@@ -0,0 +1,13 @@
+defaults:
+ - text2semantic_pretrain_small
+ - _self_
+
+project: text2semantic_pretrain_large_4_in_8_codebooks
+# Model Configuration
+model:
+ model:
+ config:
+ n_layer: 36
+ n_head: 20
+ dim: 1280
+project: text2semantic_pretrain_medium_4_in_8_codebooks
+ n_layer: 24
+ n_head: 16
+ dim: 1024
@@ -2,7 +2,7 @@ defaults:
- base
- _self_
-project: text2semantic_pretrain_400m_4_in_8_codebooks
+project: text2semantic_pretrain_small_4_in_8_codebooks
max_length: 2048
# Lightning Trainer
@@ -58,9 +58,9 @@ model:
_target_: fish_speech.models.text2semantic.llama.ModelArgs
max_seq_len: 4096
vocab_size: 36408
- n_layer: 24
- n_head: 16
- dim: 1024
+ n_layer: 12
+ n_head: 12
+ dim: 768
rope_base: 10000
norm_eps: 1e-5
num_in_codebooks: 4 # input codebook size