mimi_name = "mimi-dbaa9758@125.safetensors"
moshi_name = "hibiki-rs-220b12c0@200.safetensors"
tokenizer_name = "tokenizer_spm_48k_multi6_2.model"

[model]
text_in_vocab_size = 48001
text_out_vocab_size = 48000
audio_vocab_size = 2049
audio_codebooks = 32

[model.transformer]
d_model = 2560
num_heads = 20
num_layers = 24
dim_feedforward = 10240
causal = true
norm_first = true
bias_ff = false
bias_attn = false
context = 500
max_period = 100000
use_conv_block = false
use_conv_bias = true
gating = "silu"
norm = "RmsNorm"
positional_embedding = "Rope"
conv_layout = false
conv_kernel_size = 3
kv_repeat = 1
max_seq_len = 4096

[model.depformer]
num_slices = 16
low_rank_embeddings = 128

[model.depformer.transformer]
d_model = 1024
num_heads = 16
num_layers = 4
dim_feedforward = 3072
causal = true
norm_first = true
bias_ff = false
bias_attn = false
context = 16
max_period = 10000
use_conv_block = false
use_conv_bias = true
gating = "silu"
norm = "RmsNorm"
positional_embedding = "None"
conv_layout = false
conv_kernel_size = 3
kv_repeat = 1
max_seq_len = 4096

[model.conditioners.description]
type = "Lut"
n_bins = 31
dim = 16
possible_values = ["very_bad", "bad", "neutral", "good", "very_good"]