| |
| |
|
|
| from transformers.configuration_utils import PretrainedConfig |
|
|
|
|
| class MiniCPMTTSConfig(PretrainedConfig): |
| model_type = "minicpmtts" |
|
|
| def __init__( |
| self, |
| llm_dim: int = 2560, |
| llm_intermediate_size: int = 768, |
| llm_down_scale: bool = False, |
| llm_dim_model_base: int = 256, |
| projector_type: str = "mlp", |
| hidden_act: str = "silu", |
| aug_loss_weight: bool = False, |
| aug_layer_loss_weight: bool = False, |
| filter_tts_loss: bool = False, |
| tts_filter_loss_fix: bool = False, |
| long_weight: float = 0.1, |
| short_weight: float = 0.1, |
| hidden_size: int = 768, |
| intermediate_size: int = 3072, |
| num_attention_heads: int = 12, |
| num_hidden_layers: int = 20, |
| num_key_value_heads: int = 12, |
| max_position_embeddings: int = 4096, |
| num_audio_tokens: int = 4097, |
| num_text_tokens: int = 21178, |
| num_mel_bins: int = 100, |
| num_vq: int = 1, |
| use_llm_hidden_state: bool = False, |
| audio_bos_token_id: int = 21132, |
| text_eos_token_id: int = 21133, |
| use_text: bool = True, |
| streaming: bool = False, |
| streaming_text_chunk_min: int = 3, |
| streaming_text_chunk_max: int = 7, |
| streaming_text_reserved_len: int = 300, |
| streaming_audio_chunk_size: int = 50, |
| attn_implementation: str = "sdpa", |
| condition_type: str = "llm_hidden", |
| backbone_model: str = "llama", |
| audio_tokenizer_type: str = "wavtokenizer", |
| audio_tokenizer_sample_rate: int = 24000, |
| streaming_sliding_window: bool = False, |
| streaming_sliding_window_max_text_len: int = 500, |
| streaming_sliding_window_average_speed: int = 5, |
| streaming_sliding_window_fast_speed: int = 7, |
| streaming_sliding_window_slow_speed: int = 3, |
| streaming_sliding_window_audio_frame_rate: int = 50, |
| streaming_sliding_window_text_window_size: int = 50, |
| streaming_sliding_window_audio_init_text_length: int = 10, |
| streaming_sliding_window_audio_window_size: int = 300, |
| normalize_projected_hidden: bool = False, |
| interleaved: bool = False, |
| attention_type: str = "sliding_recompute", |
| recomputed_chunks: int = 1, |
| window_size: int = 2, |
| **kwargs, |
| ): |
| super().__init__(**kwargs) |
|
|
| self.llm_dim = llm_dim |
| self.llm_hidden_size = llm_dim |
| self.llm_intermediate_size = llm_intermediate_size |
| self.llm_down_scale = llm_down_scale |
| self.llm_dim_model_base = llm_dim_model_base |
| self.projector_type = projector_type |
| self.aug_loss_weight = aug_loss_weight |
| self.aug_layer_loss_weight = aug_layer_loss_weight |
| self.tts_filter_loss_fix = tts_filter_loss_fix |
| self.filter_tts_loss = filter_tts_loss |
| self.long_weight = long_weight |
| self.short_weight = short_weight |
| self.hidden_act = hidden_act |
|
|
| self.hidden_size = hidden_size |
| self.intermediate_size = intermediate_size |
| self.num_attention_heads = num_attention_heads |
| self.num_hidden_layers = num_hidden_layers |
| self.num_key_value_heads = num_key_value_heads |
| self.max_position_embeddings = max_position_embeddings |
| self.num_audio_tokens = num_audio_tokens |
| self.num_text_tokens = num_text_tokens |
| self.num_mel_bins = num_mel_bins |
| self.num_vq = num_vq |
| self.use_llm_hidden_state = use_llm_hidden_state |
| self.audio_bos_token_id = audio_bos_token_id |
| self.text_eos_token_id = text_eos_token_id |
| self.use_text = use_text |
| self.streaming = streaming |
| self.streaming_text_chunk_min = streaming_text_chunk_min |
| self.streaming_text_chunk_max = streaming_text_chunk_max |
| self.streaming_text_reserved_len = streaming_text_reserved_len |
| self.streaming_audio_chunk_size = streaming_audio_chunk_size |
| self.attn_implementation = attn_implementation |
| self.condition_type = condition_type |
| self.backbone_model = backbone_model |
| self.audio_tokenizer_type = audio_tokenizer_type |
| self.audio_tokenizer_sample_rate = audio_tokenizer_sample_rate |
|
|
| self.streaming_sliding_window = streaming_sliding_window |
| self.streaming_sliding_window_max_text_len = streaming_sliding_window_max_text_len |
| self.streaming_sliding_window_average_speed = streaming_sliding_window_average_speed |
| self.streaming_sliding_window_fast_speed = streaming_sliding_window_fast_speed |
| self.streaming_sliding_window_slow_speed = streaming_sliding_window_slow_speed |
| self.streaming_sliding_window_audio_frame_rate = streaming_sliding_window_audio_frame_rate |
| self.streaming_sliding_window_text_window_size = streaming_sliding_window_text_window_size |
| self.streaming_sliding_window_audio_init_text_length = streaming_sliding_window_audio_init_text_length |
| self.streaming_sliding_window_audio_window_size = streaming_sliding_window_audio_window_size |
|
|
| self.normalize_projected_hidden = normalize_projected_hidden |
|
|
| self.interleaved = interleaved |
| self.attention_type = attention_type |
| self.recomputed_chunks = recomputed_chunks |
| self.window_size = window_size |
|
|
