| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import math |
| 5 | from collections.abc import Iterable, Mapping, Sequence |
| 6 | from typing import Annotated, cast |
| 7 | |
| 8 | import torch |
| 9 | import torch.nn.functional as F |
| 10 | from torch import nn |
| 11 | from transformers import ( |
| 12 | BatchFeature, |
| 13 | Qwen3Config, |
| 14 | ) |
| 15 | |
| 16 | from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig |
| 17 | from vllm.config.multimodal import MultiModalDummyOptions |
| 18 | from vllm.config.speech_to_text import SpeechToTextParams |
| 19 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 20 | from vllm.inputs import MultiModalDataDict, PromptType |
| 21 | from vllm.logger import init_logger |
| 22 | from vllm.model_executor.layers.activation import _ACTIVATION_REGISTRY |
| 23 | from vllm.model_executor.layers.attention.mm_encoder_attention import ( |
| 24 | MMEncoderAttention, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.linear import ( |
| 27 | ColumnParallelLinear, |
| 28 | QKVParallelLinear, |
| 29 | ReplicatedLinear, |
| 30 | RowParallelLinear, |
| 31 | ) |
| 32 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 33 | from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 34 | from vllm.model_executor.models.whisper_utils import ( |
| 35 | ISO639_1_SUPPORTED_LANGS, |
| 36 | ) |
| 37 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 38 | from vllm.multimodal.inputs import ( |
| 39 | MultiModalFieldConfig, |
| 40 | MultiModalKwargsItems, |
| 41 | ) |
| 42 | from vllm.multimodal.parse import MultiModalDataItems, MultiModalDataParser |
| 43 | from vllm.multimodal.processing import ( |
| 44 | BaseDummyInputsBuilder, |
| 45 | BaseMultiModalProcessor, |
| 46 | BaseProcessingInfo, |
| 47 | ProcessorInputs, |
| 48 | PromptReplacement, |
| 49 | PromptUpdate, |
| 50 | TimingContext, |
| 51 | ) |
| 52 | from vllm.multimodal.processing.processor import ( |
| 53 | HFMultiModalInputs, |
| 54 | MultiModalProcessingResult, |
| 55 | ) |
| 56 | from vllm.transformers_utils.processor import cached_processor_from_config |
| 57 | from vllm.transformers_utils.processors.funasr import FunASRFeatureExtractor |
| 58 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 59 | |
| 60 | from .interfaces import ( |
| 61 | MultiModalEmbeddings, |
| 62 | SupportsMultiModal, |
| 63 | SupportsTranscription, |
| 64 | _require_is_multimodal, |
| 65 | ) |
| 66 | from .qwen3 import Qwen3Model |
| 67 | from .utils import ( |
| 68 | AutoWeightsLoader, |
| 69 | WeightsMapper, |
| 70 | _merge_multimodal_embeddings, |
| 71 | maybe_prefix, |
| 72 | ) |
| 73 | |
| 74 | logger = init_logger(__name__) |
| 75 | |
| 76 | |
| 77 | def sequence_mask(lengths, maxlen=None, dtype=torch.float32, device=None): |
| 78 | if maxlen is None: |
| 79 | maxlen = lengths.max() |
| 80 | row_vector = torch.arange(0, maxlen, 1).to(lengths.device) |
| 81 | matrix = torch.unsqueeze(lengths, dim=-1) |
| 82 | mask = row_vector < matrix |
| 83 | mask = mask.detach() |
| 84 | |
| 85 | return mask.type(dtype).to(device) if device is not None else mask.type(dtype) |
| 86 | |
| 87 | |
| 88 | class LayerNorm(torch.nn.LayerNorm): |
| 89 | def __init__(self, nout, dim=-1): |
| 90 | super().__init__(nout, eps=1e-12) |
| 91 | self.dim = dim |
| 92 | |
| 93 | def forward(self, x: torch.Tensor): |
| 94 | if self.dim == -1: |
| 95 | return super().forward(x) |
| 96 | return super().forward(x.transpose(self.dim, -1)).transpose(self.dim, -1) |
| 97 | |
| 98 | |
| 99 | class EncoderLayerSANM(nn.Module): |
| 100 | def __init__( |
| 101 | self, |
| 102 | in_size: int, |
| 103 | size: int, |
| 104 | self_attn: nn.Module, |
| 105 | feed_forward: nn.Module, |
| 106 | normalize_before=True, |
| 107 | ): |
| 108 | super().__init__() |
| 109 | self.self_attn = self_attn |
| 110 | self.feed_forward = feed_forward |
| 111 | self.norm1 = LayerNorm(in_size) |
| 112 | self.norm2 = LayerNorm(size) |
| 113 | self.in_size = in_size |
| 114 | self.size = size |
| 115 | self.normalize_before = normalize_before |
| 116 | |
| 117 | def forward( |
| 118 | self, |
| 119 | hidden_states: torch.Tensor, |
| 120 | mask: torch.Tensor | None = None, |
| 121 | cache=None, |
| 122 | mask_shift_chunk=None, |
| 123 | mask_att_chunk_encoder=None, |
| 124 | ): |
| 125 | residual = hidden_states |
| 126 | hidden_states = self.norm1(hidden_states) |
| 127 | |
| 128 | if self.in_size == self.size: |
| 129 | hidden_states = residual + self.self_attn( |
| 130 | hidden_states, |
| 131 | mask, |
| 132 | mask_shift_chunk=mask_shift_chunk, |
| 133 | mask_att_chunk_encoder=mask_att_chunk_encoder, |
| 134 | ) |
| 135 | else: |
| 136 | hidden_states = self.self_attn( |
| 137 | hidden_states, |
| 138 | mask, |
| 139 | mask_shift_chunk=mask_shift_chunk, |
| 140 | mask_att_chunk_encoder=mask_att_chunk_encoder, |
| 141 | ) |
| 142 | |
| 143 | residual = hidden_states |
| 144 | hidden_states = self.norm2(hidden_states) |
| 145 | hidden_states = residual + self.feed_forward(hidden_states) |
| 146 | |
| 147 | return hidden_states, mask, cache, mask_shift_chunk, mask_att_chunk_encoder |
| 148 | |
| 149 | |
| 150 | class MultiHeadedAttentionSANM(nn.Module): |
| 151 | def __init__( |
| 152 | self, |
| 153 | n_head: int, |
| 154 | in_feat: int, |
| 155 | n_feat: int, |
| 156 | kernel_size: int, |
| 157 | sanm_shift: int = 0, |
| 158 | ): |
| 159 | super().__init__() |
| 160 | assert n_feat % n_head == 0 |
| 161 | # We assume d_v always equals d_k |
| 162 | self.d_k = n_feat // n_head |
| 163 | self.h = n_head |
| 164 | self.out_proj = ReplicatedLinear( |
| 165 | input_size=n_feat, |
| 166 | output_size=n_feat, |
| 167 | bias=True, |
| 168 | ) |
| 169 | self.linear_q_k_v = ReplicatedLinear( |
| 170 | input_size=in_feat, |
| 171 | output_size=n_feat * 3, |
| 172 | bias=True, |
| 173 | ) |
| 174 | self.attn = None |
| 175 | |
| 176 | self.fsmn_block = nn.Conv1d( |
| 177 | n_feat, n_feat, kernel_size, stride=1, padding=0, groups=n_feat, bias=False |
| 178 | ) |
| 179 | # padding |
| 180 | left_padding = (kernel_size - 1) // 2 |
| 181 | if sanm_shift > 0: |
| 182 | left_padding = left_padding + sanm_shift |
| 183 | right_padding = kernel_size - 1 - left_padding |
| 184 | self.pad_fn = nn.ConstantPad1d((left_padding, right_padding), 0.0) |
| 185 | |
| 186 | def forward_fsmn( |
| 187 | self, |
| 188 | inputs: torch.Tensor, |
| 189 | mask: torch.Tensor, |
| 190 | mask_shift_chunk: torch.Tensor = None, |
| 191 | ): |
| 192 | b, t, d = inputs.size() |
| 193 | if mask is not None: |
| 194 | mask = torch.reshape(mask, (b, -1, 1)) |
| 195 | if mask_shift_chunk is not None: |
| 196 | mask = mask * mask_shift_chunk |
| 197 | inputs = inputs * mask |
| 198 | |
| 199 | x = inputs.transpose(1, 2) |
| 200 | x = self.pad_fn(x) |
| 201 | x = self.fsmn_block(x) |
| 202 | x = x.transpose(1, 2) |
| 203 | x += inputs |
| 204 | if mask is not None: |
| 205 | x = x * mask |
| 206 | return x |
| 207 | |
| 208 | def forward_qkv(self, x: torch.Tensor): |
| 209 | b, t, d = x.size() |
| 210 | |
| 211 | q_k_v, _ = self.linear_q_k_v(x) |
| 212 | q, k, v = torch.split(q_k_v, int(self.h * self.d_k), dim=-1) |
| 213 | q_h = torch.reshape(q, (b, t, self.h, self.d_k)).transpose(1, 2) |
| 214 | k_h = torch.reshape(k, (b, t, self.h, self.d_k)).transpose(1, 2) |
| 215 | v_h = torch.reshape(v, (b, t, self.h, self.d_k)).transpose(1, 2) |
| 216 | |
| 217 | return q_h, k_h, v_h, v |
| 218 | |
| 219 | def forward_attention( |
| 220 | self, |
| 221 | value: torch.Tensor, |
| 222 | scores: torch.Tensor, |
| 223 | mask: torch.Tensor, |
| 224 | mask_att_chunk_encoder: torch.Tensor = None, |
| 225 | ): |
| 226 | n_batch = value.size(0) |
| 227 | if mask is not None: |
| 228 | if mask_att_chunk_encoder is not None: |
| 229 | mask = mask * mask_att_chunk_encoder |
| 230 | |
| 231 | mask = mask.unsqueeze(1).eq(0) |
| 232 | |
| 233 | min_value = -float("inf") |
| 234 | scores = scores.masked_fill(mask, min_value) |
| 235 | attn = torch.softmax(scores, dim=-1).masked_fill(mask, 0.0) |
| 236 | else: |
| 237 | attn = torch.softmax(scores, dim=-1) |
| 238 | |
| 239 | p_attn = attn |
| 240 | x = torch.matmul(p_attn, value) |
| 241 | x = x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k) |
| 242 | |
| 243 | out, _ = self.out_proj(x) |
| 244 | return out |
| 245 | |
| 246 | def forward( |
| 247 | self, |
| 248 | hidden_states: torch.Tensor, |
| 249 | mask: torch.Tensor, |
| 250 | mask_shift_chunk: torch.Tensor = None, |
| 251 | mask_att_chunk_encoder: torch.Tensor = None, |
| 252 | ): |
| 253 | q_h, k_h, v_h, v = self.forward_qkv(hidden_states) |
| 254 | fsmn_memory = self.forward_fsmn(v, mask, mask_shift_chunk) |
| 255 | q_h = q_h * self.d_k ** (-0.5) |
| 256 | scores = torch.matmul(q_h, k_h.transpose(-2, -1)) |
| 257 | att_outs = self.forward_attention(v_h, scores, mask, mask_att_chunk_encoder) |
| 258 | return att_outs + fsmn_memory |
| 259 | |
| 260 | |
| 261 | class SinusoidalPositionEncoder(torch.nn.Module): |
| 262 | def __init__(self, d_model=80): |
| 263 | super().__init__() |
| 264 | |
| 265 | def encode( |
| 266 | self, |
| 267 | positions: torch.Tensor, |
| 268 | depth: int, |
| 269 | dtype: torch.dtype = torch.float32, |
| 270 | ): |
| 271 | batch_size = positions.size(0) |
| 272 | positions = positions.type(dtype) |
| 273 | device = positions.device |
| 274 | log_timescale_increment = torch.log( |
| 275 | torch.tensor([10000], dtype=dtype, device=device) |
| 276 | ) / (depth / 2 - 1) |
| 277 | inv_timescales = torch.exp( |
| 278 | torch.arange(depth / 2, device=device).type(dtype) |
| 279 | * (-log_timescale_increment) |
| 280 | ) |
| 281 | inv_timescales = torch.reshape(inv_timescales, [batch_size, -1]) |
| 282 | scaled_time = torch.reshape(positions, [1, -1, 1]) * torch.reshape( |
| 283 | inv_timescales, [1, 1, -1] |
| 284 | ) |
| 285 | encoding = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=2) |
| 286 | return encoding.type(dtype) |
| 287 | |
| 288 | def forward(self, hidden_states: torch.Tensor): |
| 289 | batch_size, timesteps, input_dim = hidden_states.size() |
| 290 | positions = torch.arange(1, timesteps + 1, device=hidden_states.device)[None, :] |
| 291 | position_encoding = self.encode(positions, input_dim, hidden_states.dtype).to( |
| 292 | hidden_states.device |
| 293 | ) |
| 294 | |
| 295 | return hidden_states + position_encoding |
| 296 | |
| 297 | |
| 298 | class SenseVoiceEncoderSmall(nn.Module): |
| 299 | def __init__( |
| 300 | self, |
| 301 | input_size: int, |
| 302 | output_size: int = 256, |
| 303 | attention_heads: int = 4, |
| 304 | linear_units: int = 2048, |
| 305 | num_blocks: int = 6, |
| 306 | tp_blocks: int = 0, |
| 307 | attention_dropout_rate: float = 0.0, |
| 308 | normalize_before: bool = True, |
| 309 | kernel_size: int = 11, |
| 310 | sanm_shift: int = 0, |
| 311 | **kwargs, |
| 312 | ): |
| 313 | super().__init__() |
| 314 | self._output_size = output_size |
| 315 | self.embed = SinusoidalPositionEncoder() |
| 316 | |
| 317 | self.normalize_before = normalize_before |
| 318 | |
| 319 | positionwise_layer = PositionwiseFeedForward |
| 320 | positionwise_layer_args = ( |
| 321 | output_size, |
| 322 | linear_units, |
| 323 | ) |
| 324 | |
| 325 | encoder_selfattn_layer = MultiHeadedAttentionSANM |
| 326 | encoder_selfattn_layer_args0 = ( |
| 327 | attention_heads, |
| 328 | input_size, |
| 329 | output_size, |
| 330 | kernel_size, |
| 331 | sanm_shift, |
| 332 | ) |
| 333 | encoder_selfattn_layer_args = ( |
| 334 | attention_heads, |
| 335 | output_size, |
| 336 | output_size, |
| 337 | kernel_size, |
| 338 | sanm_shift, |
| 339 | ) |
| 340 | |
| 341 | self.encoders0 = nn.ModuleList( |
| 342 | [ |
| 343 | EncoderLayerSANM( |
| 344 | input_size, |
| 345 | output_size, |
| 346 | encoder_selfattn_layer(*encoder_selfattn_layer_args0), |
| 347 | positionwise_layer(*positionwise_layer_args), |
| 348 | ) |
| 349 | for i in range(1) |
| 350 | ] |
| 351 | ) |
| 352 | self.encoders = nn.ModuleList( |
| 353 | [ |
| 354 | EncoderLayerSANM( |
| 355 | output_size, |
| 356 | output_size, |
| 357 | encoder_selfattn_layer(*encoder_selfattn_layer_args), |
| 358 | positionwise_layer(*positionwise_layer_args), |
| 359 | ) |
| 360 | for i in range(num_blocks - 1) |
| 361 | ] |
| 362 | ) |
| 363 | |
| 364 | self.tp_encoders = nn.ModuleList( |
| 365 | [ |
| 366 | EncoderLayerSANM( |
| 367 | output_size, |
| 368 | output_size, |
| 369 | encoder_selfattn_layer(*encoder_selfattn_layer_args), |
| 370 | positionwise_layer(*positionwise_layer_args), |
| 371 | ) |
| 372 | for i in range(tp_blocks) |
| 373 | ] |
| 374 | ) |
| 375 | |
| 376 | self.after_norm = LayerNorm(output_size) |
| 377 | |
| 378 | self.tp_norm = LayerNorm(output_size) |
| 379 | |
| 380 | def output_size(self) -> int: |
| 381 | return self._output_size |
| 382 | |
| 383 | def forward( |
| 384 | self, |
| 385 | xs_pad: torch.Tensor, |
| 386 | ilens: torch.Tensor, |
| 387 | ): |
| 388 | maxlen = xs_pad.shape[1] |
| 389 | masks = sequence_mask( |
| 390 | ilens, maxlen=maxlen, dtype=ilens.dtype, device=ilens.device |
| 391 | )[:, None, :] |
| 392 | |
| 393 | xs_pad *= self.output_size() ** 0.5 |
| 394 | |
| 395 | xs_pad = self.embed(xs_pad) |
| 396 | |
| 397 | for layer_idx, encoder_layer in enumerate(self.encoders0): |
| 398 | encoder_outs = encoder_layer(xs_pad, masks) |
| 399 | xs_pad, masks = encoder_outs[0], encoder_outs[1] |
| 400 | |
| 401 | for layer_idx, encoder_layer in enumerate(self.encoders): |
| 402 | encoder_outs = encoder_layer(xs_pad, masks) |
| 403 | xs_pad, masks = encoder_outs[0], encoder_outs[1] |
| 404 | |
| 405 | xs_pad = self.after_norm(xs_pad) |
| 406 | |
| 407 | olens = masks.squeeze(1).sum(1).int() |
| 408 | |
| 409 | for layer_idx, encoder_layer in enumerate(self.tp_encoders): |
| 410 | encoder_outs = encoder_layer(xs_pad, masks) |
| 411 | xs_pad, masks = encoder_outs[0], encoder_outs[1] |
| 412 | |
| 413 | xs_pad = self.tp_norm(xs_pad) |
| 414 | return xs_pad, olens |
| 415 | |
| 416 | |
| 417 | class PositionwiseFeedForward(nn.Module): |
| 418 | def __init__(self, idim: int, hidden_units: int): |
| 419 | super().__init__() |
| 420 | self.w_1 = ColumnParallelLinear( |
| 421 | input_size=idim, |
| 422 | output_size=hidden_units, |
| 423 | bias=True, |
| 424 | ) |
| 425 | self.w_2 = RowParallelLinear( |
| 426 | input_size=hidden_units, |
| 427 | output_size=idim, |
| 428 | bias=True, |
| 429 | ) |
| 430 | self.activation = _ACTIVATION_REGISTRY["relu"] |
| 431 | |
| 432 | def forward(self, hidden_states: torch.Tensor): |
| 433 | hidden_states, _ = self.w_1(hidden_states) |
| 434 | hidden_states = self.activation(hidden_states) |
| 435 | hidden_states, _ = self.w_2(hidden_states) |
| 436 | return hidden_states |
| 437 | |
| 438 | |
| 439 | class EncoderLayer(nn.Module): |
| 440 | def __init__( |
| 441 | self, |
| 442 | size: int, |
| 443 | self_attn: nn.Module, |
| 444 | feed_forward: nn.Module, |
| 445 | ): |
| 446 | super().__init__() |
| 447 | self.self_attn = self_attn |
| 448 | self.feed_forward = feed_forward |
| 449 | self.norm1 = LayerNorm(size) |
| 450 | self.norm2 = LayerNorm(size) |
| 451 | |
| 452 | def forward(self, hidden_states: torch.Tensor): |
| 453 | residual = hidden_states |
| 454 | hidden_states = self.norm1(hidden_states) |
| 455 | hidden_states = residual + self.self_attn(hidden_states, None, None) |
| 456 | residual = hidden_states |
| 457 | hidden_states = self.norm2(hidden_states) |
| 458 | hidden_states = residual + self.feed_forward(hidden_states) |
| 459 | |
| 460 | return hidden_states |
| 461 | |
| 462 | |
| 463 | class FunASRAudioAttention(nn.Module): |
| 464 | def __init__( |
| 465 | self, |
| 466 | num_heads: int, |
| 467 | embed_dim: int, |
| 468 | prefix: str = "", |
| 469 | ): |
| 470 | super().__init__() |
| 471 | self.embed_dim = embed_dim |
| 472 | self.num_heads = num_heads |
| 473 | self.head_dim = self.embed_dim // self.num_heads |
| 474 | tp_size = get_tensor_model_parallel_world_size() |
| 475 | self.num_local_heads = self.num_heads // tp_size |
| 476 | |
| 477 | if (self.head_dim * self.num_heads) != self.embed_dim: |
| 478 | raise ValueError( |
| 479 | f"embed_dim must be divisible by num_heads (got `embed_dim`: " |
| 480 | f"{self.embed_dim} and `num_heads`: {self.num_heads})." |
| 481 | ) |
| 482 | |
| 483 | self.scaling = self.head_dim**-0.5 |
| 484 | |
| 485 | self.qkv = QKVParallelLinear( |
| 486 | hidden_size=self.embed_dim, |
| 487 | head_size=self.head_dim, |
| 488 | total_num_heads=self.num_heads, |
| 489 | total_num_kv_heads=self.num_heads, |
| 490 | bias=True, |
| 491 | prefix=f"{prefix}.qkv", |
| 492 | ) |
| 493 | |
| 494 | self.out_proj = RowParallelLinear( |
| 495 | input_size=self.embed_dim, |
| 496 | output_size=self.embed_dim, |
| 497 | bias=True, |
| 498 | prefix=f"{prefix}.out_proj", |
| 499 | ) |
| 500 | |
| 501 | self.attn = MMEncoderAttention( |
| 502 | num_heads=self.num_local_heads, |
| 503 | head_size=self.head_dim, |
| 504 | scale=self.scaling, |
| 505 | ) |
| 506 | |
| 507 | def forward( |
| 508 | self, |
| 509 | hidden_states: torch.Tensor, |
| 510 | cu_seqlens: torch.Tensor, |
| 511 | max_seqlen: torch.Tensor | None, |
| 512 | ) -> torch.Tensor: |
| 513 | bs, seq_length, _ = hidden_states.size() |
| 514 | qkv, _ = self.qkv(hidden_states) |
| 515 | q, k, v = qkv.chunk(3, dim=-1) |
| 516 | q = q.view(bs, seq_length, -1, self.head_dim) |
| 517 | k = k.view(bs, seq_length, -1, self.head_dim) |
| 518 | v = v.view(bs, seq_length, -1, self.head_dim) |
| 519 | |
| 520 | attn_output = self.attn( |
| 521 | query=q, |
| 522 | key=k, |
| 523 | value=v, |
| 524 | cu_seqlens=cu_seqlens, |
| 525 | max_seqlen=max_seqlen, |
| 526 | ) |
| 527 | |
| 528 | attn_output = attn_output.view(bs, seq_length, -1) |
| 529 | output, _ = self.out_proj(attn_output) |
| 530 | return output |
| 531 | |
| 532 | |
| 533 | class Transformer(nn.Module): |
| 534 | def __init__( |
| 535 | self, |
| 536 | downsample_rate=2, |
| 537 | encoder_dim=1280, |
| 538 | llm_dim=4096, |
| 539 | ffn_dim: int = 2048, |
| 540 | prefix: str = "", |
| 541 | **kwargs, |
| 542 | ): |
| 543 | super().__init__() |
| 544 | self.k = downsample_rate |
| 545 | self.encoder_dim = encoder_dim |
| 546 | self.llm_dim = llm_dim |
| 547 | self.linear1 = ColumnParallelLinear( |
| 548 | input_size=self.encoder_dim * self.k, |
| 549 | output_size=ffn_dim, |
| 550 | bias=True, |
| 551 | ) |
| 552 | self.relu = nn.ReLU() |
| 553 | self.linear2 = RowParallelLinear( |
| 554 | input_size=ffn_dim, |
| 555 | output_size=self.llm_dim, |
| 556 | bias=True, |
| 557 | ) |
| 558 | |
| 559 | self.blocks = None |
| 560 | if kwargs.get("n_layer", 2) > 0: |
| 561 | self.blocks = nn.ModuleList( |
| 562 | [ |
| 563 | EncoderLayer( |
| 564 | llm_dim, |
| 565 | FunASRAudioAttention( |
| 566 | kwargs.get("attention_heads", 8), |
| 567 | llm_dim, |
| 568 | prefix=f"{prefix}.self_attn", |
| 569 | ), |
| 570 | PositionwiseFeedForward( |
| 571 | llm_dim, |
| 572 | llm_dim // 4, |
| 573 | ), |
| 574 | ) |
| 575 | for _ in range(kwargs.get("n_layer", 2)) |
| 576 | ] |
| 577 | ) |
| 578 | |
| 579 | def forward(self, hidden_states: torch.Tensor, ilens: torch.Tensor): |
| 580 | max_len = ilens.max() |
| 581 | hidden_states = hidden_states[:, :max_len, :] |
| 582 | batch_size, seq_len, dim = hidden_states.size() |
| 583 | chunk_num = (seq_len - 1) // self.k + 1 |
| 584 | pad_num = chunk_num * self.k - seq_len |
| 585 | hidden_states = F.pad(hidden_states, (0, 0, 0, pad_num, 0, 0), value=0.0) |
| 586 | seq_len = hidden_states.size(1) |
| 587 | |
| 588 | hidden_states = hidden_states.contiguous() |
| 589 | hidden_states = hidden_states.view(batch_size, chunk_num, dim * self.k) |
| 590 | hidden_states, _ = self.linear1(hidden_states) |
| 591 | hidden_states = self.relu(hidden_states) |
| 592 | hidden_states, _ = self.linear2(hidden_states) |
| 593 | |
| 594 | olens = None |
| 595 | olens = (ilens - 1) // self.k + 1 |
| 596 | |
| 597 | if self.blocks is not None: |
| 598 | for layer, block in enumerate(self.blocks): |
| 599 | hidden_states = block(hidden_states) |
| 600 | return hidden_states, olens |
| 601 | |
| 602 | |
| 603 | class FunASRAudioInputs(TensorSchema): |
| 604 | """Dimensions: |
| 605 | - b: Batch size |
| 606 | - nmb: Number of mel bins |
| 607 | - t: Time frames (M) |
| 608 | """ |
| 609 | |
| 610 | input_features: Annotated[ |
| 611 | list[torch.Tensor] | None, |
| 612 | TensorShape("b", "nmb", "t"), |
| 613 | ] |
| 614 | speech_lengths: Annotated[ |
| 615 | list[torch.Tensor] | None, |
| 616 | TensorShape("b"), |
| 617 | ] |
| 618 | fake_token_lengths: Annotated[ |
| 619 | list[torch.Tensor] | None, |
| 620 | TensorShape("b"), |
| 621 | ] |
| 622 | |
| 623 | |
| 624 | class FunASREncoder(nn.Module): |
| 625 | hf_to_vllm_mapper = WeightsMapper( |
| 626 | orig_to_new_stacked={ |
| 627 | ".self_attn.q_proj": (".self_attn.qkv", "q"), |
| 628 | ".self_attn.k_proj": (".self_attn.qkv", "k"), |
| 629 | ".self_attn.v_proj": (".self_attn.qkv", "v"), |
| 630 | } |
| 631 | ) |
| 632 | |
| 633 | def __init__( |
| 634 | self, *, vllm_config: VllmConfig, prefix: str = "", init_in_fp32: bool = False |
| 635 | ): |
| 636 | super().__init__() |
| 637 | self.audio_encoder = SenseVoiceEncoderSmall( |
| 638 | input_size=560, **vllm_config.model_config.hf_config.audio_encoder_conf |
| 639 | ) |
| 640 | self.audio_adaptor = Transformer( |
| 641 | downsample_rate=1, |
| 642 | use_low_frame_rate=True, |
| 643 | ffn_dim=2048, |
| 644 | llm_dim=1024, |
| 645 | encoder_dim=512, |
| 646 | n_layer=2, |
| 647 | freeze=True, |
| 648 | prefix=maybe_prefix(prefix, "audio_encoder"), |
| 649 | ) |
| 650 | |
| 651 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 652 | loader = AutoWeightsLoader(self) |
| 653 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 654 | |
| 655 | |
| 656 | class FunASRModel(nn.Module): |
| 657 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 658 | super().__init__() |
| 659 | self.encoder = FunASREncoder( |
| 660 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "encoder") |
| 661 | ) |
| 662 | self.decoder = Qwen3Model( |
| 663 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "decoder") |
| 664 | ) |
| 665 | |
| 666 | def forward( |
| 667 | self, |
| 668 | input_ids: torch.Tensor | None, |
| 669 | positions: torch.Tensor, |
| 670 | inputs_embeds: torch.Tensor | None = None, |
| 671 | ) -> torch.Tensor: |
| 672 | decoder_outputs = self.decoder( |
| 673 | input_ids=input_ids, |
| 674 | positions=positions, |
| 675 | inputs_embeds=inputs_embeds, |
| 676 | ) |
| 677 | return decoder_outputs |
| 678 | |
| 679 | def get_encoder_outputs( |
| 680 | self, |
| 681 | speech: torch.Tensor | list[torch.Tensor] | None, |
| 682 | speech_lengths: torch.Tensor | list[torch.Tensor] | None, |
| 683 | ) -> torch.Tensor | None: |
| 684 | self.feat_permute = False |
| 685 | |
| 686 | if self.feat_permute: |
| 687 | if not isinstance(speech, torch.Tensor): |
| 688 | raise NotImplementedError("List audio inputs cannot be permuted") |
| 689 | encoder_out, encoder_out_lens = self.encoder.audio_encoder( |
| 690 | speech.permute(0, 2, 1), speech_lengths |
| 691 | ) |
| 692 | else: |
| 693 | encoder_out, encoder_out_lens = self.encoder.audio_encoder( |
| 694 | speech, speech_lengths |
| 695 | ) |
| 696 | |
| 697 | encoder_out, encoder_out_lens = self.encoder.audio_adaptor( |
| 698 | encoder_out, encoder_out_lens |
| 699 | ) |
| 700 | return encoder_out |
| 701 | |
| 702 | |
| 703 | class FunASRProcessingInfo(BaseProcessingInfo): |
| 704 | def get_hf_config(self) -> Qwen3Config: |
| 705 | return self.ctx.get_hf_config(Qwen3Config) |
| 706 | |
| 707 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 708 | return {"audio": 1} |
| 709 | |
| 710 | def get_feature_extractor(self, **kwargs: object) -> FunASRFeatureExtractor: |
| 711 | hf_processor = self.get_hf_processor(**kwargs) |
| 712 | feature_extractor = hf_processor.feature_extractor # type: ignore |
| 713 | assert isinstance(feature_extractor, FunASRFeatureExtractor) |
| 714 | return feature_extractor |
| 715 | |
| 716 | def get_data_parser(self) -> MultiModalDataParser: |
| 717 | feature_extractor = self.get_feature_extractor() |
| 718 | return MultiModalDataParser( |
| 719 | target_sr=feature_extractor.sampling_rate, |
| 720 | target_channels=self.get_target_channels(), |
| 721 | ) |
| 722 | |
| 723 | def get_target_channels(self) -> int: |
| 724 | return 1 |
| 725 | |
| 726 | |
| 727 | class FunASRDummyInputsBuilder(BaseDummyInputsBuilder[FunASRProcessingInfo]): |
| 728 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 729 | num_audios = mm_counts.get("audio", 0) |
| 730 | |
| 731 | return "<|AUDIO|>" * num_audios |
| 732 | |
| 733 | def get_dummy_mm_data( |
| 734 | self, |
| 735 | seq_len: int, |
| 736 | mm_counts: Mapping[str, int], |
| 737 | mm_options: MultiModalDummyOptions, |
| 738 | ) -> MultiModalDataDict: |
| 739 | feature_extractor = self.info.get_feature_extractor() |
| 740 | |
| 741 | sampling_rate = feature_extractor.sampling_rate |
| 742 | audio_len = feature_extractor.chunk_length * sampling_rate |
| 743 | |
| 744 | return { |
| 745 | "audio": self._get_dummy_audios( |
| 746 | length=audio_len, |
| 747 | num_audios=mm_counts.get("audio", 0), |
| 748 | overrides=mm_options.get("audio"), |
| 749 | ), |
| 750 | } |
| 751 | |
| 752 | |
| 753 | class FunASRMultiModalProcessor(BaseMultiModalProcessor[FunASRProcessingInfo]): |
| 754 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 755 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 756 | |
| 757 | def _get_hf_mm_inputs( |
| 758 | self, |
| 759 | mm_items: MultiModalDataItems, |
| 760 | hf_kwargs: Mapping[str, object], |
| 761 | ) -> HFMultiModalInputs: |
| 762 | hf_inputs = super()._get_hf_mm_inputs(mm_items, hf_kwargs) |
| 763 | |
| 764 | feature_extractor = self.info.get_feature_extractor(**hf_kwargs) |
| 765 | return hf_inputs._replace( |
| 766 | hf_kwargs=dict( |
| 767 | hf_inputs.hf_kwargs, |
| 768 | sampling_rate=feature_extractor.sampling_rate, |
| 769 | ) |
| 770 | ) |
| 771 | |
| 772 | def _cached_apply_hf_processor( |
| 773 | self, |
| 774 | inputs: ProcessorInputs, |
| 775 | timing_ctx: TimingContext, |
| 776 | ) -> MultiModalProcessingResult: |
| 777 | # Dithering injects noise into the extracted features, so the |
| 778 | # feature extractor is not a pure function of its input. Since the |
| 779 | # processing cache assumes that processor outputs are invariant |
| 780 | # across calls, bypass the cache when dithering is active. |
| 781 | if self.info.get_feature_extractor().dither > 0: |
| 782 | return self._apply_hf_processor(inputs, timing_ctx) |
| 783 | |
| 784 | return super()._cached_apply_hf_processor(inputs, timing_ctx) |
| 785 | |
| 786 | def _get_mm_fields_config( |
| 787 | self, |
| 788 | hf_inputs: BatchFeature, |
| 789 | hf_processor_mm_kwargs: Mapping[str, object], |
| 790 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 791 | return dict( |
| 792 | input_features=MultiModalFieldConfig.batched("audio"), |
| 793 | speech_lengths=MultiModalFieldConfig.batched("audio"), |
| 794 | fake_token_lengths=MultiModalFieldConfig.batched("audio", keep_on_cpu=True), |
| 795 | ) |
| 796 | |
| 797 | def _get_prompt_updates( |
| 798 | self, |
| 799 | mm_items: MultiModalDataItems, |
| 800 | hf_processor_mm_kwargs: Mapping[str, object], |
| 801 | out_mm_kwargs: MultiModalKwargsItems, |
| 802 | ) -> Sequence[PromptUpdate]: |
| 803 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 804 | audio_token_id = processor.audio_token_id |
| 805 | |
| 806 | out_mm_data = out_mm_kwargs.get_data() |
| 807 | |
| 808 | fake_token_lengths = out_mm_data.get("fake_token_lengths") |
| 809 | if fake_token_lengths is None: |
| 810 | audio_output_lengths = [] |
| 811 | else: |
| 812 | assert isinstance(fake_token_lengths, torch.Tensor) |
| 813 | |
| 814 | audio_output_lengths = fake_token_lengths.tolist() |
| 815 | |
| 816 | def get_replacement_qwen2_audio(item_idx: int): |
| 817 | num_features = audio_output_lengths[item_idx] |
| 818 | return [audio_token_id] * num_features |
| 819 | |
| 820 | return [ |
| 821 | PromptReplacement( |
| 822 | modality="audio", |
| 823 | target=[audio_token_id], |
| 824 | replacement=get_replacement_qwen2_audio, |
| 825 | ) |
| 826 | ] |
| 827 | |
| 828 | |
| 829 | @MULTIMODAL_REGISTRY.register_processor( |
| 830 | FunASRMultiModalProcessor, |
| 831 | info=FunASRProcessingInfo, |
| 832 | dummy_inputs=FunASRDummyInputsBuilder, |
| 833 | ) |
| 834 | class FunASRForConditionalGeneration( |
| 835 | nn.Module, SupportsTranscription, SupportsMultiModal |
| 836 | ): |
| 837 | hf_to_vllm_mapper = WeightsMapper( |
| 838 | orig_to_new_substr={ |
| 839 | "linear_q.": "q_proj.", |
| 840 | "linear_k.": "k_proj.", |
| 841 | "linear_v.": "v_proj.", |
| 842 | "linear_out.": "out_proj.", |
| 843 | "audio_adaptor.": "model.encoder.audio_adaptor.", |
| 844 | "audio_encoder.": "model.encoder.audio_encoder.", |
| 845 | "llm.model.": "model.decoder.", |
| 846 | "llm.lm_head": "lm_head", |
| 847 | } |
| 848 | ) |
| 849 | |
| 850 | supports_transcription_only = True |
| 851 | supports_segment_timestamp = True |
| 852 | supported_languages = ISO639_1_SUPPORTED_LANGS |
| 853 | |
| 854 | @classmethod |
| 855 | def validate_language(cls, language: str | None) -> str | None: |
| 856 | if language is None: |
| 857 | # TODO language should be optional and can be guessed. |
| 858 | # For now we default to en. See |
| 859 | # https://github.com/huggingface/transformers/blob/main/src/transformers/models/whisper/generation_whisper.py#L1520 |
| 860 | logger.warning( |
| 861 | "Defaulting to language='en'. If you wish to transcribe " |
| 862 | "audio in a different language, pass the `language` field " |
| 863 | "in the TranscriptionRequest." |
| 864 | ) |
| 865 | language = "en" |
| 866 | return super().validate_language(language) |
| 867 | |
| 868 | @classmethod |
| 869 | def get_generation_prompt( |
| 870 | cls, |
| 871 | stt_params: SpeechToTextParams, |
| 872 | ) -> PromptType: |
| 873 | audio = stt_params.audio |
| 874 | stt_config = stt_params.stt_config |
| 875 | language = stt_params.language |
| 876 | hotwords = stt_params.hotwords |
| 877 | |
| 878 | if language is None: |
| 879 | raise ValueError( |
| 880 | "Language must be specified when creating the funasr prompt" |
| 881 | ) |
| 882 | |
| 883 | if hotwords is not None: |
| 884 | funasr_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n请结合上下文信息，更加准确地完成语音转写任务。如果没有相关信息，我们会留空。\n\n\n**上下文信息：**\n\n\n热词列表：[{}]\n语音转写：<|AUDIO|><|im_end|>\n<|im_start|>assistant\n".format( # noqa: E501 |
| 885 | hotwords |
| 886 | ) |
| 887 | else: |
| 888 | funasr_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n语音转写：<|AUDIO|><|im_end|>\n<|im_start|>assistant\n" # noqa: E501 |
| 889 | |
| 890 | prompt = { |
| 891 | "prompt": funasr_prompt, |
| 892 | "multi_modal_data": { |
| 893 | "audio": (audio, stt_config.sample_rate), |
| 894 | }, |
| 895 | } |
| 896 | return cast(PromptType, prompt) |
| 897 | |
| 898 | @classmethod |
| 899 | def get_speech_to_text_config( |
| 900 | cls, model_config: ModelConfig, task_type: str |
| 901 | ) -> SpeechToTextConfig: |
| 902 | processor = cached_processor_from_config(model_config) |
| 903 | |
| 904 | return SpeechToTextConfig( |
| 905 | max_audio_clip_s=processor.feature_extractor.chunk_length, |
| 906 | sample_rate=processor.feature_extractor.sampling_rate, |
| 907 | ) |
| 908 | |
| 909 | @classmethod |
| 910 | def get_num_audio_tokens( |
| 911 | cls, |
| 912 | audio_duration_s: float, |
| 913 | stt_config: SpeechToTextConfig, |
| 914 | model_config: ModelConfig, |
| 915 | ) -> int | None: |
| 916 | processor = cached_processor_from_config(model_config) |
| 917 | hop_length = processor.feature_extractor.hop_length |
| 918 | assert hop_length is not None |
| 919 | return math.ceil(audio_duration_s * stt_config.sample_rate / hop_length) |
| 920 | |
| 921 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 922 | super().__init__() |
| 923 | config = vllm_config.model_config.hf_config |
| 924 | quant_config = vllm_config.quant_config |
| 925 | self.config = config |
| 926 | self.dtype = vllm_config.model_config.dtype |
| 927 | |
| 928 | self.model = FunASRModel( |
| 929 | vllm_config=vllm_config, |
| 930 | prefix=maybe_prefix(prefix, "model"), |
| 931 | ) |
| 932 | logit_scale = getattr(config, "logit_scale", 1.0) |
| 933 | |
| 934 | self.lm_head = ParallelLMHead( |
| 935 | config.vocab_size, |
| 936 | config.hidden_size, |
| 937 | quant_config=quant_config, |
| 938 | prefix=maybe_prefix(prefix, "lm_head"), |
| 939 | ) |
| 940 | if config.tie_word_embeddings: |
| 941 | self.lm_head = self.lm_head.tie_weights(self.model.decoder.embed_tokens) |
| 942 | self.logits_processor = LogitsProcessor(config.vocab_size, scale=logit_scale) |
| 943 | |
| 944 | def get_language_model(self) -> torch.nn.Module: |
| 945 | # Required as part of SupportsMultiModal interface. |
| 946 | return self.model.decoder |
| 947 | |
| 948 | def forward( |
| 949 | self, |
| 950 | input_ids: torch.Tensor, |
| 951 | positions: torch.Tensor, |
| 952 | inputs_embeds: torch.Tensor | None = None, |
| 953 | **kwargs, |
| 954 | ) -> torch.Tensor: |
| 955 | decoder_outputs = self.model( |
| 956 | input_ids=input_ids, |
| 957 | positions=positions, |
| 958 | inputs_embeds=inputs_embeds, |
| 959 | ) |
| 960 | return decoder_outputs |
| 961 | |
| 962 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 963 | audio_input = self._parse_and_validate_audio_input(**kwargs) |
| 964 | |
| 965 | speech = audio_input["input_features"] |
| 966 | speech_lengths = audio_input["speech_lengths"] |
| 967 | enc_output = self.model.get_encoder_outputs( |
| 968 | speech=speech, speech_lengths=speech_lengths |
| 969 | ) |
| 970 | |
| 971 | return enc_output |
| 972 | |
| 973 | def embed_input_ids( |
| 974 | self, |
| 975 | input_ids: torch.Tensor, |
| 976 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 977 | *, |
| 978 | is_multimodal: torch.Tensor | None = None, |
| 979 | ) -> torch.Tensor: |
| 980 | inputs_embeds = self.model.decoder.embed_input_ids(input_ids) |
| 981 | |
| 982 | return _merge_multimodal_embeddings( |
| 983 | inputs_embeds=inputs_embeds, |
| 984 | multimodal_embeddings=multimodal_embeddings, |
| 985 | is_multimodal=_require_is_multimodal(is_multimodal), |
| 986 | ) |
| 987 | |
| 988 | def _parse_and_validate_audio_input(self, **kwargs: object) -> FunASRAudioInputs: |
| 989 | input_features = kwargs.pop("input_features", None) |
| 990 | speech_lengths = kwargs.pop("speech_lengths", None) |
| 991 | fake_token_lengths = kwargs.pop("fake_token_lengths", None) |
| 992 | |
| 993 | return FunASRAudioInputs( |
| 994 | input_features=input_features, |
| 995 | speech_lengths=speech_lengths, |
| 996 | fake_token_lengths=fake_token_lengths, |
| 997 | ) |
| 998 | |
| 999 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 1000 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1001 | return logits |
| 1002 | |
| 1003 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1004 | loader = AutoWeightsLoader( |
| 1005 | self, |
| 1006 | ) |
| 1007 | |
| 1008 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |