*=IQ3_S
|
|
output.weight=Q5_K
|