OpenMOSS-Team/MOSS-TTS-Realtime
10711k
1# Copyright 2026 OpenMOSS and the HuggingFace Inc. team. All rights reserved.2#3# Licensed under the Apache License, Version 2.0 (the "License");4# you may not use this file except in compliance with the License.5# You may obtain a copy of the License at6#7# http://www.apache.org/licenses/LICENSE-2.08#9# Unless required by applicable law or agreed to in writing, software10# distributed under the License is distributed on an "AS IS" BASIS,11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.12# See the License for the specific language governing permissions and13# limitations under the License.14"""Local transformer used by MossTTSRealtime for RVQ codebook decoding."""15 16from __future__ import annotations17 18from typing import Optional, Union19 20import torch21import torch.nn as nn22 23from transformers.activations import ACT2FN24from transformers.cache_utils import Cache, StaticCache25from transformers.generation import GenerationMixin26from transformers.modeling_flash_attention_utils import FlashAttentionKwargs27from transformers.modeling_layers import GradientCheckpointingLayer28from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast29from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update30from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel31from transformers.masking_utils import create_causal_mask32from transformers.processing_utils import Unpack33from transformers.loss.loss_utils import ForCausalLMLoss34from transformers.utils import TransformersKwargs, logging35from .configuration_mossttsrealtime import MossTTSRealtimeLocalTransformerConfig36 37logger = logging.get_logger(__name__)38 39 40class MossTTSRealtimeLocalTransformerRMSNorm(nn.Module):41 def __init__(self, hidden_size, eps=1e-6) -> None:42 super().__init__()43 self.weight = nn.Parameter(torch.ones(hidden_size))44 self.variance_epsilon = eps45 46 def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:47 input_dtype = hidden_states.dtype48 hidden_states = hidden_states.to(torch.float32)49 variance = hidden_states.pow(2).mean(-1, keepdim=True)50 hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)51 return self.weight * hidden_states.to(input_dtype)52 53 def extra_repr(self):54 return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"55 56 57class MossTTSRealtimeLocalTransformerMLP(nn.Module):58 def __init__(self, config: MossTTSRealtimeLocalTransformerConfig):59 super().__init__()60 self.config = config61 self.hidden_size = config.hidden_size62 self.intermediate_size = config.intermediate_size63 self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)64 self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)65 self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)66 self.act_fn = ACT2FN[config.hidden_act]67 68 def forward(self, x):69 down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))70 return down_proj71 72 73def rotate_half(x):74 x1 = x[..., : x.shape[-1] // 2]75 x2 = x[..., x.shape[-1] // 2 :]76 return torch.cat((-x2, x1), dim=-1)77 78 79def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):80 cos = cos.unsqueeze(unsqueeze_dim)81 sin = sin.unsqueeze(unsqueeze_dim)82 q_embed = (q * cos) + (rotate_half(q) * sin)83 k_embed = (k * cos) + (rotate_half(k) * sin)84 return q_embed, k_embed85 86 87def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:88 batch, num_key_value_heads, slen, head_dim = hidden_states.shape89 if n_rep == 1:90 return hidden_states91 hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)92 return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)93 94 95def eager_attention_forward(96 module: nn.Module,97 query: torch.Tensor,98 key: torch.Tensor,99 value: torch.Tensor,100 attention_mask: Optional[torch.Tensor],101 scaling: float,102 dropout: float = 0.0,103 **kwargs: Unpack[TransformersKwargs],104):105 key_states = repeat_kv(key, module.num_key_value_groups)106 value_states = repeat_kv(value, module.num_key_value_groups)107 attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling108 if attention_mask is not None:109 causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]110 attn_weights = attn_weights + causal_mask111 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)112 attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)113 attn_output = torch.matmul(attn_weights, value_states)114 attn_output = attn_output.transpose(1, 2).contiguous()115 return attn_output, attn_weights116 117 118class MossTTSRealtimeLocalTransformerAttention(nn.Module):119 def __init__(self, config: MossTTSRealtimeLocalTransformerConfig, layer_idx: int):120 super().__init__()121 self.config = config122 self.layer_idx = layer_idx123 self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)124 self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads125 self.scaling = self.head_dim**-0.5126 self.attention_dropout = config.attention_dropout127 self.is_causal = True128 129 self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias)130 self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias)131 self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias)132 self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias)133 self.q_norm = MossTTSRealtimeLocalTransformerRMSNorm(self.head_dim, eps=config.rms_norm_eps)134 self.k_norm = MossTTSRealtimeLocalTransformerRMSNorm(self.head_dim, eps=config.rms_norm_eps)135 self.sliding_window = None136 137 def forward(138 self,139 hidden_states: torch.Tensor,140 position_embeddings: tuple[torch.Tensor, torch.Tensor],141 attention_mask: Optional[torch.Tensor],142 past_key_values: Optional[Cache] = None,143 cache_position: Optional[torch.LongTensor] = None,144 **kwargs: Unpack[FlashAttentionKwargs],145 ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:146 input_shape = hidden_states.shape[:-1]147 hidden_shape = (*input_shape, -1, self.head_dim)148 query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)149 key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)150 value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)151 cos, sin = position_embeddings152 153 query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)154 155 if past_key_values is not None:156 cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}157 key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)158 159 attention_interface = eager_attention_forward160 if self.config._attn_implementation != "eager":161 attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]162 163 attn_output, attn_weights = attention_interface(164 self,165 query_states,166 key_states,167 value_states,168 attention_mask,169 dropout=0.0 if not self.training else self.attention_dropout,170 scaling=self.scaling,171 sliding_window=self.sliding_window,172 **kwargs,173 )174 175 attn_output = attn_output.reshape(*input_shape, -1).contiguous()176 attn_output = self.o_proj(attn_output)177 return attn_output, attn_weights178 179 180class MossTTSRealtimeLocalTransformerDecoderLayer(GradientCheckpointingLayer):181 def __init__(self, config: MossTTSRealtimeLocalTransformerConfig, layer_idx: int):182 super().__init__()183 self.hidden_size = config.hidden_size184 self.self_attn = MossTTSRealtimeLocalTransformerAttention(config=config, layer_idx=layer_idx)185 self.mlp = MossTTSRealtimeLocalTransformerMLP(config)186 self.input_layernorm = MossTTSRealtimeLocalTransformerRMSNorm(config.hidden_size, eps=config.rms_norm_eps)187 self.post_attention_layernorm = MossTTSRealtimeLocalTransformerRMSNorm(config.hidden_size, eps=config.rms_norm_eps)188 self.attention_type = "full_attention"189 190 def forward(191 self,192 hidden_states: torch.Tensor,193 attention_mask: Optional[torch.Tensor] = None,194 position_ids: Optional[torch.LongTensor] = None,195 past_key_values: Optional[Cache] = None,196 use_cache: Optional[bool] = False,197 cache_position: Optional[torch.LongTensor] = None,198 position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,199 **kwargs: Unpack[TransformersKwargs],200 ) -> torch.Tensor:201 residual = hidden_states202 hidden_states = self.input_layernorm(hidden_states)203 hidden_states, _ = self.self_attn(204 hidden_states=hidden_states,205 attention_mask=attention_mask,206 position_ids=position_ids,207 past_key_values=past_key_values,208 use_cache=use_cache,209 cache_position=cache_position,210 position_embeddings=position_embeddings,211 **kwargs,212 )213 hidden_states = residual + hidden_states214 residual = hidden_states215 hidden_states = self.post_attention_layernorm(hidden_states)216 hidden_states = self.mlp(hidden_states)217 hidden_states = residual + hidden_states218 return hidden_states219 220 221class MossTTSRealtimeLocalTransformerPreTrainedModel(PreTrainedModel):222 223 config_class = MossTTSRealtimeLocalTransformerConfig224 config: MossTTSRealtimeLocalTransformerConfig225 226 base_model_prefix = "local_transformer"227 supports_gradient_checkpointing = True228 _no_split_modules = ["MossTTSRealtimeLocalTransformerDecoderLayer"]229 _skip_keys_device_placement = ["past_key_values"]230 _supports_sdpa = True231 _supports_flex_attn = True232 _supports_flash_attn = True233 _can_compile_fullgraph = True234 _supports_attention_backend = True235 236 _can_record_outputs = {237 "hidden_states": MossTTSRealtimeLocalTransformerDecoderLayer,238 "attentions": MossTTSRealtimeLocalTransformerAttention,239 }240 241 242class MossTTSRealtimeLocalTransformerRotaryEmbedding(nn.Module):243 inv_freq: torch.Tensor244 245 def __init__(self, config: MossTTSRealtimeLocalTransformerConfig, device=None):246 super().__init__()247 self.config = config248 self.rope_type = getattr(config, "rope_type", "linear")249 self.max_seq_len_cached = config.max_position_embeddings250 self.original_max_seq_len = config.max_position_embeddings251 self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]252 inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)253 self.register_buffer("inv_freq", inv_freq, persistent=False)254 self.original_inv_freq = self.inv_freq255 256 @torch.no_grad()257 @dynamic_rope_update258 def forward(self, x, position_ids):259 inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)260 position_ids_expanded = position_ids[:, None, :].float()261 device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"262 with torch.autocast(device_type=device_type, enabled=False):263 freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)264 emb = torch.cat((freqs, freqs), dim=-1)265 cos = emb.cos() * self.attention_scaling266 sin = emb.sin() * self.attention_scaling267 return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)268 269 270class MossTTSRealtimeLocalTransformer(MossTTSRealtimeLocalTransformerPreTrainedModel):271 def __init__(self, config: MossTTSRealtimeLocalTransformerConfig):272 super().__init__(config)273 self.padding_idx = config.pad_token_id274 self.embed_tokens = nn.ModuleList(275 [nn.Embedding(config.audio_vocab_size, config.hidden_size, config.audio_pad_token) for _ in range(config.rvq - 1)]276 )277 self.layers = nn.ModuleList(278 [MossTTSRealtimeLocalTransformerDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]279 )280 self.norm = MossTTSRealtimeLocalTransformerRMSNorm(config.hidden_size, eps=config.rms_norm_eps)281 self.rotary_emb = MossTTSRealtimeLocalTransformerRotaryEmbedding(config=config)282 self.gradient_checkpointing = False283 self.has_sliding_layers = None284 self.post_init()285 286 def forward(287 self,288 input_ids: Optional[torch.LongTensor] = None,289 backbone_last_hidden_state: Optional[torch.FloatTensor] = None,290 attention_mask: Optional[torch.Tensor] = None,291 position_ids: Optional[torch.LongTensor] = None,292 past_key_values: Optional[Cache] = None,293 inputs_embeds: Optional[torch.FloatTensor] = None,294 use_cache: Optional[bool] = None,295 cache_position: Optional[torch.LongTensor] = None,296 codebook_idx: Optional[int] = None,297 **kwargs: Unpack[TransformersKwargs],298 ) -> BaseModelOutputWithPast:299 if position_ids is not None and not torch.compiler.is_compiling():300 position_ids = None301 302 if (input_ids is None) == (inputs_embeds is None):303 raise ValueError("You must specify exactly one of input_ids or inputs_embeds.")304 305 if use_cache and past_key_values is None:306 device = inputs_embeds.device if inputs_embeds is not None else input_ids.device307 past_key_values = StaticCache(config=self.config, max_cache_len=self.config.rvq, device=device)308 309 if cache_position is None:310 past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0311 inputs_seq_length = inputs_embeds.shape[1] if inputs_embeds is not None else input_ids.shape[1]312 device = inputs_embeds.device if inputs_embeds is not None else input_ids.device313 cache_position = torch.arange(past_seen_tokens, past_seen_tokens + inputs_seq_length, device=device)314 315 if inputs_embeds is None:316 if codebook_idx is not None:317 if codebook_idx <= 0:318 raise ValueError(f"`codebook_idx` must be in [1, {len(self.embed_tokens)}], got {codebook_idx}.")319 if codebook_idx > len(self.embed_tokens):320 raise ValueError(f"`codebook_idx` must be in [1, {len(self.embed_tokens)}], got {codebook_idx}.")321 if input_ids.ndim == 1:322 input_ids = input_ids.unsqueeze(1)323 token_emb = self.embed_tokens[codebook_idx - 1](input_ids[:, 0]).unsqueeze(1) # [B,1,H]324 inputs_embeds = token_emb325 else:326 if input_ids.shape[1] != cache_position.shape[0]:327 raise ValueError(328 "`input_ids` and `cache_position` must align in sequence length: "329 f"got {input_ids.shape[1]} and {cache_position.shape[0]}."330 )331 codebook_idxs = torch.clamp(cache_position - 1, min=0, max=len(self.embed_tokens) - 1)332 inputs_embeds = torch.stack(333 [334 self.embed_tokens[codebook_idx](input_ids[:, seq_idx])335 for seq_idx, codebook_idx in enumerate(codebook_idxs.tolist())336 ],337 dim=1,338 )339 340 input_ids_are_first_codebook = bool(cache_position[0] == 0)341 if backbone_last_hidden_state is not None:342 inputs_embeds[:, 0, :] = backbone_last_hidden_state[:, 0, :]343 else:344 if not torch.compiler.is_compiling() and input_ids_are_first_codebook:345 logger.warning(346 "When the first codebook token is provided, `backbone_last_hidden_state` should also be provided for correct inference."347 )348 349 causal_mask = create_causal_mask(350 config=self.config,351 input_embeds=inputs_embeds,352 attention_mask=attention_mask,353 cache_position=cache_position,354 past_key_values=past_key_values,355 position_ids=position_ids,356 )357 358 hidden_states = inputs_embeds359 position_ids = cache_position.unsqueeze(0)360 position_embeddings = self.rotary_emb(hidden_states, position_ids)361 362 for decoder_layer in self.layers[: self.config.num_hidden_layers]:363 hidden_states = decoder_layer(364 hidden_states,365 attention_mask=causal_mask,366 position_ids=position_ids,367 past_key_values=past_key_values,368 use_cache=use_cache,369 cache_position=cache_position,370 position_embeddings=position_embeddings,371 **kwargs,372 )373 hidden_states = self.norm(hidden_states)374 return BaseModelOutputWithPast(375 last_hidden_state=hidden_states,376 past_key_values=past_key_values if use_cache else None,377 )378 379 380class MossTTSRealtimeLocalTransformerForCausalLM(MossTTSRealtimeLocalTransformerPreTrainedModel, GenerationMixin):381 _tied_weights_keys = None382 _tp_plan = None383 _pp_plan = None384 385 def __init__(self, config):386 super().__init__(config)387 self.model = MossTTSRealtimeLocalTransformer(config)388 self.audio_vocab_size = self.config.audio_vocab_size389 390 self.local_lm_heads = nn.ModuleList(391 [nn.Linear(config.hidden_size, config.audio_vocab_size, bias=False) for _ in range(config.rvq)]392 )393 self.post_init()394 395 def forward(396 self,397 input_ids: Optional[torch.LongTensor] = None,398 backbone_last_hidden_state: Optional[torch.FloatTensor] = None,399 attention_mask: Optional[torch.Tensor] = None,400 position_ids: Optional[torch.LongTensor] = None,401 past_key_values: Optional[Union[Cache, list[torch.FloatTensor]]] = None,402 inputs_embeds: Optional[torch.FloatTensor] = None,403 labels: Optional[torch.LongTensor] = None,404 use_cache: Optional[bool] = None,405 cache_position: Optional[torch.LongTensor] = None,406 codebook_idx: Optional[int] = None,407 logits_to_keep: Union[int, torch.Tensor] = 0,408 **kwargs: Unpack[TransformersKwargs],409 ) -> Union[tuple, CausalLMOutputWithPast]:410 outputs = self.model(411 input_ids=input_ids,412 backbone_last_hidden_state=backbone_last_hidden_state,413 inputs_embeds=inputs_embeds,414 attention_mask=attention_mask,415 position_ids=position_ids,416 past_key_values=past_key_values,417 use_cache=use_cache,418 cache_position=cache_position,419 codebook_idx=codebook_idx,420 **kwargs,421 )422 423 hidden_states = outputs.last_hidden_state424 425 if isinstance(logits_to_keep, int):426 if logits_to_keep == 0:427 slice_indices = slice(0, None)428 else:429 slice_indices = slice(-logits_to_keep, None)430 else:431 slice_indices = logits_to_keep432 hs = hidden_states[:, slice_indices, :]433 434 if cache_position is not None:435 if codebook_idx is None:436 raise ValueError("`codebook_idx` must be provided when `cache_position` is provided.")437 logits = self.local_lm_heads[codebook_idx](hs[:, 0, :]).unsqueeze(1)438 else:439 if hs.shape[1] > len(self.local_lm_heads):440 raise ValueError(441 f"Cannot project {hs.shape[1]} codebooks with only {len(self.local_lm_heads)} LM heads."442 )443 logits_list = []444 for i in range(hs.shape[1]):445 logits_list.append(self.local_lm_heads[i](hs[:, i, :]))446 logits = torch.stack(logits_list, dim=1)447 448 logits = logits.contiguous()449 loss = None450 if labels is not None:451 loss = ForCausalLMLoss(logits, None, self.audio_vocab_size, shift_labels=labels.contiguous())452 453 return CausalLMOutputWithPast(454 loss=loss,455 logits=logits,456 past_key_values=outputs.past_key_values,457 hidden_states=outputs.hidden_states,458 attentions=outputs.attentions,459 )460 461__all__ = [462 "MossTTSRealtimeLocalTransformer",463 "MossTTSRealtimeLocalTransformerAttention",464 "MossTTSRealtimeLocalTransformerConfig",465 "MossTTSRealtimeLocalTransformerDecoderLayer",466 "MossTTSRealtimeLocalTransformerForCausalLM",467 "MossTTSRealtimeLocalTransformerPreTrainedModel",468 "MossTTSRealtimeLocalTransformerRMSNorm",469 "MossTTSRealtimeLocalTransformerRotaryEmbedding",470]471