| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """MiMo audio: tokenizer, encoding utilities, and audio encoder. |
| 4 | |
| 5 | Ported from SGLang's mimo_audio.py. |
| 6 | Audio tokenizer adapted from https://github.com/XiaomiMiMo/MiMo-Audio-Tokenizer.git |
| 7 | """ |
| 8 | |
| 9 | import dataclasses |
| 10 | import json |
| 11 | import logging |
| 12 | import math |
| 13 | import os |
| 14 | import typing as tp |
| 15 | from dataclasses import dataclass |
| 16 | from functools import wraps |
| 17 | |
| 18 | import torch |
| 19 | import torch.distributed as dist |
| 20 | import torch.nn as nn |
| 21 | import torch.nn.functional as F |
| 22 | from einops import rearrange, repeat |
| 23 | from transformers.activations import ACT2FN |
| 24 | from transformers.configuration_utils import PreTrainedConfig |
| 25 | from transformers.modeling_utils import PreTrainedModel |
| 26 | from transformers.models.qwen2.configuration_qwen2 import Qwen2Config |
| 27 | from transformers.models.qwen2.modeling_qwen2 import Qwen2Model |
| 28 | |
| 29 | logger = logging.getLogger(__name__) |
| 30 | |
| 31 | |
| 32 | # --------------------------------------------------------------------------- |
| 33 | # Vector quantization (from MiMo-Audio-Tokenizer) |
| 34 | # --------------------------------------------------------------------------- |
| 35 | |
| 36 | |
| 37 | def _vq_default(val: tp.Any, d: tp.Any) -> tp.Any: |
| 38 | return val if val is not None else d |
| 39 | |
| 40 | |
| 41 | def _ema_inplace(moving_avg, new, decay: float): |
| 42 | if dist.is_initialized(): |
| 43 | dist.all_reduce(new, op=dist.ReduceOp.SUM) |
| 44 | moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay)) |
| 45 | |
| 46 | |
| 47 | def _laplace_smoothing(x, n_categories: int, epsilon: float = 1e-5): |
| 48 | return (x + epsilon) / (x.sum() + n_categories * epsilon) |
| 49 | |
| 50 | |
| 51 | def _uniform_init(*shape: int): |
| 52 | t = torch.empty(shape) |
| 53 | nn.init.kaiming_uniform_(t) |
| 54 | return t |
| 55 | |
| 56 | |
| 57 | def _sample_vectors(samples, num: int): |
| 58 | num_samples, device = samples.shape[0], samples.device |
| 59 | |
| 60 | if num_samples >= num: |
| 61 | indices = torch.randperm(num_samples, device=device)[:num] |
| 62 | else: |
| 63 | indices = torch.randint(0, num_samples, (num,), device=device) |
| 64 | |
| 65 | selected_samples = samples[indices] |
| 66 | |
| 67 | if dist.is_initialized(): |
| 68 | dist.broadcast(selected_samples, src=0) |
| 69 | |
| 70 | return selected_samples |
| 71 | |
| 72 | |
| 73 | def _kmeans(samples, num_clusters: int, num_iters: int = 10): |
| 74 | dim, dtype = samples.shape[-1], samples.dtype |
| 75 | |
| 76 | means = _sample_vectors(samples, num_clusters) |
| 77 | |
| 78 | for _ in range(num_iters): |
| 79 | dists = -( |
| 80 | samples.pow(2).sum(1, keepdim=True) |
| 81 | - 2 * samples @ means.t() |
| 82 | + means.t().pow(2).sum(0, keepdim=True) |
| 83 | ) |
| 84 | |
| 85 | buckets = dists.max(dim=-1).indices |
| 86 | bins = torch.bincount(buckets, minlength=num_clusters) |
| 87 | |
| 88 | new_means = buckets.new_zeros(num_clusters, dim, dtype=dtype) |
| 89 | new_means = new_means.scatter_add_( |
| 90 | 0, repeat(buckets, "n -> n d", d=dim), samples |
| 91 | ) |
| 92 | |
| 93 | if dist.is_initialized(): |
| 94 | dist.all_reduce(bins, op=dist.ReduceOp.SUM) |
| 95 | dist.all_reduce(new_means, op=dist.ReduceOp.SUM) |
| 96 | |
| 97 | zero_mask = bins == 0 |
| 98 | bins_min_clamped = bins.masked_fill(zero_mask, 1) |
| 99 | |
| 100 | new_means = new_means / bins_min_clamped[..., None] |
| 101 | |
| 102 | means = torch.where(zero_mask[..., None], means, new_means) |
| 103 | |
| 104 | return means, bins |
| 105 | |
| 106 | |
| 107 | def _rotate_half(x): |
| 108 | x1 = x[..., : x.shape[-1] // 2] |
| 109 | x2 = x[..., x.shape[-1] // 2 :] |
| 110 | return torch.cat((-x2, x1), dim=-1) |
| 111 | |
| 112 | |
| 113 | def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1): |
| 114 | cos = cos.unsqueeze(unsqueeze_dim) |
| 115 | sin = sin.unsqueeze(unsqueeze_dim) |
| 116 | q_embed = (q * cos) + (_rotate_half(q) * sin) |
| 117 | k_embed = (k * cos) + (_rotate_half(k) * sin) |
| 118 | return q_embed, k_embed |
| 119 | |
| 120 | |
| 121 | def _compute_default_rope_parameters( |
| 122 | config=None, device=None, seq_len=None, **rope_kwargs |
| 123 | ): |
| 124 | if config is not None and len(rope_kwargs) > 0: |
| 125 | raise ValueError( |
| 126 | "Unexpected arguments: `**rope_kwargs` and `config` are mutually exclusive" |
| 127 | ) |
| 128 | if len(rope_kwargs) > 0: |
| 129 | base = rope_kwargs["base"] |
| 130 | dim = rope_kwargs["dim"] |
| 131 | elif config is not None: |
| 132 | base = config.rope_theta |
| 133 | partial_rotary_factor = ( |
| 134 | config.partial_rotary_factor |
| 135 | if hasattr(config, "partial_rotary_factor") |
| 136 | else 1.0 |
| 137 | ) |
| 138 | head_dim = ( |
| 139 | getattr(config, "head_dim", None) |
| 140 | or config.hidden_size // config.num_attention_heads |
| 141 | ) |
| 142 | dim = int(head_dim * partial_rotary_factor) |
| 143 | attention_factor = 1.0 |
| 144 | inv_freq = 1.0 / ( |
| 145 | base |
| 146 | ** ( |
| 147 | torch.arange(0, dim, 2, dtype=torch.int64).to( |
| 148 | device=device, dtype=torch.float |
| 149 | ) |
| 150 | / dim |
| 151 | ) |
| 152 | ) |
| 153 | return inv_freq, attention_factor |
| 154 | |
| 155 | |
| 156 | _ROPE_INIT_FUNCTIONS = { |
| 157 | "default": _compute_default_rope_parameters, |
| 158 | } |
| 159 | |
| 160 | |
| 161 | def _dynamic_rope_update(rope_forward): |
| 162 | def dynamic_frequency_update(self, position_ids, device): |
| 163 | seq_len = torch.max(position_ids) + 1 |
| 164 | if seq_len > self.max_seq_len_cached: |
| 165 | inv_freq, self.attention_scaling = self.rope_init_fn( |
| 166 | self.config, device, seq_len=seq_len |
| 167 | ) |
| 168 | self.register_buffer("inv_freq", inv_freq, persistent=False) |
| 169 | self.max_seq_len_cached = seq_len |
| 170 | |
| 171 | if ( |
| 172 | seq_len < self.original_max_seq_len |
| 173 | and self.max_seq_len_cached > self.original_max_seq_len |
| 174 | ): |
| 175 | self.original_inv_freq = self.original_inv_freq.to(device) |
| 176 | self.register_buffer("inv_freq", self.original_inv_freq, persistent=False) |
| 177 | self.max_seq_len_cached = self.original_max_seq_len |
| 178 | |
| 179 | @wraps(rope_forward) |
| 180 | def wrapper(self, x, position_ids): |
| 181 | if "dynamic" in self.rope_type: |
| 182 | dynamic_frequency_update(self, position_ids, device=x.device) |
| 183 | return rope_forward(self, x, position_ids) |
| 184 | |
| 185 | return wrapper |
| 186 | |
| 187 | |
| 188 | class AudioRotaryEmbedding(nn.Module): |
| 189 | def __init__(self, base, dim, max_seq_len, rope_type="default", device=None): |
| 190 | super().__init__() |
| 191 | self.max_seq_len = max_seq_len |
| 192 | self.rope_type = rope_type |
| 193 | self.rope_init_fn = _ROPE_INIT_FUNCTIONS[self.rope_type] |
| 194 | inv_freq, self.attention_scaling = self.rope_init_fn( |
| 195 | device=device, base=base, dim=dim |
| 196 | ) |
| 197 | self.register_buffer("inv_freq", inv_freq, persistent=False) |
| 198 | self.original_inv_freq = self.inv_freq |
| 199 | |
| 200 | @torch.no_grad() |
| 201 | @_dynamic_rope_update |
| 202 | def forward(self, x, position_ids): |
| 203 | inv_freq_expanded = self.inv_freq[:, None].float().expand(-1, 1).to(x.device) |
| 204 | position_ids_expanded = position_ids[None, :].float() |
| 205 | device_type = ( |
| 206 | x.device.type |
| 207 | if isinstance(x.device.type, str) and x.device.type != "mps" |
| 208 | else "cpu" |
| 209 | ) |
| 210 | with torch.autocast(device_type=device_type, enabled=False): |
| 211 | freqs = ( |
| 212 | inv_freq_expanded.float() @ position_ids_expanded.float() |
| 213 | ).transpose(0, 1) |
| 214 | emb = torch.cat((freqs, freqs), dim=-1) |
| 215 | cos = emb.cos() * self.attention_scaling |
| 216 | sin = emb.sin() * self.attention_scaling |
| 217 | return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) |
| 218 | |
| 219 | |
| 220 | class EuclideanCodebook(nn.Module): |
| 221 | def __init__( |
| 222 | self, |
| 223 | dim: int, |
| 224 | codebook_size: int, |
| 225 | kmeans_init: int = False, |
| 226 | kmeans_iters: int = 10, |
| 227 | decay: float = 0.99, |
| 228 | epsilon: float = 1e-5, |
| 229 | threshold_ema_dead_code: int = 2, |
| 230 | ): |
| 231 | super().__init__() |
| 232 | self.decay = decay |
| 233 | init_fn: tp.Callable[..., torch.Tensor] | tp.Any = ( |
| 234 | _uniform_init if not kmeans_init else torch.zeros |
| 235 | ) |
| 236 | embed = init_fn(codebook_size, dim) |
| 237 | |
| 238 | self.codebook_size = codebook_size |
| 239 | self.kmeans_iters = kmeans_iters |
| 240 | self.epsilon = epsilon |
| 241 | self.threshold_ema_dead_code = threshold_ema_dead_code |
| 242 | |
| 243 | self.register_buffer("inited", torch.Tensor([not kmeans_init])) |
| 244 | self.register_buffer("cluster_size", torch.zeros(codebook_size)) |
| 245 | self.register_buffer("embed", embed) |
| 246 | self.register_buffer("embed_avg", embed.clone()) |
| 247 | |
| 248 | @torch.jit.ignore |
| 249 | def init_embed_(self, data): |
| 250 | if self.inited: |
| 251 | return |
| 252 | |
| 253 | embed, cluster_size = _kmeans(data, self.codebook_size, self.kmeans_iters) |
| 254 | self.embed.data.copy_(embed) |
| 255 | self.embed_avg.data.copy_(embed.clone()) |
| 256 | self.cluster_size.data.copy_(cluster_size) |
| 257 | self.inited.data.copy_(torch.Tensor([True])) |
| 258 | |
| 259 | def replace_(self, samples, mask): |
| 260 | replace_num = mask.sum() |
| 261 | modified_codebook = self.embed.clone() |
| 262 | modified_codebook[mask] = _sample_vectors(samples, replace_num) |
| 263 | self.embed.data.copy_(modified_codebook) |
| 264 | |
| 265 | def expire_codes_(self, batch_samples): |
| 266 | if self.threshold_ema_dead_code == 0: |
| 267 | return |
| 268 | |
| 269 | expired_codes = self.cluster_size < self.threshold_ema_dead_code |
| 270 | if not torch.any(expired_codes): |
| 271 | return |
| 272 | |
| 273 | batch_samples = rearrange(batch_samples, "... d -> (...) d") |
| 274 | self.replace_(batch_samples, mask=expired_codes) |
| 275 | |
| 276 | def preprocess(self, x): |
| 277 | x = rearrange(x, "... d -> (...) d") |
| 278 | return x |
| 279 | |
| 280 | def quantize(self, x): |
| 281 | embed = self.embed.t() |
| 282 | dist_val = -( |
| 283 | x.pow(2).sum(1, keepdim=True) |
| 284 | - 2 * x @ embed |
| 285 | + embed.pow(2).sum(0, keepdim=True) |
| 286 | ) |
| 287 | embed_ind = dist_val.max(dim=-1).indices |
| 288 | return embed_ind |
| 289 | |
| 290 | def postprocess_emb(self, embed_ind, shape): |
| 291 | return embed_ind.view(*shape[:-1]) |
| 292 | |
| 293 | def dequantize(self, embed_ind): |
| 294 | quantize = F.embedding(embed_ind, self.embed) |
| 295 | return quantize |
| 296 | |
| 297 | def encode(self, x): |
| 298 | shape = x.shape |
| 299 | x = self.preprocess(x) |
| 300 | embed_ind = self.quantize(x) |
| 301 | embed_ind = self.postprocess_emb(embed_ind, shape) |
| 302 | return embed_ind |
| 303 | |
| 304 | def decode(self, embed_ind): |
| 305 | quantize = self.dequantize(embed_ind) |
| 306 | return quantize |
| 307 | |
| 308 | def forward(self, x): |
| 309 | shape, dtype = x.shape, x.dtype |
| 310 | x = self.preprocess(x) |
| 311 | |
| 312 | self.init_embed_(x) |
| 313 | |
| 314 | embed_ind = self.quantize(x) |
| 315 | embed_onehot = F.one_hot(embed_ind, self.codebook_size).type(dtype) |
| 316 | embed_ind = self.postprocess_emb(embed_ind, shape) |
| 317 | quantize = self.dequantize(embed_ind) |
| 318 | |
| 319 | if self.training: |
| 320 | self.expire_codes_(x) |
| 321 | _ema_inplace(self.cluster_size, embed_onehot.sum(0), self.decay) |
| 322 | embed_sum = x.t() @ embed_onehot |
| 323 | _ema_inplace(self.embed_avg, embed_sum.t().contiguous(), self.decay) |
| 324 | cluster_size = ( |
| 325 | _laplace_smoothing(self.cluster_size, self.codebook_size, self.epsilon) |
| 326 | * self.cluster_size.sum() |
| 327 | ) |
| 328 | embed_normalized = self.embed_avg / cluster_size.unsqueeze(1) |
| 329 | self.embed.data.copy_(embed_normalized) |
| 330 | |
| 331 | return quantize, embed_ind |
| 332 | |
| 333 | |
| 334 | class VectorQuantization(nn.Module): |
| 335 | def __init__( |
| 336 | self, |
| 337 | dim: int, |
| 338 | codebook_size: int, |
| 339 | codebook_dim: int | None = None, |
| 340 | decay: float = 0.99, |
| 341 | epsilon: float = 1e-5, |
| 342 | kmeans_init: bool = True, |
| 343 | kmeans_iters: int = 50, |
| 344 | threshold_ema_dead_code: int = 2, |
| 345 | commitment_weight: float = 1.0, |
| 346 | ): |
| 347 | super().__init__() |
| 348 | _codebook_dim: int = _vq_default(codebook_dim, dim) |
| 349 | |
| 350 | requires_projection = _codebook_dim != dim |
| 351 | self.project_in = ( |
| 352 | nn.Linear(dim, _codebook_dim) if requires_projection else nn.Identity() |
| 353 | ) |
| 354 | self.project_out = ( |
| 355 | nn.Linear(_codebook_dim, dim) if requires_projection else nn.Identity() |
| 356 | ) |
| 357 | |
| 358 | self.epsilon = epsilon |
| 359 | self.commitment_weight = commitment_weight |
| 360 | |
| 361 | self._codebook = EuclideanCodebook( |
| 362 | dim=_codebook_dim, |
| 363 | codebook_size=codebook_size, |
| 364 | kmeans_init=kmeans_init, |
| 365 | kmeans_iters=kmeans_iters, |
| 366 | decay=decay, |
| 367 | epsilon=epsilon, |
| 368 | threshold_ema_dead_code=threshold_ema_dead_code, |
| 369 | ) |
| 370 | self.codebook_size = codebook_size |
| 371 | |
| 372 | @property |
| 373 | def codebook(self): |
| 374 | return self._codebook.embed |
| 375 | |
| 376 | def encode(self, x): |
| 377 | x = self.project_in(x) |
| 378 | embed_in = self._codebook.encode(x) |
| 379 | return embed_in |
| 380 | |
| 381 | def decode(self, embed_ind): |
| 382 | quantize = self._codebook.decode(embed_ind) |
| 383 | quantize = self.project_out(quantize) |
| 384 | return quantize |
| 385 | |
| 386 | def forward(self, x): |
| 387 | device = x.device |
| 388 | x = self.project_in(x) |
| 389 | |
| 390 | quantize, embed_ind = self._codebook(x) |
| 391 | |
| 392 | if self.training: |
| 393 | quantize = x + (quantize - x).detach() |
| 394 | |
| 395 | loss = torch.tensor([0.0], device=device, requires_grad=self.training) |
| 396 | |
| 397 | quantize = self.project_out(quantize) |
| 398 | return quantize, embed_ind, loss |
| 399 | |
| 400 | |
| 401 | class ResidualVectorQuantization(nn.Module): |
| 402 | def __init__(self, *, num_quantizers, codebook_size, **kwargs): |
| 403 | super().__init__() |
| 404 | if isinstance(codebook_size, int): |
| 405 | codebook_size = [codebook_size] * num_quantizers |
| 406 | elif len(codebook_size) < num_quantizers: |
| 407 | codebook_size += [codebook_size[-1]] * (num_quantizers - len(codebook_size)) |
| 408 | self.layers = nn.ModuleList( |
| 409 | [ |
| 410 | VectorQuantization(codebook_size=codebook_size[i], **kwargs) |
| 411 | for i in range(num_quantizers) |
| 412 | ] |
| 413 | ) |
| 414 | |
| 415 | def forward(self, x, n_q: int | None = None, layers: list | None = None): |
| 416 | quantized_out = 0.0 |
| 417 | residual = x |
| 418 | |
| 419 | all_losses = [] |
| 420 | all_indices = [] |
| 421 | out_quantized = [] |
| 422 | |
| 423 | n_q = n_q or len(self.layers) |
| 424 | |
| 425 | for i, layer in enumerate(self.layers[:n_q]): |
| 426 | quantized, indices, loss = layer(residual) |
| 427 | residual = residual - quantized |
| 428 | quantized_out = quantized_out + quantized |
| 429 | |
| 430 | all_indices.append(indices) |
| 431 | all_losses.append(loss) |
| 432 | if layers and i in layers: |
| 433 | out_quantized.append(quantized_out) |
| 434 | |
| 435 | out_losses, out_indices = map(torch.stack, (all_losses, all_indices)) |
| 436 | return quantized_out, out_indices, out_losses, out_quantized |
| 437 | |
| 438 | def encode( |
| 439 | self, x: torch.Tensor, n_q: int | None = None, st: int | None = None |
| 440 | ) -> torch.Tensor: |
| 441 | residual = x |
| 442 | all_indices = [] |
| 443 | n_q = len(self.layers) if n_q is None else n_q |
| 444 | st = 0 if st is None else st |
| 445 | for layer in self.layers[st:n_q]: |
| 446 | indices = layer.encode(residual) |
| 447 | quantized = layer.decode(indices) |
| 448 | residual = residual - quantized |
| 449 | all_indices.append(indices) |
| 450 | out_indices = torch.stack(all_indices) |
| 451 | return out_indices |
| 452 | |
| 453 | def decode(self, q_indices: torch.Tensor, st: int = 0) -> torch.Tensor: |
| 454 | quantized_out = self.layers[st].decode(q_indices[0]) |
| 455 | for i in range(1, len(q_indices)): |
| 456 | layer = self.layers[st + i] |
| 457 | quantized = layer.decode(q_indices[i]) |
| 458 | quantized_out = quantized_out + quantized |
| 459 | return quantized_out |
| 460 | |
| 461 | |
| 462 | class ResidualVectorQuantizer(nn.Module): |
| 463 | def __init__( |
| 464 | self, |
| 465 | dimension: int = 256, |
| 466 | n_q: int = 8, |
| 467 | bins: int | list = 1024, |
| 468 | decay: float = 0.99, |
| 469 | kmeans_init: bool = True, |
| 470 | kmeans_iters: int = 50, |
| 471 | threshold_ema_dead_code: int = 2, |
| 472 | ): |
| 473 | super().__init__() |
| 474 | self.n_q = n_q |
| 475 | self.dimension = dimension |
| 476 | self.bins = bins |
| 477 | self.decay = decay |
| 478 | self.kmeans_init = kmeans_init |
| 479 | self.kmeans_iters = kmeans_iters |
| 480 | self.threshold_ema_dead_code = threshold_ema_dead_code |
| 481 | self.vq = ResidualVectorQuantization( |
| 482 | dim=self.dimension, |
| 483 | codebook_size=self.bins, |
| 484 | num_quantizers=self.n_q, |
| 485 | decay=self.decay, |
| 486 | kmeans_init=self.kmeans_init, |
| 487 | kmeans_iters=self.kmeans_iters, |
| 488 | threshold_ema_dead_code=self.threshold_ema_dead_code, |
| 489 | ) |
| 490 | |
| 491 | def forward( |
| 492 | self, |
| 493 | x: torch.Tensor, |
| 494 | n_q: int | None = None, |
| 495 | layers: list | None = None, |
| 496 | ): |
| 497 | n_q = n_q if n_q else self.n_q |
| 498 | quantized, codes, commit_loss, quantized_list = self.vq( |
| 499 | x, n_q=n_q, layers=layers |
| 500 | ) |
| 501 | return quantized, codes, torch.mean(commit_loss), quantized_list |
| 502 | |
| 503 | def encode( |
| 504 | self, x: torch.Tensor, n_q: int | None = None, st: int | None = None |
| 505 | ) -> torch.Tensor: |
| 506 | n_q = n_q if n_q else self.n_q |
| 507 | st = st or 0 |
| 508 | codes = self.vq.encode(x, n_q=n_q, st=st) |
| 509 | return codes |
| 510 | |
| 511 | def decode(self, codes: torch.Tensor, st: int = 0) -> torch.Tensor: |
| 512 | quantized = self.vq.decode(codes, st=st) |
| 513 | return quantized |
| 514 | |
| 515 | |
| 516 | # --------------------------------------------------------------------------- |
| 517 | # Audio tokenizer |
| 518 | # --------------------------------------------------------------------------- |
| 519 | |
| 520 | |
| 521 | class MiMoAudioTokenizerConfig(PreTrainedConfig): |
| 522 | model_type = "mimo_audio_tokenizer" |
| 523 | |
| 524 | def __init__( |
| 525 | self, |
| 526 | max_audio_seconds: int = 1800, |
| 527 | stride_size: int = 2, |
| 528 | avg_pooler: int = 1, |
| 529 | d_model: int = 768, |
| 530 | scale_embedding: bool = True, |
| 531 | kernel_size: int = 3, |
| 532 | activation_function: str = "gelu", |
| 533 | encoder_layers: int = 8, |
| 534 | encoder_skip_layer_id: int | None = None, |
| 535 | encoder_attention_heads: int = 12, |
| 536 | encoder_ffn_dim: int = 3072, |
| 537 | encoder_causal: bool = False, |
| 538 | encoder_attn_window_size: list[int] | None = None, |
| 539 | decoder_layers: int = 8, |
| 540 | decoder_attention_heads: int = 12, |
| 541 | decoder_ffn_dim: int = 3072, |
| 542 | decoder_kernel_size: int = 3, |
| 543 | decoder_stride_size: int = 2, |
| 544 | decoder_causal: bool = True, |
| 545 | decoder_attn_window_size: list[int] | None = None, |
| 546 | nfft: int = 1024, |
| 547 | vocoder_dim: int = 512, |
| 548 | vocoder_intermediate_dim: int = 4096, |
| 549 | vocoder_num_layers: int = 30, |
| 550 | n_mels: int = 80, |
| 551 | sampling_rate: int = 24000, |
| 552 | hop_length: int = 240, |
| 553 | window_size: int = 1024, |
| 554 | vocoder_padding: str = "same", |
| 555 | fmin: int = 0, |
| 556 | fmax: int | None = None, |
| 557 | num_quantizers: int = 12, |
| 558 | codebook_size: list[int] | None = None, |
| 559 | threshold_ema_dead_code: int = 10, |
| 560 | position_embedding_type: str = "rope", |
| 561 | rope_theta: int = 10000, |
| 562 | rope_type: str = "default", |
| 563 | ln_type: str = "LayerNorm", |
| 564 | vocoder_attention_heads: int = 4, |
| 565 | vocoder_attn_window_size: list[int] | None = None, |
| 566 | use_istft_only: bool = False, |
| 567 | hybrid_attention: bool = False, |
| 568 | hybrid_block_size: int = 8, |
| 569 | swa_per_block: int = 2, |
| 570 | **kwargs, |
| 571 | ): |
| 572 | super().__init__(**kwargs) |
| 573 | self.max_audio_seconds = max_audio_seconds |
| 574 | self.stride_size = stride_size |
| 575 | self.avg_pooler = avg_pooler |
| 576 | self.d_model = d_model |
| 577 | self.scale_embedding = scale_embedding |
| 578 | self.kernel_size = kernel_size |
| 579 | self.activation_function = activation_function |
| 580 | self.encoder_layers = encoder_layers |
| 581 | self.encoder_skip_layer_id = encoder_skip_layer_id |
| 582 | self.encoder_attention_heads = encoder_attention_heads |
| 583 | self.encoder_ffn_dim = encoder_ffn_dim |
| 584 | self.encoder_causal = encoder_causal |
| 585 | self.encoder_attn_window_size = ( |
| 586 | encoder_attn_window_size |
| 587 | if encoder_attn_window_size is not None |
| 588 | else [-1, -1] |
| 589 | ) |
| 590 | self.decoder_layers = decoder_layers |
| 591 | self.decoder_attention_heads = decoder_attention_heads |
| 592 | self.decoder_ffn_dim = decoder_ffn_dim |
| 593 | self.decoder_kernel_size = decoder_kernel_size |
| 594 | self.decoder_stride_size = decoder_stride_size |
| 595 | self.decoder_causal = decoder_causal |
| 596 | self.decoder_attn_window_size = ( |
| 597 | decoder_attn_window_size |
| 598 | if decoder_attn_window_size is not None |
| 599 | else [-1, -1] |
| 600 | ) |
| 601 | self.nfft = nfft |
| 602 | self.vocoder_dim = vocoder_dim |
| 603 | self.vocoder_intermediate_dim = vocoder_intermediate_dim |
| 604 | self.vocoder_num_layers = vocoder_num_layers |
| 605 | self.n_mels = n_mels |
| 606 | self.sampling_rate = sampling_rate |
| 607 | self.hop_length = hop_length |
| 608 | self.window_size = window_size |
| 609 | self.vocoder_padding = vocoder_padding |
| 610 | self.fmin = fmin |
| 611 | self.fmax = fmax |
| 612 | self.num_quantizers = num_quantizers |
| 613 | self.codebook_size = codebook_size if codebook_size is not None else [1024] |
| 614 | self.threshold_ema_dead_code = threshold_ema_dead_code |
| 615 | self.position_embedding_type = position_embedding_type |
| 616 | self.rope_theta = rope_theta |
| 617 | self.rope_type = rope_type |
| 618 | self.ln_type = ln_type |
| 619 | self.vocoder_attention_heads = vocoder_attention_heads |
| 620 | self.vocoder_attn_window_size = ( |
| 621 | vocoder_attn_window_size |
| 622 | if vocoder_attn_window_size is not None |
| 623 | else [40, 10] |
| 624 | ) |
| 625 | self.use_istft_only = use_istft_only |
| 626 | self.hybrid_attention = hybrid_attention |
| 627 | self.hybrid_block_size = hybrid_block_size |
| 628 | self.swa_per_block = swa_per_block |
| 629 | |
| 630 | |
| 631 | def get_sequence_mask(inputs, inputs_length): |
| 632 | if inputs.dim() == 3: |
| 633 | bsz, tgt_len, _ = inputs.size() |
| 634 | else: |
| 635 | bsz, tgt_len = inputs_length.shape[0], torch.max(inputs_length) |
| 636 | sequence_mask = torch.arange(0, tgt_len).to(inputs.device) |
| 637 | sequence_mask = torch.lt(sequence_mask, inputs_length.reshape(bsz, 1)).view( |
| 638 | bsz, tgt_len, 1 |
| 639 | ) |
| 640 | unpacking_index = torch.cumsum(sequence_mask.to(torch.int64).view(-1), dim=0) - 1 |
| 641 | return sequence_mask, unpacking_index |
| 642 | |
| 643 | |
| 644 | def unpack_hidden_states( |
| 645 | hidden_states, lengths, sequence_mask=None, unpacking_index=None |
| 646 | ): |
| 647 | bsz = lengths.shape[0] |
| 648 | if sequence_mask is None or unpacking_index is None: |
| 649 | sequence_mask, unpacking_index = get_sequence_mask(hidden_states, lengths) |
| 650 | hidden_states = torch.index_select(hidden_states, 0, unpacking_index).view( |
| 651 | bsz, torch.max(lengths), hidden_states.shape[-1] |
| 652 | ) |
| 653 | return torch.where(sequence_mask, hidden_states, 0) |
| 654 | |
| 655 | |
| 656 | def get_position_ids(lengths): |
| 657 | total_len = lengths.sum() |
| 658 | offset = torch.cat([torch.zeros(1).to(lengths), lengths[:-1].cumsum(dim=0)]) |
| 659 | offset = torch.repeat_interleave(offset, lengths) |
| 660 | return torch.arange(0, total_len).to(offset) - offset |
| 661 | |
| 662 | |
| 663 | LAYER_NORM = {"LayerNorm": nn.LayerNorm} |
| 664 | |
| 665 | |
| 666 | class AudioEncoderAttention(nn.Module): |
| 667 | def __init__( |
| 668 | self, |
| 669 | embed_dim: int, |
| 670 | num_heads: int, |
| 671 | window_size: tuple[int, int] = (-1, -1), |
| 672 | causal: bool = False, |
| 673 | ): |
| 674 | super().__init__() |
| 675 | self.embed_dim = embed_dim |
| 676 | self.num_heads = num_heads |
| 677 | self.head_dim = embed_dim // num_heads |
| 678 | self.window_size = window_size |
| 679 | self.causal = causal |
| 680 | |
| 681 | self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False) |
| 682 | self.v_proj = nn.Linear(embed_dim, embed_dim, bias=True) |
| 683 | self.q_proj = nn.Linear(embed_dim, embed_dim, bias=True) |
| 684 | self.out_proj = nn.Linear(embed_dim, embed_dim, bias=True) |
| 685 | |
| 686 | def forward( |
| 687 | self, |
| 688 | hidden_states: torch.Tensor, |
| 689 | cu_seqlens: torch.Tensor, |
| 690 | max_seqlen: int, |
| 691 | rope_position_embeddings=None, |
| 692 | ): |
| 693 | from vllm.vllm_flash_attn import flash_attn_varlen_func |
| 694 | |
| 695 | bsz, _ = hidden_states.size() |
| 696 | |
| 697 | query_states = self.q_proj(hidden_states).view( |
| 698 | bsz, self.num_heads, self.head_dim |
| 699 | ) |
| 700 | key_states = self.k_proj(hidden_states).view(bsz, self.num_heads, self.head_dim) |
| 701 | value_states = self.v_proj(hidden_states).view( |
| 702 | bsz, self.num_heads, self.head_dim |
| 703 | ) |
| 704 | |
| 705 | if rope_position_embeddings is not None: |
| 706 | cos, sin = rope_position_embeddings |
| 707 | query_states, key_states = apply_rotary_pos_emb( |
| 708 | query_states, key_states, cos, sin |
| 709 | ) |
| 710 | |
| 711 | attn_output = flash_attn_varlen_func( |
| 712 | query_states, |
| 713 | key_states, |
| 714 | value_states, |
| 715 | cu_seqlens_q=cu_seqlens, |
| 716 | cu_seqlens_k=cu_seqlens, |
| 717 | max_seqlen_q=max_seqlen, |
| 718 | max_seqlen_k=max_seqlen, |
| 719 | causal=self.causal, |
| 720 | window_size=list(self.window_size), |
| 721 | ) |
| 722 | |
| 723 | attn_output = attn_output.reshape(bsz, self.embed_dim) |
| 724 | attn_output = self.out_proj(attn_output) |
| 725 | return attn_output |
| 726 | |
| 727 | |
| 728 | class AudioEncoderTransformerLayer(nn.Module): |
| 729 | def __init__( |
| 730 | self, |
| 731 | config: MiMoAudioTokenizerConfig, |
| 732 | causal: bool, |
| 733 | attn_window_size: tuple[int, int] = (-1, -1), |
| 734 | ): |
| 735 | super().__init__() |
| 736 | self.embed_dim = config.d_model |
| 737 | |
| 738 | self.self_attn = AudioEncoderAttention( |
| 739 | embed_dim=self.embed_dim, |
| 740 | num_heads=config.encoder_attention_heads, |
| 741 | window_size=attn_window_size, |
| 742 | causal=causal, |
| 743 | ) |
| 744 | self.self_attn_layer_norm = LAYER_NORM[config.ln_type](self.embed_dim) |
| 745 | |
| 746 | self.activation_fn = ACT2FN[config.activation_function] |
| 747 | self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim) |
| 748 | self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim) |
| 749 | self.final_layer_norm = LAYER_NORM[config.ln_type](self.embed_dim) |
| 750 | |
| 751 | def forward( |
| 752 | self, |
| 753 | hidden_states: torch.Tensor, |
| 754 | cu_seqlens: torch.Tensor, |
| 755 | max_seqlen: int, |
| 756 | rope_position_embeddings: tuple[torch.Tensor, torch.Tensor], |
| 757 | ) -> torch.Tensor: |
| 758 | residual = hidden_states |
| 759 | hidden_states = self.self_attn_layer_norm(hidden_states) |
| 760 | hidden_states = self.self_attn( |
| 761 | hidden_states, |
| 762 | cu_seqlens, |
| 763 | max_seqlen, |
| 764 | rope_position_embeddings=rope_position_embeddings, |
| 765 | ) |
| 766 | hidden_states = residual + hidden_states |
| 767 | |
| 768 | residual = hidden_states |
| 769 | hidden_states = self.final_layer_norm(hidden_states) |
| 770 | hidden_states = self.activation_fn(self.fc1(hidden_states)) |
| 771 | hidden_states = self.fc2(hidden_states) |
| 772 | hidden_states = residual + hidden_states |
| 773 | |
| 774 | return hidden_states |
| 775 | |
| 776 | |
| 777 | class AudioEncoder(nn.Module): |
| 778 | def __init__( |
| 779 | self, |
| 780 | config: MiMoAudioTokenizerConfig, |
| 781 | ): |
| 782 | super().__init__() |
| 783 | self.config = config |
| 784 | self.max_source_positions = ( |
| 785 | config.max_audio_seconds * config.sampling_rate // config.hop_length |
| 786 | ) // config.stride_size |
| 787 | self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0 |
| 788 | self.skip_layer_idx = config.encoder_skip_layer_id |
| 789 | |
| 790 | self.conv1 = nn.Conv1d( |
| 791 | config.n_mels, |
| 792 | config.d_model, |
| 793 | kernel_size=config.kernel_size, |
| 794 | padding=1, |
| 795 | ) |
| 796 | self.conv2 = nn.Conv1d( |
| 797 | config.d_model, |
| 798 | config.d_model, |
| 799 | kernel_size=config.kernel_size, |
| 800 | stride=config.stride_size, |
| 801 | padding=1, |
| 802 | ) |
| 803 | |
| 804 | self.position_embedding = AudioRotaryEmbedding( |
| 805 | config.rope_theta, |
| 806 | config.d_model // config.encoder_attention_heads, |
| 807 | self.max_source_positions, |
| 808 | config.rope_type, |
| 809 | ) |
| 810 | |
| 811 | encoder_attn_window_size = config.encoder_attn_window_size |
| 812 | assert encoder_attn_window_size is not None |
| 813 | assert len(encoder_attn_window_size) == 2 |
| 814 | attn_window_size = ( |
| 815 | encoder_attn_window_size[0], |
| 816 | encoder_attn_window_size[1], |
| 817 | ) |
| 818 | attn_window_sizes: list[tuple[int, int]] = [] |
| 819 | if config.hybrid_attention: |
| 820 | for i in range(config.encoder_layers): |
| 821 | if i % config.swa_per_block < config.swa_per_block - 1: |
| 822 | attn_window_sizes.append(attn_window_size) |
| 823 | else: |
| 824 | attn_window_sizes.append((-1, -1)) |
| 825 | else: |
| 826 | attn_window_sizes = [attn_window_size] * config.encoder_layers |
| 827 | |
| 828 | self.layers = nn.ModuleList( |
| 829 | [ |
| 830 | AudioEncoderTransformerLayer( |
| 831 | config=config, |
| 832 | causal=config.encoder_causal, |
| 833 | attn_window_size=attn_window_sizes[i], |
| 834 | ) |
| 835 | for i in range(config.encoder_layers) |
| 836 | ] |
| 837 | ) |
| 838 | |
| 839 | self.layer_norm = LAYER_NORM[config.ln_type](config.d_model) |
| 840 | |
| 841 | if config.avg_pooler != 1: |
| 842 | self.down_sample_layer = nn.Sequential( |
| 843 | nn.Conv1d( |
| 844 | config.d_model, |
| 845 | config.d_model, |
| 846 | config.avg_pooler, |
| 847 | config.avg_pooler, |
| 848 | bias=False, |
| 849 | ), |
| 850 | nn.GELU(), |
| 851 | ) |
| 852 | self.down_sample_norm = LAYER_NORM[config.ln_type](config.d_model) |
| 853 | else: |
| 854 | self.down_sample_layer = None |
| 855 | |
| 856 | self.quantizer: ResidualVectorQuantizer | None |
| 857 | if config.num_quantizers != 0: |
| 858 | self.quantizer = ResidualVectorQuantizer( |
| 859 | dimension=config.d_model, |
| 860 | n_q=config.num_quantizers, |
| 861 | bins=config.codebook_size, |
| 862 | threshold_ema_dead_code=config.threshold_ema_dead_code, |
| 863 | ) |
| 864 | else: |
| 865 | self.quantizer = None |
| 866 | |
| 867 | def get_features(self, input_features, output_length): |
| 868 | input_features = input_features.to(self.conv1.weight) |
| 869 | inputs_embeds = nn.functional.gelu(self.conv1(input_features)) |
| 870 | inputs_embeds = nn.functional.gelu(self.conv2(inputs_embeds)) |
| 871 | inputs_embeds = inputs_embeds.permute(0, 2, 1) |
| 872 | bsz, tgt_len, _ = inputs_embeds.size() |
| 873 | hidden_states = inputs_embeds |
| 874 | |
| 875 | position_ids = get_position_ids(output_length).long().to(input_features.device) |
| 876 | rope_position_embeddings = self.position_embedding(input_features, position_ids) |
| 877 | |
| 878 | attention_mask, unpacking_index = get_sequence_mask( |
| 879 | hidden_states, output_length |
| 880 | ) |
| 881 | hidden_states = torch.masked_select(hidden_states, attention_mask).view( |
| 882 | torch.sum(output_length), self.config.d_model |
| 883 | ) |
| 884 | |
| 885 | cu_seqlens = F.pad( |
| 886 | torch.cumsum(output_length, dim=0), (1, 0), "constant", 0 |
| 887 | ).to(device=hidden_states.device, dtype=torch.int32) |
| 888 | max_seqlen = torch.max(output_length).to(torch.int32).item() |
| 889 | |
| 890 | skip_connect_hidden_states = 0.0 |
| 891 | for idx, encoder_layer in enumerate(self.layers): |
| 892 | hidden_states = encoder_layer( |
| 893 | hidden_states, |
| 894 | cu_seqlens, |
| 895 | max_seqlen, |
| 896 | rope_position_embeddings=rope_position_embeddings, |
| 897 | ) |
| 898 | if (self.skip_layer_idx is not None) and idx == self.skip_layer_idx - 1: |
| 899 | skip_connect_hidden_states = hidden_states.clone() |
| 900 | |
| 901 | hidden_states += skip_connect_hidden_states |
| 902 | hidden_states = self.layer_norm(hidden_states) |
| 903 | |
| 904 | if self.down_sample_layer is not None: |
| 905 | hidden_states = torch.index_select(hidden_states, 0, unpacking_index).view( |
| 906 | bsz, tgt_len, self.config.d_model |
| 907 | ) |
| 908 | if hidden_states.size(1) % self.config.avg_pooler: |
| 909 | pad_len = ( |
| 910 | self.config.avg_pooler |
| 911 | - hidden_states.size(1) % self.config.avg_pooler |
| 912 | ) |
| 913 | hidden_states = torch.nn.functional.pad( |
| 914 | hidden_states, (0, 0, 0, pad_len), mode="constant", value=0.0 |
| 915 | ) |
| 916 | tgt_len += pad_len |
| 917 | tgt_len = tgt_len // self.config.avg_pooler |
| 918 | hidden_states = self.down_sample_layer(hidden_states.transpose(1, 2)) |
| 919 | output_length = ( |
| 920 | output_length // self.config.avg_pooler |
| 921 | + (output_length % self.config.avg_pooler != 0).int() |
| 922 | ) |
| 923 | hidden_states = hidden_states.transpose(1, 2) |
| 924 | attention_mask, unpacking_index = get_sequence_mask( |
| 925 | hidden_states, output_length |
| 926 | ) |
| 927 | hidden_states = torch.masked_select(hidden_states, attention_mask).view( |
| 928 | torch.sum(output_length), self.config.d_model |
| 929 | ) |
| 930 | hidden_states = self.down_sample_norm(hidden_states) |
| 931 | |
| 932 | return ( |
| 933 | hidden_states, |
| 934 | output_length, |
| 935 | attention_mask, |
| 936 | unpacking_index, |
| 937 | tgt_len, |
| 938 | bsz, |
| 939 | ) |
| 940 | |
| 941 | def get_output_length(self, mel_len): |
| 942 | tgt_len = mel_len + 3 - self.config.kernel_size |
| 943 | return (tgt_len + 2 - self.config.kernel_size) // self.config.stride_size + 1 |
| 944 | |
| 945 | @torch.no_grad() |
| 946 | def encode( |
| 947 | self, |
| 948 | input_features, |
| 949 | input_lens=None, |
| 950 | output_length=None, |
| 951 | return_codes_only=False, |
| 952 | n_q=None, |
| 953 | use_quantizer=True, |
| 954 | ): |
| 955 | if output_length is None: |
| 956 | output_length = self.get_output_length(input_lens) |
| 957 | input_features = unpack_hidden_states(input_features, input_lens) |
| 958 | hidden_states, output_length, attention_mask, unpacking_index, tgt_len, bsz = ( |
| 959 | self.get_features( |
| 960 | input_features=input_features.transpose(1, 2), |
| 961 | output_length=output_length, |
| 962 | ) |
| 963 | ) |
| 964 | |
| 965 | dtype = hidden_states.dtype |
| 966 | if use_quantizer and self.quantizer is not None: |
| 967 | self.quantizer.float() |
| 968 | codes = self.quantizer.encode(hidden_states.float(), n_q=n_q) |
| 969 | if return_codes_only: |
| 970 | return codes, output_length |
| 971 | hidden_states = self.quantizer.decode(codes) |
| 972 | hidden_states = hidden_states.to(dtype) |
| 973 | else: |
| 974 | codes = None |
| 975 | |
| 976 | hidden_states_packed = hidden_states.clone() |
| 977 | hidden_states = torch.index_select(hidden_states, 0, unpacking_index).view( |
| 978 | bsz, tgt_len, self.config.d_model |
| 979 | ) |
| 980 | hidden_states = torch.where(attention_mask, hidden_states, 0) |
| 981 | return hidden_states, hidden_states_packed, output_length, codes |
| 982 | |
| 983 | @torch.no_grad() |
| 984 | def decode_vq(self, codes): |
| 985 | assert self.quantizer is not None |
| 986 | self.quantizer.float() |
| 987 | return self.quantizer.decode(codes) |
| 988 | |
| 989 | |
| 990 | class MiMoAudioTokenizer(PreTrainedModel): |
| 991 | config_class = MiMoAudioTokenizerConfig |
| 992 | |
| 993 | def __init__(self, config: MiMoAudioTokenizerConfig): |
| 994 | super().__init__(config) |
| 995 | self.config = config |
| 996 | self.sampling_rate = config.sampling_rate |
| 997 | self.encoder = AudioEncoder(config=config) |
| 998 | self.downsample_rate = int(config.hop_length * 2 * config.avg_pooler) |
| 999 | |
| 1000 | def get_output_length(self, mel_len): |
| 1001 | tgt_len = mel_len + 3 - self.config.kernel_size |
| 1002 | return (tgt_len + 2 - self.config.kernel_size) // self.config.stride_size + 1 |
| 1003 | |
| 1004 | @torch.no_grad() |
| 1005 | def encode(self, mels, input_lens, use_quantizer=True): |
| 1006 | input_features = mels |
| 1007 | encoder_output_length = self.get_output_length(input_lens) |
| 1008 | hidden_states, hidden_states_packed, encoder_output_length, codes = ( |
| 1009 | self.encoder.encode( |
| 1010 | input_features, input_lens=input_lens, use_quantizer=use_quantizer |
| 1011 | ) |
| 1012 | ) |
| 1013 | return hidden_states, hidden_states_packed, encoder_output_length, codes |
| 1014 | |
| 1015 | |
| 1016 | # --------------------------------------------------------------------------- |
| 1017 | # Audio encoding utilities |
| 1018 | # --------------------------------------------------------------------------- |
| 1019 | |
| 1020 | |
| 1021 | def group_by_length(features: torch.Tensor, lengths: torch.Tensor, max_length: int): |
| 1022 | if features.size(0) != lengths.sum().item(): |
| 1023 | raise ValueError( |
| 1024 | f"Feature size mismatch: {features.size(0)} vs {lengths.sum().item()}" |
| 1025 | ) |
| 1026 | |
| 1027 | split_points = [] |
| 1028 | current_sum = 0 |
| 1029 | |
| 1030 | for i, seq_len in enumerate(lengths): |
| 1031 | if current_sum + seq_len > max_length and current_sum > 0: |
| 1032 | split_points.append(i) |
| 1033 | current_sum = seq_len.item() |
| 1034 | else: |
| 1035 | current_sum += seq_len.item() |
| 1036 | |
| 1037 | group_sizes = [] |
| 1038 | prev = 0 |
| 1039 | for point in split_points: |
| 1040 | group_sizes.append(point - prev) |
| 1041 | prev = point |
| 1042 | if prev < len(lengths): |
| 1043 | group_sizes.append(len(lengths) - prev) |
| 1044 | |
| 1045 | len_groups = torch.split(lengths, group_sizes) |
| 1046 | feature_sizes = [group.sum().item() for group in len_groups] |
| 1047 | feature_groups = torch.split(features, feature_sizes) |
| 1048 | |
| 1049 | return feature_groups, len_groups |
| 1050 | |
| 1051 | |
| 1052 | @torch.no_grad() |
| 1053 | def encode_batch( |
| 1054 | audio_tokenizer_encoder, |
| 1055 | input_features: torch.Tensor, |
| 1056 | input_lens: torch.Tensor, |
| 1057 | max_length: int = 256000, |
| 1058 | ): |
| 1059 | feature_groups, len_groups = group_by_length(input_features, input_lens, max_length) |
| 1060 | |
| 1061 | encoded_parts = [] |
| 1062 | for features, lengths in zip(feature_groups, len_groups): |
| 1063 | codes, _ = audio_tokenizer_encoder.encode( |
| 1064 | input_features=features, input_lens=lengths, return_codes_only=True |
| 1065 | ) |
| 1066 | encoded_parts.append(codes) |
| 1067 | |
| 1068 | return torch.cat(encoded_parts, dim=-1) |
| 1069 | |
| 1070 | |
| 1071 | def _segment_lengths_for_mel(mel: torch.Tensor, segment_size: int): |
| 1072 | """Split mel into segments of segment_size with a possible shorter remainder.""" |
| 1073 | input_len = mel.size(0) |
| 1074 | segs = [segment_size] * (input_len // segment_size) |
| 1075 | if input_len % segment_size > 0: |
| 1076 | segs.append(input_len % segment_size) |
| 1077 | return segs |
| 1078 | |
| 1079 | |
| 1080 | @torch.no_grad() |
| 1081 | def tokenize_audio_batch(mels, audio_tokenizer_encoder, segment_size=6000, device=None): |
| 1082 | """Tokenize multiple mels in one encode_batch call. |
| 1083 | |
| 1084 | Returns list of code tensors, each [T_i, C] for that mel. |
| 1085 | """ |
| 1086 | if not mels: |
| 1087 | return [] |
| 1088 | if device is None: |
| 1089 | device = next(audio_tokenizer_encoder.parameters()).device |
| 1090 | input_len_seg_per_mel = [_segment_lengths_for_mel(m, segment_size) for m in mels] |
| 1091 | input_lens_flat = [s for segs in input_len_seg_per_mel for s in segs] |
| 1092 | input_features = torch.cat([m.to(device) for m in mels], dim=0) |
| 1093 | input_lens_t = torch.tensor(input_lens_flat, dtype=torch.long, device=device) |
| 1094 | codes_packed = encode_batch( |
| 1095 | audio_tokenizer_encoder, |
| 1096 | input_features=input_features, |
| 1097 | input_lens=input_lens_t, |
| 1098 | ) |
| 1099 | codes = codes_packed.transpose(0, 1).detach() # [total_code_T, C] |
| 1100 | code_lengths = [] |
| 1101 | for segs in input_len_seg_per_mel: |
| 1102 | out_len = audio_tokenizer_encoder.get_output_length( |
| 1103 | torch.tensor(segs, dtype=torch.long, device=device) |
| 1104 | ) |
| 1105 | if getattr(audio_tokenizer_encoder, "down_sample_layer", None) is not None: |
| 1106 | avg = audio_tokenizer_encoder.config.avg_pooler |
| 1107 | out_len = out_len // avg + (out_len % avg != 0).long() |
| 1108 | code_lengths.append(out_len.sum().item()) |
| 1109 | code_list = torch.split(codes, code_lengths) |
| 1110 | return list(code_list) |
| 1111 | |
| 1112 | |
| 1113 | # --------------------------------------------------------------------------- |
| 1114 | # MimoAudioEncoderConfig |
| 1115 | # --------------------------------------------------------------------------- |
| 1116 | |
| 1117 | |
| 1118 | @dataclass |
| 1119 | class MimoAudioEncoderConfig: |
| 1120 | """Config for MimoAudioEncoder. |
| 1121 | |
| 1122 | Field names match the audio_config dict in the model checkpoint. |
| 1123 | """ |
| 1124 | |
| 1125 | speech_vocab_size: str = "1025-1025-129-129-129-129-129-129" |
| 1126 | speech_zeroemb_idx: str = "1024-1024-128-128-128-128-128-128" |
| 1127 | group_size: int = 4 |
| 1128 | audio_channels: int = 8 |
| 1129 | input_local_layers: int = 6 |
| 1130 | input_local_dim: int = 1024 |
| 1131 | input_full_attention: bool = True |
| 1132 | input_local_attn_heads: int = 64 |
| 1133 | input_local_head_dim: int = 16 |
| 1134 | input_local_intermediate_size: int = 4096 |
| 1135 | input_local_hidden_dropout: float = 0.0 |
| 1136 | out_hidden_size: int = 4096 |
| 1137 | rope_theta: float = 640000.0 |
| 1138 | partial_rotary_factor: float = 0.334 |
| 1139 | projection_layers: int = 1 |
| 1140 | add_post_norm: bool = False |
| 1141 | audio_segment_size: int = 6000 |
| 1142 | |
| 1143 | @classmethod |
| 1144 | def from_dict(cls, d: dict) -> "MimoAudioEncoderConfig": |
| 1145 | known = {f.name for f in dataclasses.fields(cls)} |
| 1146 | return cls(**{k: v for k, v in d.items() if k in known}) |
| 1147 | |
| 1148 | |
| 1149 | # --------------------------------------------------------------------------- |
| 1150 | # AudioProjection |
| 1151 | # --------------------------------------------------------------------------- |
| 1152 | |
| 1153 | |
| 1154 | class AudioProjection(nn.Module): |
| 1155 | def __init__( |
| 1156 | self, |
| 1157 | input_size: int, |
| 1158 | hidden_size: int, |
| 1159 | output_size: int, |
| 1160 | ) -> None: |
| 1161 | super().__init__() |
| 1162 | self.mlp = nn.Sequential( |
| 1163 | nn.Linear(input_size, hidden_size, bias=False), |
| 1164 | nn.GELU(), |
| 1165 | nn.Linear(hidden_size, output_size, bias=False), |
| 1166 | ) |
| 1167 | |
| 1168 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 1169 | return self.mlp(x) |
| 1170 | |
| 1171 | |
| 1172 | # --------------------------------------------------------------------------- |
| 1173 | # MimoAudioEncoder |
| 1174 | # --------------------------------------------------------------------------- |
| 1175 | |
| 1176 | |
| 1177 | class MimoAudioEncoder(nn.Module): |
| 1178 | """Audio encoder for MiMo-V2-Omni. |
| 1179 | |
| 1180 | Encodes mel spectrograms into LLM-compatible embeddings via: |
| 1181 | 1. Audio tokenizer (VQ codes) |
| 1182 | 2. Speech embeddings lookup |
| 1183 | 3. Local Qwen2 transformer |
| 1184 | 4. Linear projection |
| 1185 | """ |
| 1186 | |
| 1187 | def __init__(self, config, model_path: str = "") -> None: |
| 1188 | super().__init__() |
| 1189 | if isinstance(config, dict): |
| 1190 | config = MimoAudioEncoderConfig.from_dict(config) |
| 1191 | self.config = config |
| 1192 | self.audio_channels = config.audio_channels |
| 1193 | self.audio_group_size = config.group_size |
| 1194 | self.audio_segment_size = config.audio_segment_size |
| 1195 | |
| 1196 | speech_vocab_sizes = self._parse_maybe_list( |
| 1197 | config.speech_vocab_size, config.audio_channels |
| 1198 | ) |
| 1199 | speech_empty_ids = self._parse_maybe_list( |
| 1200 | config.speech_zeroemb_idx, config.audio_channels |
| 1201 | ) |
| 1202 | |
| 1203 | input_local_config = Qwen2Config( |
| 1204 | hidden_size=config.input_local_dim, |
| 1205 | num_hidden_layers=config.input_local_layers, |
| 1206 | num_attention_heads=config.input_local_attn_heads, |
| 1207 | num_key_value_heads=config.input_local_attn_heads, |
| 1208 | intermediate_size=config.input_local_intermediate_size, |
| 1209 | attention_dropout=config.input_local_hidden_dropout, |
| 1210 | rope_theta=config.rope_theta, |
| 1211 | partial_rotary_factor=config.partial_rotary_factor, |
| 1212 | ) |
| 1213 | |
| 1214 | self.input_local_transformer = Qwen2Model(input_local_config) |
| 1215 | |
| 1216 | if not config.add_post_norm: |
| 1217 | self.input_local_transformer.norm = nn.Identity() |
| 1218 | |
| 1219 | self.speech_embeddings = nn.ModuleList( |
| 1220 | [ |
| 1221 | nn.Embedding( |
| 1222 | speech_vocab_sizes[i], |
| 1223 | config.input_local_dim, |
| 1224 | padding_idx=speech_empty_ids[i], |
| 1225 | ) |
| 1226 | for i in range(config.audio_channels) |
| 1227 | ] |
| 1228 | ) |
| 1229 | |
| 1230 | if config.projection_layers == 1: |
| 1231 | self.projection = nn.Linear( |
| 1232 | config.input_local_dim * config.group_size, |
| 1233 | config.out_hidden_size, |
| 1234 | bias=False, |
| 1235 | ) |
| 1236 | elif config.projection_layers == 2: |
| 1237 | self.projection = AudioProjection( |
| 1238 | config.input_local_dim * config.group_size, |
| 1239 | config.input_local_dim * config.group_size * 4, |
| 1240 | config.out_hidden_size, |
| 1241 | ) |
| 1242 | else: |
| 1243 | raise ValueError(f"Invalid projection_layers: {config.projection_layers}") |
| 1244 | |
| 1245 | self.audio_tokenizer: MiMoAudioTokenizer | None = None |
| 1246 | if model_path: |
| 1247 | audio_tokenizer_path = os.path.join(model_path, "audio_tokenizer") |
| 1248 | if os.path.exists(audio_tokenizer_path): |
| 1249 | dev = torch.get_default_device() |
| 1250 | self.audio_tokenizer = self._load_audio_tokenizer( |
| 1251 | audio_tokenizer_path, dev |
| 1252 | ) |
| 1253 | else: |
| 1254 | logger.warning( |
| 1255 | "Audio tokenizer not found at %s, audio encoding disabled", |
| 1256 | audio_tokenizer_path, |
| 1257 | ) |
| 1258 | |
| 1259 | @staticmethod |
| 1260 | def _load_audio_tokenizer(path: str, device: torch.device) -> MiMoAudioTokenizer: |
| 1261 | """Load MiMoAudioTokenizer from directory.""" |
| 1262 | from safetensors.torch import load_file |
| 1263 | |
| 1264 | config_path = os.path.join(path, "config.json") |
| 1265 | with open(config_path) as f: |
| 1266 | config_dict = json.load(f) |
| 1267 | config = MiMoAudioTokenizer.config_class(**config_dict) |
| 1268 | model = MiMoAudioTokenizer(config) |
| 1269 | safetensors_path = os.path.join(path, "model.safetensors") |
| 1270 | bin_path = os.path.join(path, "pytorch_model.bin") |
| 1271 | if os.path.exists(safetensors_path): |
| 1272 | state_dict = load_file(safetensors_path, device="cpu") |
| 1273 | elif os.path.exists(bin_path): |
| 1274 | state_dict = torch.load(bin_path, map_location="cpu", weights_only=True) |
| 1275 | else: |
| 1276 | raise FileNotFoundError( |
| 1277 | f"No model weights found in {path} " |
| 1278 | "(expected model.safetensors or pytorch_model.bin)" |
| 1279 | ) |
| 1280 | model.load_state_dict(state_dict, strict=False) |
| 1281 | model = model.to(device=device, dtype=torch.bfloat16) |
| 1282 | model.eval() |
| 1283 | model.requires_grad_(False) |
| 1284 | return model |
| 1285 | |
| 1286 | def _parse_maybe_list(self, value, length: int) -> list[int]: |
| 1287 | if isinstance(value, str) and "-" in value: |
| 1288 | return [int(s) for s in value.split("-")] |
| 1289 | return [int(value)] * length |
| 1290 | |
| 1291 | def apply_input_local_transformer(self, speech_embeddings: torch.Tensor): |
| 1292 | output = self.input_local_transformer( |
| 1293 | inputs_embeds=speech_embeddings, |
| 1294 | return_dict=True, |
| 1295 | is_causal=not self.config.input_full_attention, |
| 1296 | ) |
| 1297 | return output.last_hidden_state |
| 1298 | |
| 1299 | def apply_speech_embeddings(self, audio_codes: torch.Tensor) -> torch.Tensor: |
| 1300 | num_segments = audio_codes.shape[0] |
| 1301 | _audio_embeddings = torch.zeros( |
| 1302 | (num_segments, self.config.group_size, self.config.input_local_dim), |
| 1303 | dtype=next(self.speech_embeddings[0].parameters()).dtype, |
| 1304 | device=audio_codes.device, |
| 1305 | ) |
| 1306 | for i in range(self.config.audio_channels): |
| 1307 | _audio_embeddings.add_(self.speech_embeddings[i](audio_codes[:, :, i])) |
| 1308 | return _audio_embeddings |
| 1309 | |
| 1310 | def process_audio(self, audio: torch.Tensor) -> torch.Tensor: |
| 1311 | """Pad audio codes to group_size boundary. |
| 1312 | |
| 1313 | Args: |
| 1314 | audio: [T, audio_channels] code tensor |
| 1315 | |
| 1316 | Returns: |
| 1317 | [T//group_size, group_size, audio_channels] |
| 1318 | |
| 1319 | """ |
| 1320 | T = audio.shape[0] |
| 1321 | audio = audio[:, : self.audio_channels] |
| 1322 | padded_T = ( |
| 1323 | (T + self.audio_group_size - 1) |
| 1324 | // self.audio_group_size |
| 1325 | * self.audio_group_size |
| 1326 | ) |
| 1327 | padded_audio = torch.cat( |
| 1328 | [ |
| 1329 | audio, |
| 1330 | torch.zeros( |
| 1331 | padded_T - T, |
| 1332 | self.audio_channels, |
| 1333 | dtype=torch.int32, |
| 1334 | device=audio.device, |
| 1335 | ) |
| 1336 | + audio[-1, :], |
| 1337 | ], |
| 1338 | dim=0, |
| 1339 | ) |
| 1340 | padded_audio = padded_audio.reshape( |
| 1341 | padded_T // self.audio_group_size, |
| 1342 | self.audio_group_size, |
| 1343 | self.audio_channels, |
| 1344 | ) |
| 1345 | return padded_audio |
| 1346 | |
| 1347 | def get_audio_feature( |
| 1348 | self, mel_specs: list[torch.Tensor] |
| 1349 | ) -> tuple[torch.Tensor, list[int]]: |
| 1350 | """Encode mel spectrograms into LLM embedding space. |
| 1351 | |
| 1352 | Args: |
| 1353 | mel_specs: list of mel spectrogram tensors, each [T, n_mels] |
| 1354 | |
| 1355 | Returns: |
| 1356 | Tuple of: |
| 1357 | - audio_embeds: [total_tokens, out_hidden_size] concatenated embeddings |
| 1358 | - item_token_lens: list of int, number of tokens per input item |
| 1359 | |
| 1360 | """ |
| 1361 | if self.audio_tokenizer is None: |
| 1362 | raise RuntimeError( |
| 1363 | "audio_tokenizer is not loaded. " |
| 1364 | "Ensure model_path points to a directory containing audio_tokenizer/." |
| 1365 | ) |
| 1366 | |
| 1367 | if not mel_specs: |
| 1368 | device = next(self.projection.parameters()).device |
| 1369 | dtype = next(self.projection.parameters()).dtype |
| 1370 | return ( |
| 1371 | torch.empty(0, self.config.out_hidden_size, device=device, dtype=dtype), |
| 1372 | [], |
| 1373 | ) |
| 1374 | |
| 1375 | device = next(self.audio_tokenizer.encoder.parameters()).device |
| 1376 | code_list = tokenize_audio_batch( |
| 1377 | mel_specs, |
| 1378 | self.audio_tokenizer.encoder, |
| 1379 | segment_size=self.audio_segment_size, |
| 1380 | device=device, |
| 1381 | ) |
| 1382 | |
| 1383 | item_token_lens: list[int] = [] |
| 1384 | codecs_to_concat = [] |
| 1385 | for codecs in code_list: |
| 1386 | padded_codes = self.process_audio(codecs) |
| 1387 | codecs_to_concat.append(padded_codes) |
| 1388 | item_token_lens.append(padded_codes.shape[0]) |
| 1389 | |
| 1390 | audio_codes = torch.cat( |
| 1391 | codecs_to_concat, dim=0 |
| 1392 | ) # [total_T//group_size, group_size, audio_channels] |
| 1393 | |
| 1394 | _audio_embeddings = self.apply_speech_embeddings(audio_codes) |
| 1395 | audio_embeds = self.apply_input_local_transformer(_audio_embeddings) |
| 1396 | B = audio_embeds.shape[0] |
| 1397 | audio_embeds = self.projection(audio_embeds.reshape(B, -1)) |
| 1398 | return audio_embeds, item_token_lens |