| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterable, Mapping, Sequence |
| 5 | from typing import Annotated, Any, Literal, TypeAlias |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | from transformers import BatchFeature |
| 10 | from transformers.models.glmasr import GlmAsrConfig, GlmAsrProcessor |
| 11 | from transformers.models.whisper import WhisperFeatureExtractor |
| 12 | |
| 13 | from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig |
| 14 | from vllm.config.multimodal import MultiModalDummyOptions |
| 15 | from vllm.config.speech_to_text import SpeechToTextParams |
| 16 | from vllm.distributed.parallel_state import get_tensor_model_parallel_world_size |
| 17 | from vllm.inputs import ModalityData, MultiModalDataDict, PromptType, TokensPrompt |
| 18 | from vllm.model_executor.layers.activation import get_act_fn |
| 19 | from vllm.model_executor.layers.attention import MMEncoderAttention |
| 20 | from vllm.model_executor.layers.linear import ( |
| 21 | ColumnParallelLinear, |
| 22 | QKVParallelLinear, |
| 23 | RowParallelLinear, |
| 24 | ) |
| 25 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 26 | from vllm.model_executor.layers.rotary_embedding.common import ApplyRotaryEmb |
| 27 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 28 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 29 | from vllm.multimodal.inputs import ( |
| 30 | MultiModalFieldConfig, |
| 31 | MultiModalKwargsItems, |
| 32 | ) |
| 33 | from vllm.multimodal.parse import ( |
| 34 | DictEmbeddingItems, |
| 35 | ModalityDataItems, |
| 36 | MultiModalDataItems, |
| 37 | MultiModalDataParser, |
| 38 | ) |
| 39 | from vllm.multimodal.processing import ( |
| 40 | BaseDummyInputsBuilder, |
| 41 | BaseMultiModalProcessor, |
| 42 | BaseProcessingInfo, |
| 43 | PromptReplacement, |
| 44 | PromptUpdate, |
| 45 | PromptUpdateDetails, |
| 46 | ) |
| 47 | from vllm.multimodal.processing.processor import HFMultiModalInputs |
| 48 | from vllm.sequence import IntermediateTensors |
| 49 | from vllm.tokenizers import cached_tokenizer_from_config |
| 50 | from vllm.transformers_utils.processor import cached_processor_from_config |
| 51 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 52 | |
| 53 | from .glmasr_utils import ( |
| 54 | DEFAULT_CONV_PARAMS, |
| 55 | DEFAULT_MAX_AUDIO_LEN_S, |
| 56 | DEFAULT_MERGE_FACTOR, |
| 57 | _flatten_audio_features_by_length, |
| 58 | _get_audio_output_lengths_for_tower, |
| 59 | _group_audio_embeddings, |
| 60 | _normalize_chunk_counts, |
| 61 | ) |
| 62 | from .interfaces import ( |
| 63 | MultiModalEmbeddings, |
| 64 | SupportsLoRA, |
| 65 | SupportsMultiModal, |
| 66 | SupportsPP, |
| 67 | SupportsTranscription, |
| 68 | ) |
| 69 | from .utils import ( |
| 70 | AutoWeightsLoader, |
| 71 | WeightsMapper, |
| 72 | init_vllm_registered_model, |
| 73 | maybe_prefix, |
| 74 | ) |
| 75 | from .whisper import ISO639_1_SUPPORTED_LANGS, _create_fake_bias_for_k_proj |
| 76 | |
| 77 | |
| 78 | class GlmAsrEncoderRotaryEmbedding(nn.Module): |
| 79 | """Rotary Position Embedding for GLM-ASR encoder. |
| 80 | |
| 81 | Computes rotary position embeddings on-demand for efficiency. |
| 82 | Only caches inv_freq as a buffer; cos/sin are computed during forward |
| 83 | to avoid wasted computation during initialization and ensure correct |
| 84 | device placement. |
| 85 | """ |
| 86 | |
| 87 | def __init__(self, config) -> None: |
| 88 | super().__init__() |
| 89 | |
| 90 | # Compute inverse frequencies following transformers implementation |
| 91 | head_dim = getattr( |
| 92 | config, "head_dim", config.hidden_size // config.num_attention_heads |
| 93 | ) |
| 94 | |
| 95 | # Handle rope_parameters if present (for compatibility with transformers config) |
| 96 | if hasattr(config, "rope_parameters") and config.rope_parameters: |
| 97 | base = config.rope_parameters.get("rope_theta", 10000.0) |
| 98 | partial_rotary_factor = config.rope_parameters.get( |
| 99 | "partial_rotary_factor", 1.0 |
| 100 | ) |
| 101 | dim = int(head_dim * partial_rotary_factor) |
| 102 | self.attention_scaling = config.rope_parameters.get( |
| 103 | "attention_scaling", 1.0 |
| 104 | ) |
| 105 | else: |
| 106 | base = getattr(config, "rope_theta", 10000.0) |
| 107 | dim = head_dim |
| 108 | self.attention_scaling = 1.0 |
| 109 | |
| 110 | self.dim = dim |
| 111 | self.head_dim = head_dim |
| 112 | |
| 113 | # Only cache inv_freq; cos/sin computed on-demand in correct device |
| 114 | inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float) / dim)) |
| 115 | self.register_buffer("inv_freq", inv_freq, persistent=False) |
| 116 | |
| 117 | def forward(self, seq_len: int) -> torch.Tensor: |
| 118 | """Compute rotary position frequencies for given sequence length. |
| 119 | |
| 120 | Args: |
| 121 | seq_len: The sequence length to compute embeddings for. |
| 122 | |
| 123 | Returns: |
| 124 | Frequency tensor with shape [seq_len, dim/2]. Use .cos() and |
| 125 | .sin() to get the rotary embedding components. |
| 126 | |
| 127 | """ |
| 128 | # Compute on the same device as inv_freq (automatically correct after .to()) |
| 129 | seq = torch.arange( |
| 130 | seq_len, device=self.inv_freq.device, dtype=self.inv_freq.dtype |
| 131 | ) |
| 132 | freqs = torch.outer(seq, self.inv_freq) |
| 133 | return freqs * self.attention_scaling |
| 134 | |
| 135 | |
| 136 | class GlmAsrEncoderAttention(nn.Module): |
| 137 | """Optimized Multi-headed Grouped Query Attention for GLM-ASR encoder. |
| 138 | |
| 139 | Uses vLLM's QKVParallelLinear for fused projections, ApplyRotaryEmb for |
| 140 | rotary position embeddings, and MMEncoderAttention for hardware-optimized |
| 141 | attention computation with automatic backend selection. |
| 142 | """ |
| 143 | |
| 144 | def __init__( |
| 145 | self, |
| 146 | config, |
| 147 | quant_config: QuantizationConfig | None = None, |
| 148 | prefix: str = "", |
| 149 | ): |
| 150 | super().__init__() |
| 151 | self.config = config |
| 152 | self.hidden_size = config.hidden_size |
| 153 | self.num_heads = config.num_attention_heads |
| 154 | self.num_kv_heads = getattr( |
| 155 | config, "num_key_value_heads", config.num_attention_heads |
| 156 | ) |
| 157 | self.head_dim = self.hidden_size // self.num_heads |
| 158 | |
| 159 | self.tp_size = get_tensor_model_parallel_world_size() |
| 160 | self.num_heads_per_rank = self.num_heads // self.tp_size |
| 161 | self.num_kv_heads_per_rank = max(1, self.num_kv_heads // self.tp_size) |
| 162 | |
| 163 | # Use QKVParallelLinear for fused QKV projection |
| 164 | # Note: GLM-ASR uses bias on Q and V, but not K |
| 165 | # For simplicity with QKVParallelLinear, we use bias=True for all |
| 166 | self.qkv_proj = QKVParallelLinear( |
| 167 | self.hidden_size, |
| 168 | self.head_dim, |
| 169 | self.num_heads, |
| 170 | self.num_kv_heads, |
| 171 | bias=True, |
| 172 | quant_config=quant_config, |
| 173 | prefix=f"{prefix}.qkv_proj", |
| 174 | ) |
| 175 | |
| 176 | self.o_proj = RowParallelLinear( |
| 177 | self.hidden_size, |
| 178 | self.hidden_size, |
| 179 | bias=True, |
| 180 | quant_config=quant_config, |
| 181 | prefix=f"{prefix}.o_proj", |
| 182 | ) |
| 183 | |
| 184 | # Use vLLM's ApplyRotaryEmb CustomOp |
| 185 | # enforce_enable=True ensures the op is always enabled (important for ViT) |
| 186 | rope_params = getattr(config, "rope_parameters", None) |
| 187 | if rope_params: |
| 188 | partial_rotary_factor = rope_params.get("partial_rotary_factor", 0.5) |
| 189 | else: |
| 190 | partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5) |
| 191 | self.rotary_dim = int(self.head_dim * partial_rotary_factor) |
| 192 | self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True) |
| 193 | |
| 194 | # Use vLLM's MMEncoderAttention for hardware-optimized attention |
| 195 | # Automatically selects Flash Attention, SDPA, or Pallas based on device |
| 196 | self.attn = MMEncoderAttention( |
| 197 | num_heads=self.num_heads_per_rank, |
| 198 | head_size=self.head_dim, |
| 199 | scale=self.head_dim**-0.5, |
| 200 | num_kv_heads=self.num_kv_heads_per_rank, |
| 201 | prefix=f"{prefix}.attn", |
| 202 | ) |
| 203 | |
| 204 | def forward( |
| 205 | self, |
| 206 | hidden_states: torch.Tensor, |
| 207 | rotary_pos_emb_cos: torch.Tensor, |
| 208 | rotary_pos_emb_sin: torch.Tensor, |
| 209 | ) -> torch.Tensor: |
| 210 | """Args: |
| 211 | hidden_states: [batch_size, seq_len, hidden_size] |
| 212 | rotary_pos_emb_cos: [seq_len, rotary_dim/2] - cosine of rotary embeddings |
| 213 | rotary_pos_emb_sin: [seq_len, rotary_dim/2] - sine of rotary embeddings |
| 214 | |
| 215 | Returns: |
| 216 | [batch_size, seq_len, hidden_size] |
| 217 | |
| 218 | """ |
| 219 | batch_size, seq_len, _ = hidden_states.shape |
| 220 | |
| 221 | # QKV projection - fused for efficiency |
| 222 | qkv, _ = self.qkv_proj(hidden_states) |
| 223 | |
| 224 | # Split into q, k, v |
| 225 | q_size = self.num_heads_per_rank * self.head_dim |
| 226 | kv_size = self.num_kv_heads_per_rank * self.head_dim |
| 227 | q, k, v = qkv.split([q_size, kv_size, kv_size], dim=-1) |
| 228 | |
| 229 | # Reshape to [batch, seq, num_heads, head_dim] for ApplyRotaryEmb |
| 230 | q = q.view(batch_size, seq_len, self.num_heads_per_rank, self.head_dim) |
| 231 | k = k.view(batch_size, seq_len, self.num_kv_heads_per_rank, self.head_dim) |
| 232 | v = v.view(batch_size, seq_len, self.num_kv_heads_per_rank, self.head_dim) |
| 233 | |
| 234 | # Apply rotary position embeddings using vLLM's ApplyRotaryEmb |
| 235 | # ApplyRotaryEmb expects x: [batch, seq, heads, head_dim] |
| 236 | # cos/sin: [seq_len, rotary_dim/2] |
| 237 | q[..., : self.rotary_dim] = self.apply_rotary_emb( |
| 238 | q[..., : self.rotary_dim], rotary_pos_emb_cos, rotary_pos_emb_sin |
| 239 | ) |
| 240 | k[..., : self.rotary_dim] = self.apply_rotary_emb( |
| 241 | k[..., : self.rotary_dim], rotary_pos_emb_cos, rotary_pos_emb_sin |
| 242 | ) |
| 243 | |
| 244 | # MMEncoderAttention expects [batch, seq, num_heads, head_dim] |
| 245 | # It handles GQA internally via repeat_interleave |
| 246 | attn_output = self.attn(q, k, v) |
| 247 | |
| 248 | # Reshape back to [batch, seq, hidden_size] |
| 249 | attn_output = attn_output.view(batch_size, seq_len, -1) |
| 250 | |
| 251 | # Output projection |
| 252 | output, _ = self.o_proj(attn_output) |
| 253 | return output |
| 254 | |
| 255 | |
| 256 | class GlmAsrEncoderMLP(nn.Module): |
| 257 | """Optimized MLP for GLM-ASR encoder. |
| 258 | Uses vLLM's parallel linear layers for better performance. |
| 259 | """ |
| 260 | |
| 261 | def __init__( |
| 262 | self, |
| 263 | config, |
| 264 | quant_config: QuantizationConfig | None = None, |
| 265 | prefix: str = "", |
| 266 | ): |
| 267 | super().__init__() |
| 268 | self.config = config |
| 269 | self.hidden_size = config.hidden_size |
| 270 | self.intermediate_size = config.intermediate_size |
| 271 | |
| 272 | self.fc1 = ColumnParallelLinear( |
| 273 | self.hidden_size, |
| 274 | self.intermediate_size, |
| 275 | bias=True, |
| 276 | quant_config=quant_config, |
| 277 | prefix=f"{prefix}.fc1", |
| 278 | ) |
| 279 | |
| 280 | self.act_fn = get_act_fn(config.hidden_act) |
| 281 | |
| 282 | self.fc2 = RowParallelLinear( |
| 283 | self.intermediate_size, |
| 284 | self.hidden_size, |
| 285 | bias=True, |
| 286 | quant_config=quant_config, |
| 287 | prefix=f"{prefix}.fc2", |
| 288 | ) |
| 289 | |
| 290 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 291 | hidden_states, _ = self.fc1(hidden_states) |
| 292 | hidden_states = self.act_fn(hidden_states) |
| 293 | hidden_states, _ = self.fc2(hidden_states) |
| 294 | return hidden_states |
| 295 | |
| 296 | |
| 297 | class GlmAsrEncoderLayer(nn.Module): |
| 298 | """Optimized Transformer encoder layer for GLM-ASR. |
| 299 | Combines attention and MLP with residual connections and layer norms. |
| 300 | """ |
| 301 | |
| 302 | def __init__( |
| 303 | self, |
| 304 | config, |
| 305 | quant_config: QuantizationConfig | None = None, |
| 306 | prefix: str = "", |
| 307 | ): |
| 308 | super().__init__() |
| 309 | self.hidden_size = config.hidden_size |
| 310 | |
| 311 | self.self_attn = GlmAsrEncoderAttention( |
| 312 | config, |
| 313 | quant_config=quant_config, |
| 314 | prefix=f"{prefix}.self_attn", |
| 315 | ) |
| 316 | |
| 317 | self.mlp = GlmAsrEncoderMLP( |
| 318 | config, |
| 319 | quant_config=quant_config, |
| 320 | prefix=f"{prefix}.mlp", |
| 321 | ) |
| 322 | |
| 323 | layer_norm_eps = getattr(config, "layer_norm_eps", 1e-5) |
| 324 | self.input_layernorm = nn.LayerNorm(self.hidden_size, eps=layer_norm_eps) |
| 325 | self.post_attention_layernorm = nn.LayerNorm( |
| 326 | self.hidden_size, eps=layer_norm_eps |
| 327 | ) |
| 328 | |
| 329 | def forward( |
| 330 | self, |
| 331 | hidden_states: torch.Tensor, |
| 332 | rotary_pos_emb_cos: torch.Tensor, |
| 333 | rotary_pos_emb_sin: torch.Tensor, |
| 334 | ) -> torch.Tensor: |
| 335 | """Args: |
| 336 | hidden_states: [batch_size, seq_len, hidden_size] |
| 337 | rotary_pos_emb_cos: [seq_len, rotary_dim/2] - cosine of rotary embeddings |
| 338 | rotary_pos_emb_sin: [seq_len, rotary_dim/2] - sine of rotary embeddings |
| 339 | |
| 340 | Returns: |
| 341 | [batch_size, seq_len, hidden_size] |
| 342 | |
| 343 | """ |
| 344 | # Self-attention with residual |
| 345 | residual = hidden_states |
| 346 | hidden_states = self.input_layernorm(hidden_states) |
| 347 | hidden_states = self.self_attn( |
| 348 | hidden_states=hidden_states, |
| 349 | rotary_pos_emb_cos=rotary_pos_emb_cos, |
| 350 | rotary_pos_emb_sin=rotary_pos_emb_sin, |
| 351 | ) |
| 352 | hidden_states = residual + hidden_states |
| 353 | |
| 354 | # MLP with residual |
| 355 | residual = hidden_states |
| 356 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 357 | hidden_states = self.mlp(hidden_states) |
| 358 | hidden_states = residual + hidden_states |
| 359 | |
| 360 | return hidden_states |
| 361 | |
| 362 | |
| 363 | class _GlmAsrEncoderOutput: |
| 364 | """Simple output container compatible with transformers' BaseModelOutput. |
| 365 | |
| 366 | This lightweight container holds the encoder output and is compatible |
| 367 | with the transformers library's output format while being more efficient |
| 368 | than a full dataclass. |
| 369 | |
| 370 | Attributes: |
| 371 | last_hidden_state: Final layer hidden states from the encoder. |
| 372 | Shape: [batch_size, seq_len, hidden_size] |
| 373 | |
| 374 | """ |
| 375 | |
| 376 | __slots__ = ("last_hidden_state",) |
| 377 | |
| 378 | def __init__(self, last_hidden_state: torch.Tensor): |
| 379 | self.last_hidden_state = last_hidden_state |
| 380 | |
| 381 | |
| 382 | class GlmAsrEncoder(nn.Module): |
| 383 | """Optimized GLM-ASR Audio Encoder with vLLM native implementation. |
| 384 | |
| 385 | This encoder processes audio features through convolutional layers |
| 386 | followed by transformer layers with rotary position embeddings. |
| 387 | Optimized for performance with: |
| 388 | - QKVParallelLinear for fused attention projections |
| 389 | - Tensor parallelism support via ColumnParallelLinear/RowParallelLinear |
| 390 | - Quantization support |
| 391 | - Flash Attention (SDPA) |
| 392 | """ |
| 393 | |
| 394 | # Mapping for weight loading: transformers uses separate q/k/v, we use fused qkv |
| 395 | packed_modules_mapping = { |
| 396 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 397 | } |
| 398 | |
| 399 | hf_to_vllm_mapper = WeightsMapper( |
| 400 | orig_to_new_stacked={ |
| 401 | ".q_proj": (".qkv_proj", "q"), |
| 402 | ".k_proj": (".qkv_proj", "k"), |
| 403 | ".v_proj": (".qkv_proj", "v"), |
| 404 | } |
| 405 | ) |
| 406 | |
| 407 | def __init__( |
| 408 | self, |
| 409 | config, |
| 410 | quant_config: QuantizationConfig | None = None, |
| 411 | prefix: str = "", |
| 412 | ): |
| 413 | super().__init__() |
| 414 | self.config = config |
| 415 | |
| 416 | # Convolutional feature extraction layers |
| 417 | self.conv1 = nn.Conv1d( |
| 418 | config.num_mel_bins, |
| 419 | config.hidden_size, |
| 420 | kernel_size=3, |
| 421 | padding=1, |
| 422 | ) |
| 423 | self.conv2 = nn.Conv1d( |
| 424 | config.hidden_size, |
| 425 | config.hidden_size, |
| 426 | kernel_size=3, |
| 427 | stride=2, |
| 428 | padding=1, |
| 429 | ) |
| 430 | |
| 431 | # Transformer encoder layers |
| 432 | self.layers = nn.ModuleList( |
| 433 | [ |
| 434 | GlmAsrEncoderLayer( |
| 435 | config, |
| 436 | quant_config=quant_config, |
| 437 | prefix=f"{prefix}.layers.{layer_idx}", |
| 438 | ) |
| 439 | for layer_idx in range(config.num_hidden_layers) |
| 440 | ] |
| 441 | ) |
| 442 | |
| 443 | # Final layer norm |
| 444 | layer_norm_eps = getattr(config, "layer_norm_eps", 1e-5) |
| 445 | self.norm = nn.LayerNorm(config.hidden_size, eps=layer_norm_eps) |
| 446 | |
| 447 | # Rotary position embeddings |
| 448 | self.rotary_emb = GlmAsrEncoderRotaryEmbedding(config) |
| 449 | |
| 450 | def _get_feat_extract_output_lengths( |
| 451 | self, input_lengths: torch.Tensor |
| 452 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 453 | """Compute the output length after convolutions. |
| 454 | |
| 455 | Args: |
| 456 | input_lengths: Input sequence lengths [batch_size] |
| 457 | |
| 458 | Returns: |
| 459 | Tuple of (output after conv1, output after conv2) |
| 460 | |
| 461 | """ |
| 462 | # Conv1: kernel=3, stride=1, padding=1 |
| 463 | output_lengths_conv1 = (input_lengths + 2 * 1 - 3) // 1 + 1 |
| 464 | |
| 465 | # Conv2: kernel=3, stride=2, padding=1 |
| 466 | output_lengths_conv2 = (output_lengths_conv1 + 2 * 1 - 3) // 2 + 1 |
| 467 | |
| 468 | return output_lengths_conv1, output_lengths_conv2 |
| 469 | |
| 470 | def forward(self, input_features: torch.Tensor) -> _GlmAsrEncoderOutput: |
| 471 | """Forward pass through the encoder. |
| 472 | |
| 473 | Args: |
| 474 | input_features: [batch_size, num_mel_bins, seq_len] |
| 475 | |
| 476 | Returns: |
| 477 | _GlmAsrEncoderOutput: Object with .last_hidden_state attribute \ |
| 478 | containing [batch_size, seq_len', hidden_size] where seq_len' \ |
| 479 | is the sequence length after convolutions |
| 480 | |
| 481 | """ |
| 482 | # Apply convolutional layers with GELU activation |
| 483 | hidden_states = torch.nn.functional.gelu(self.conv1(input_features)) |
| 484 | hidden_states = torch.nn.functional.gelu(self.conv2(hidden_states)) |
| 485 | |
| 486 | # Transpose to [batch_size, seq_len, hidden_size] |
| 487 | hidden_states = hidden_states.transpose(1, 2) |
| 488 | output_seq_len = hidden_states.shape[1] |
| 489 | |
| 490 | # Compute rotary position embeddings on-demand |
| 491 | rotary_pos_emb = self.rotary_emb(output_seq_len) |
| 492 | rotary_pos_emb_cos = rotary_pos_emb.cos().to(dtype=hidden_states.dtype) |
| 493 | rotary_pos_emb_sin = rotary_pos_emb.sin().to(dtype=hidden_states.dtype) |
| 494 | |
| 495 | # Apply transformer layers |
| 496 | for encoder_layer in self.layers: |
| 497 | hidden_states = encoder_layer( |
| 498 | hidden_states, rotary_pos_emb_cos, rotary_pos_emb_sin |
| 499 | ) |
| 500 | |
| 501 | # Final layer norm |
| 502 | hidden_states = self.norm(hidden_states) |
| 503 | |
| 504 | # Return in a format compatible with transformers' BaseModelOutput |
| 505 | return _GlmAsrEncoderOutput(last_hidden_state=hidden_states) |
| 506 | |
| 507 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 508 | weights = _create_fake_bias_for_k_proj(weights, ".k_proj.weight") |
| 509 | loader = AutoWeightsLoader(self) |
| 510 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 511 | |
| 512 | |
| 513 | class GlmAsrFeatureInputs(TensorSchema): |
| 514 | """Dimensions: |
| 515 | - num_chunks: Number of audio chunks (flattened) |
| 516 | - nmb: Number of mel bins |
| 517 | - num_audios: Number of original audio files |
| 518 | """ |
| 519 | |
| 520 | type: Literal["audio_features"] |
| 521 | input_features: Annotated[ |
| 522 | torch.Tensor | list[torch.Tensor], |
| 523 | TensorShape("num_chunks", "nmb", "chunk_length", dynamic_dims={"chunk_length"}), |
| 524 | ] |
| 525 | feature_attention_mask: Annotated[ |
| 526 | torch.Tensor | list[torch.Tensor], |
| 527 | TensorShape("num_chunks", "chunk_length", dynamic_dims={"chunk_length"}), |
| 528 | ] |
| 529 | chunk_counts: Annotated[ |
| 530 | torch.Tensor | list[torch.Tensor], |
| 531 | TensorShape("num_audios"), |
| 532 | ] |
| 533 | |
| 534 | |
| 535 | class GlmAsrEmbeddingInputs(TensorSchema): |
| 536 | """Dimensions: |
| 537 | - bn: Batch size |
| 538 | - naf: Number of audio features |
| 539 | - hs: Hidden size (must match the hidden size of language model |
| 540 | backbone) |
| 541 | """ |
| 542 | |
| 543 | type: Literal["audio_embeds"] = "audio_embeds" |
| 544 | audio_embeds: Annotated[ |
| 545 | list[torch.Tensor], |
| 546 | TensorShape("bn", "naf", "hs", dynamic_dims={"naf"}), |
| 547 | ] |
| 548 | |
| 549 | |
| 550 | GlmAsrInputs: TypeAlias = GlmAsrFeatureInputs | GlmAsrEmbeddingInputs |
| 551 | |
| 552 | |
| 553 | class GlmAsrMultiModalProjector(nn.Module): |
| 554 | """Projects audio encoder outputs to language model hidden space. |
| 555 | |
| 556 | This projector uses a two-layer MLP to map audio features from the |
| 557 | encoder's intermediate size to the language model's hidden size. |
| 558 | Uses vLLM's parallel linear layers for tensor parallelism support. |
| 559 | |
| 560 | Architecture: |
| 561 | - Linear layer: intermediate_size -> hidden_size * 2 |
| 562 | - Activation function (e.g., GELU) |
| 563 | - Linear layer: hidden_size * 2 -> hidden_size |
| 564 | """ |
| 565 | |
| 566 | def __init__( |
| 567 | self, |
| 568 | config: GlmAsrConfig, |
| 569 | quant_config: QuantizationConfig | None = None, |
| 570 | prefix: str = "", |
| 571 | ): |
| 572 | super().__init__() |
| 573 | self.linear_1 = ColumnParallelLinear( |
| 574 | input_size=config.audio_config.intermediate_size, |
| 575 | output_size=config.text_config.hidden_size * 2, |
| 576 | quant_config=quant_config, |
| 577 | prefix=f"{prefix}.linear_1", |
| 578 | ) |
| 579 | self.act = get_act_fn(config.projector_hidden_act) |
| 580 | self.linear_2 = RowParallelLinear( |
| 581 | input_size=config.text_config.hidden_size * 2, |
| 582 | output_size=config.text_config.hidden_size, |
| 583 | quant_config=quant_config, |
| 584 | prefix=f"{prefix}.linear_2", |
| 585 | ) |
| 586 | |
| 587 | def forward(self, audio_features: torch.Tensor) -> torch.Tensor: |
| 588 | hidden_states, _ = self.linear_1(audio_features) |
| 589 | hidden_states = self.act(hidden_states) |
| 590 | hidden_states, _ = self.linear_2(hidden_states) |
| 591 | return hidden_states |
| 592 | |
| 593 | |
| 594 | def _glmasr_field_config( |
| 595 | hf_inputs: Mapping[str, torch.Tensor], |
| 596 | ) -> dict[str, MultiModalFieldConfig]: |
| 597 | """Configure multimodal field batching strategy for GLM-ASR. |
| 598 | |
| 599 | Determines how to batch audio inputs based on whether chunking is used. |
| 600 | When chunk_counts is present, features are flattened across chunks; |
| 601 | otherwise, they are batched normally. |
| 602 | |
| 603 | Args: |
| 604 | hf_inputs: Dictionary of preprocessed inputs from HuggingFace processor. |
| 605 | |
| 606 | Returns: |
| 607 | Dictionary mapping field names to MultiModalFieldConfig objects \ |
| 608 | that specify batching behavior. |
| 609 | |
| 610 | """ |
| 611 | chunk_counts = hf_inputs.get("chunk_counts") |
| 612 | if chunk_counts is not None: |
| 613 | return dict( |
| 614 | audio_embeds=MultiModalFieldConfig.batched("audio"), |
| 615 | input_features=MultiModalFieldConfig.flat_from_sizes( |
| 616 | "audio", chunk_counts, dim=0 |
| 617 | ), |
| 618 | feature_attention_mask=MultiModalFieldConfig.flat_from_sizes( |
| 619 | "audio", chunk_counts, dim=0 |
| 620 | ), |
| 621 | chunk_counts=MultiModalFieldConfig.batched("audio", keep_on_cpu=True), |
| 622 | ) |
| 623 | return dict( |
| 624 | audio_embeds=MultiModalFieldConfig.batched("audio"), |
| 625 | input_features=MultiModalFieldConfig.batched("audio"), |
| 626 | feature_attention_mask=MultiModalFieldConfig.batched("audio"), |
| 627 | chunk_counts=MultiModalFieldConfig.batched("audio", keep_on_cpu=True), |
| 628 | ) |
| 629 | |
| 630 | |
| 631 | class GlmAsrMultiModalDataParser(MultiModalDataParser): |
| 632 | """Custom parser for GLM-ASR multimodal data. |
| 633 | |
| 634 | Extends the base parser to handle GLM-ASR specific audio data formats, |
| 635 | including both pre-computed audio embeddings and raw audio features. |
| 636 | """ |
| 637 | |
| 638 | def _parse_audio_data( |
| 639 | self, |
| 640 | data: dict[str, torch.Tensor] | ModalityData[Any], |
| 641 | ) -> ModalityDataItems[Any, Any] | None: |
| 642 | if isinstance(data, dict): |
| 643 | return DictEmbeddingItems( |
| 644 | data, |
| 645 | modality="audio", |
| 646 | required_fields={"audio_embeds"}, |
| 647 | fields_factory=_glmasr_field_config, |
| 648 | ) |
| 649 | return super()._parse_audio_data(data) |
| 650 | |
| 651 | |
| 652 | class GlmAsrProcessingInfo(BaseProcessingInfo): |
| 653 | """Processing information provider for GLM-ASR model. |
| 654 | |
| 655 | Provides access to model configuration, processor, and feature extractor |
| 656 | needed for audio preprocessing and multimodal integration. |
| 657 | """ |
| 658 | |
| 659 | def get_hf_config(self) -> GlmAsrConfig: |
| 660 | return self.ctx.get_hf_config(GlmAsrConfig) |
| 661 | |
| 662 | def get_hf_processor(self, **kwargs: object) -> GlmAsrProcessor: |
| 663 | return self.ctx.get_hf_processor(GlmAsrProcessor, **kwargs) |
| 664 | |
| 665 | def get_feature_extractor(self, **kwargs: object) -> WhisperFeatureExtractor: |
| 666 | return self.get_hf_processor(**kwargs).feature_extractor |
| 667 | |
| 668 | def get_data_parser(self): |
| 669 | feature_extractor = self.get_feature_extractor() |
| 670 | |
| 671 | return GlmAsrMultiModalDataParser( |
| 672 | target_sr=feature_extractor.sampling_rate, |
| 673 | expected_hidden_size=self._get_expected_hidden_size(), |
| 674 | ) |
| 675 | |
| 676 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 677 | return {"audio": None} |
| 678 | |
| 679 | |
| 680 | class GlmAsrDummyInputsBuilder(BaseDummyInputsBuilder[GlmAsrProcessingInfo]): |
| 681 | """Builder for dummy inputs used in profiling and testing. |
| 682 | |
| 683 | Generates dummy text prompts and audio data that match the expected |
| 684 | format for GLM-ASR model inputs. Used for memory profiling and |
| 685 | performance benchmarking. |
| 686 | """ |
| 687 | |
| 688 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 689 | num_audios = mm_counts.get("audio", 0) |
| 690 | hf_processor = self.info.get_hf_processor() |
| 691 | return hf_processor.audio_token * num_audios |
| 692 | |
| 693 | def get_dummy_mm_data( |
| 694 | self, |
| 695 | seq_len: int, |
| 696 | mm_counts: Mapping[str, int], |
| 697 | mm_options: MultiModalDummyOptions, |
| 698 | ) -> MultiModalDataDict: |
| 699 | feature_extractor = self.info.get_feature_extractor() |
| 700 | sampling_rate = feature_extractor.sampling_rate |
| 701 | |
| 702 | max_audio_len = getattr( |
| 703 | self.info.get_hf_processor(), "max_audio_len", DEFAULT_MAX_AUDIO_LEN_S |
| 704 | ) |
| 705 | audio_len = int(max_audio_len * sampling_rate) |
| 706 | |
| 707 | return { |
| 708 | "audio": self._get_dummy_audios( |
| 709 | length=audio_len, |
| 710 | num_audios=mm_counts.get("audio", 0), |
| 711 | overrides=mm_options.get("audio"), |
| 712 | ) |
| 713 | } |
| 714 | |
| 715 | |
| 716 | class GlmAsrMultiModalProcessor(BaseMultiModalProcessor["GlmAsrProcessingInfo"]): |
| 717 | """GLM-ASR processor that inherits directly from BaseMultiModalProcessor |
| 718 | for better performance and cleaner implementation. |
| 719 | """ |
| 720 | |
| 721 | def _calculate_chunk_counts( |
| 722 | self, |
| 723 | audio_list: list[Any], |
| 724 | feature_extractor: WhisperFeatureExtractor, |
| 725 | processor: GlmAsrProcessor, |
| 726 | ) -> list[int]: |
| 727 | sampling_rate = feature_extractor.sampling_rate |
| 728 | chunk_length = feature_extractor.chunk_length |
| 729 | max_audio_len = getattr(processor, "max_audio_len", DEFAULT_MAX_AUDIO_LEN_S) |
| 730 | window_size = int(sampling_rate * chunk_length) |
| 731 | max_windows = int(max_audio_len // chunk_length) |
| 732 | |
| 733 | chunk_counts = [] |
| 734 | for audio in audio_list: |
| 735 | n_samples = len(audio) if isinstance(audio, list) else audio.shape[0] |
| 736 | n_chunks = max(1, (n_samples + window_size - 1) // window_size) |
| 737 | chunk_counts.append(min(n_chunks, max_windows)) |
| 738 | return chunk_counts |
| 739 | |
| 740 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 741 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 742 | |
| 743 | def _get_hf_mm_inputs( |
| 744 | self, |
| 745 | mm_items: MultiModalDataItems, |
| 746 | hf_kwargs: Mapping[str, object], |
| 747 | ) -> HFMultiModalInputs: |
| 748 | hf_inputs = super()._get_hf_mm_inputs(mm_items, hf_kwargs) |
| 749 | |
| 750 | feature_extractor = self.info.get_feature_extractor(**hf_kwargs) |
| 751 | return hf_inputs._replace( |
| 752 | hf_kwargs=dict( |
| 753 | hf_inputs.hf_kwargs, |
| 754 | sampling_rate=feature_extractor.sampling_rate, |
| 755 | ) |
| 756 | ) |
| 757 | |
| 758 | def _postprocess_hf_mm_data( |
| 759 | self, |
| 760 | mm_data: Mapping[str, object], |
| 761 | hf_processor_mm_kwargs: Mapping[str, object], |
| 762 | processed_data: BatchFeature, |
| 763 | ) -> BatchFeature: |
| 764 | # Handle different key names from different transformers versions |
| 765 | if "input_features_mask" in processed_data: |
| 766 | processed_data["feature_attention_mask"] = processed_data.pop( |
| 767 | "input_features_mask" |
| 768 | ) |
| 769 | elif "input_features" in processed_data: |
| 770 | # If no mask is provided, create one from input_features |
| 771 | input_features = processed_data["input_features"] |
| 772 | if isinstance(input_features, torch.Tensor): |
| 773 | # Create a mask of all ones matching the sequence length |
| 774 | mask = torch.ones( |
| 775 | input_features.shape[0], |
| 776 | input_features.shape[-1], |
| 777 | dtype=torch.long, |
| 778 | ) |
| 779 | processed_data["feature_attention_mask"] = mask |
| 780 | |
| 781 | audio = mm_data.get("audio") |
| 782 | if audio is None: |
| 783 | audio_list: list[Any] = [] |
| 784 | elif isinstance(audio, list): |
| 785 | audio_list = audio |
| 786 | else: |
| 787 | audio_list = [audio] |
| 788 | if audio_list: |
| 789 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 790 | |
| 791 | # Override chunk counts calculation with GLM-ASR specific logic |
| 792 | chunk_counts = self._calculate_chunk_counts( |
| 793 | audio_list, processor.feature_extractor, processor |
| 794 | ) |
| 795 | processed_data["chunk_counts"] = torch.tensor( |
| 796 | chunk_counts, dtype=torch.long |
| 797 | ) |
| 798 | |
| 799 | return processed_data |
| 800 | |
| 801 | def _get_mm_fields_config( |
| 802 | self, |
| 803 | hf_inputs: BatchFeature, |
| 804 | hf_processor_mm_kwargs: Mapping[str, object], |
| 805 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 806 | return _glmasr_field_config(hf_inputs) |
| 807 | |
| 808 | def _get_prompt_updates( |
| 809 | self, |
| 810 | mm_items: MultiModalDataItems, |
| 811 | hf_processor_mm_kwargs: Mapping[str, object], |
| 812 | out_mm_kwargs: MultiModalKwargsItems, |
| 813 | ) -> Sequence[PromptUpdate]: |
| 814 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 815 | tokenizer = self.info.get_tokenizer() |
| 816 | vocab = tokenizer.get_vocab() |
| 817 | config = self.info.get_hf_config() |
| 818 | |
| 819 | audio_token = getattr(processor, "audio_token", "<|pad|>") |
| 820 | audio_token_id = vocab.get(audio_token) |
| 821 | if audio_token_id is None: |
| 822 | audio_token_id = processor.audio_token_id |
| 823 | |
| 824 | merge_factor = getattr(config, "merge_factor", DEFAULT_MERGE_FACTOR) |
| 825 | conv_params = getattr(config, "conv_params", DEFAULT_CONV_PARAMS) |
| 826 | out_mm_data = out_mm_kwargs.get_data() |
| 827 | feature_attention_mask = out_mm_data.get("feature_attention_mask") |
| 828 | chunk_counts = out_mm_data.get("chunk_counts") |
| 829 | |
| 830 | # Pre-compute audio output lengths if feature_attention_mask is available |
| 831 | audio_output_lengths: list[int] = [] |
| 832 | if feature_attention_mask is not None: |
| 833 | # Compute output lengths for all audio items |
| 834 | from .glmasr_utils import ( |
| 835 | _as_list_chunk_counts, |
| 836 | _get_audio_output_lengths_from_mask, |
| 837 | ) |
| 838 | |
| 839 | if chunk_counts is not None: |
| 840 | start_idx = 0 |
| 841 | for count in _as_list_chunk_counts(chunk_counts): |
| 842 | end_idx = start_idx + count |
| 843 | mask = feature_attention_mask[start_idx:end_idx] |
| 844 | if isinstance(mask, list): |
| 845 | mask = torch.stack(mask) |
| 846 | |
| 847 | lengths = _get_audio_output_lengths_from_mask( |
| 848 | mask, merge_factor, conv_params |
| 849 | ) |
| 850 | audio_output_lengths.append(int(lengths.sum().item())) |
| 851 | start_idx = end_idx |
| 852 | else: |
| 853 | # Single chunk per audio |
| 854 | for idx in range(len(feature_attention_mask)): |
| 855 | mask = feature_attention_mask[idx : idx + 1] |
| 856 | if isinstance(mask, list): |
| 857 | mask = torch.tensor(mask).unsqueeze(0) |
| 858 | lengths = _get_audio_output_lengths_from_mask( |
| 859 | mask, merge_factor, conv_params |
| 860 | ) |
| 861 | audio_output_lengths.append(int(lengths.sum().item())) |
| 862 | |
| 863 | def get_replacement_glmasr(item_idx: int): |
| 864 | # Use pre-computed lengths if available, otherwise fall back to audio_embeds |
| 865 | if audio_output_lengths: |
| 866 | num_features = audio_output_lengths[item_idx] |
| 867 | else: |
| 868 | audio_embeds = out_mm_data.get("audio_embeds") |
| 869 | if audio_embeds is not None: |
| 870 | embed = audio_embeds[item_idx] |
| 871 | assert isinstance(embed, torch.Tensor) |
| 872 | num_features = embed.shape[0] |
| 873 | else: |
| 874 | raise ValueError( |
| 875 | "Either feature_attention_mask or audio_embeds must be provided" |
| 876 | ) |
| 877 | |
| 878 | if num_features == 0: |
| 879 | raise ValueError("Audio is too short") |
| 880 | |
| 881 | audio_tokens = [audio_token_id] * int(num_features) |
| 882 | return PromptUpdateDetails.select_token_id( |
| 883 | audio_tokens, |
| 884 | embed_token_id=audio_token_id, |
| 885 | ) |
| 886 | |
| 887 | return [ |
| 888 | PromptReplacement( |
| 889 | modality="audio", |
| 890 | target=[audio_token_id], |
| 891 | replacement=get_replacement_glmasr, |
| 892 | ) |
| 893 | ] |
| 894 | |
| 895 | |
| 896 | @MULTIMODAL_REGISTRY.register_processor( |
| 897 | GlmAsrMultiModalProcessor, |
| 898 | info=GlmAsrProcessingInfo, |
| 899 | dummy_inputs=GlmAsrDummyInputsBuilder, |
| 900 | ) |
| 901 | class GlmAsrForConditionalGeneration( |
| 902 | nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsTranscription |
| 903 | ): |
| 904 | supported_languages = ISO639_1_SUPPORTED_LANGS |
| 905 | |
| 906 | hf_to_vllm_mapper = WeightsMapper( |
| 907 | orig_to_new_prefix={"audio_tower.embed_positions": None} |
| 908 | ) |
| 909 | |
| 910 | packed_modules_mapping = { |
| 911 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 912 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 913 | } |
| 914 | |
| 915 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 916 | super().__init__() |
| 917 | config = vllm_config.model_config.hf_config |
| 918 | quant_config = vllm_config.quant_config |
| 919 | multimodal_config = vllm_config.model_config.multimodal_config |
| 920 | self.config = config |
| 921 | self.multimodal_config = multimodal_config |
| 922 | self.quant_config = quant_config |
| 923 | |
| 924 | with self._mark_tower_model(vllm_config, "audio"): |
| 925 | self.audio_tower = GlmAsrEncoder( |
| 926 | config.audio_config, |
| 927 | quant_config=quant_config, |
| 928 | prefix=maybe_prefix(prefix, "audio_tower"), |
| 929 | ) |
| 930 | self.multi_modal_projector = GlmAsrMultiModalProjector( |
| 931 | config, |
| 932 | quant_config=quant_config, |
| 933 | prefix=maybe_prefix(prefix, "multi_modal_projector"), |
| 934 | ) |
| 935 | |
| 936 | with self._mark_language_model(vllm_config): |
| 937 | self.language_model = init_vllm_registered_model( |
| 938 | vllm_config=vllm_config, |
| 939 | hf_config=config.text_config, |
| 940 | prefix=maybe_prefix(prefix, "language_model"), |
| 941 | architectures=["LlamaForCausalLM"], |
| 942 | ) |
| 943 | |
| 944 | self.make_empty_intermediate_tensors = ( |
| 945 | self.language_model.make_empty_intermediate_tensors |
| 946 | ) |
| 947 | |
| 948 | @classmethod |
| 949 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 950 | if modality.startswith("audio"): |
| 951 | return "<|begin_of_audio|><|pad|><|end_of_audio|>" |
| 952 | |
| 953 | raise ValueError("Only audio modality is supported") |
| 954 | |
| 955 | def get_mm_mapping(self) -> MultiModelKeys: |
| 956 | return MultiModelKeys.from_string_field( |
| 957 | language_model="language_model.", |
| 958 | connector="multi_modal_projector.", |
| 959 | tower_model="audio_tower.", |
| 960 | ) |
| 961 | |
| 962 | def _parse_and_validate_audio_input(self, **kwargs: object) -> GlmAsrInputs | None: |
| 963 | audio_embeds = kwargs.pop("audio_embeds", None) |
| 964 | if audio_embeds is not None: |
| 965 | return GlmAsrEmbeddingInputs(type="audio_embeds", audio_embeds=audio_embeds) |
| 966 | |
| 967 | input_features = kwargs.pop("input_features", None) |
| 968 | if input_features is None: |
| 969 | return None |
| 970 | |
| 971 | return GlmAsrFeatureInputs( |
| 972 | type="audio_features", |
| 973 | input_features=input_features, |
| 974 | feature_attention_mask=kwargs.pop("feature_attention_mask", None), |
| 975 | chunk_counts=kwargs.pop("chunk_counts", None), |
| 976 | ) |
| 977 | |
| 978 | def _process_audio_input( |
| 979 | self, audio_input: GlmAsrInputs |
| 980 | ) -> torch.Tensor | tuple[torch.Tensor, ...]: |
| 981 | if audio_input["type"] == "audio_embeds": |
| 982 | return tuple(audio_input["audio_embeds"]) |
| 983 | |
| 984 | input_features = audio_input["input_features"] |
| 985 | feature_attention_mask = audio_input["feature_attention_mask"] |
| 986 | |
| 987 | if isinstance(input_features, list): |
| 988 | input_features = torch.cat(input_features, dim=0) |
| 989 | feature_attention_mask = torch.cat(feature_attention_mask, dim=0) |
| 990 | |
| 991 | num_chunks = input_features.shape[0] |
| 992 | chunk_counts = _normalize_chunk_counts( |
| 993 | audio_input.get("chunk_counts"), num_chunks=num_chunks |
| 994 | ) |
| 995 | |
| 996 | # Convert input_features to model dtype (e.g., bfloat16) to match model weights |
| 997 | input_features = input_features.to(dtype=self.audio_tower.conv1.weight.dtype) |
| 998 | |
| 999 | # audio_tower returns [batch_size, seq_len, hidden_size] where hidden_size=1280 |
| 1000 | audio_hidden_states = self.audio_tower(input_features).last_hidden_state |
| 1001 | |
| 1002 | # GLM-ASR merges consecutive frames: 4 frames with hidden_size=1280 |
| 1003 | # -> 1 frame with intermediate_size=5120 |
| 1004 | hidden_size = self.config.audio_config.hidden_size |
| 1005 | intermediate_size = self.config.audio_config.intermediate_size |
| 1006 | merge_ratio = intermediate_size // hidden_size |
| 1007 | |
| 1008 | # Truncate sequence length to be divisible by merge_ratio |
| 1009 | seq_len = audio_hidden_states.shape[1] |
| 1010 | seq_len_truncated = (seq_len // merge_ratio) * merge_ratio |
| 1011 | if seq_len_truncated < seq_len: |
| 1012 | audio_hidden_states = audio_hidden_states[:, :seq_len_truncated, :] |
| 1013 | |
| 1014 | # Reshape to merge consecutive frames |
| 1015 | audio_hidden_states = audio_hidden_states.reshape( |
| 1016 | num_chunks, |
| 1017 | -1, |
| 1018 | intermediate_size, |
| 1019 | ) |
| 1020 | |
| 1021 | audio_features = self.multi_modal_projector(audio_hidden_states) |
| 1022 | |
| 1023 | merge_factor = getattr(self.config, "merge_factor", DEFAULT_MERGE_FACTOR) |
| 1024 | conv_params = getattr(self.config, "conv_params", DEFAULT_CONV_PARAMS) |
| 1025 | |
| 1026 | audio_output_lengths = _get_audio_output_lengths_for_tower( |
| 1027 | self.audio_tower, |
| 1028 | feature_attention_mask.sum(-1), |
| 1029 | merge_factor, |
| 1030 | conv_params, |
| 1031 | ) |
| 1032 | |
| 1033 | masked_audio_features = _flatten_audio_features_by_length( |
| 1034 | audio_features, audio_output_lengths |
| 1035 | ) |
| 1036 | |
| 1037 | chunk_embeddings = torch.split( |
| 1038 | masked_audio_features, audio_output_lengths.flatten().tolist() |
| 1039 | ) |
| 1040 | return _group_audio_embeddings(chunk_embeddings, chunk_counts) |
| 1041 | |
| 1042 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 1043 | audio_input = self._parse_and_validate_audio_input(**kwargs) |
| 1044 | if audio_input is None: |
| 1045 | return [] |
| 1046 | |
| 1047 | masked_audio_features = self._process_audio_input(audio_input) |
| 1048 | |
| 1049 | return masked_audio_features |
| 1050 | |
| 1051 | def forward( |
| 1052 | self, |
| 1053 | input_ids: torch.Tensor | None, |
| 1054 | positions: torch.Tensor, |
| 1055 | intermediate_tensors: IntermediateTensors | None = None, |
| 1056 | inputs_embeds: torch.Tensor | None = None, |
| 1057 | **kwargs: object, |
| 1058 | ) -> torch.Tensor | IntermediateTensors: |
| 1059 | if intermediate_tensors is not None: |
| 1060 | inputs_embeds = None |
| 1061 | |
| 1062 | hidden_states = self.language_model.model( |
| 1063 | input_ids, |
| 1064 | positions, |
| 1065 | intermediate_tensors, |
| 1066 | inputs_embeds=inputs_embeds, |
| 1067 | ) |
| 1068 | return hidden_states |
| 1069 | |
| 1070 | def compute_logits( |
| 1071 | self, |
| 1072 | hidden_states: torch.Tensor, |
| 1073 | ) -> torch.Tensor | None: |
| 1074 | return self.language_model.compute_logits(hidden_states) |
| 1075 | |
| 1076 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1077 | loader = AutoWeightsLoader(self) |
| 1078 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1079 | |
| 1080 | @classmethod |
| 1081 | def _get_audio_token(cls, model_config: ModelConfig) -> str: |
| 1082 | """Get the audio token from processor. |
| 1083 | |
| 1084 | Similar to get_placeholder_str but returns single token. |
| 1085 | """ |
| 1086 | processor = cached_processor_from_config(model_config) |
| 1087 | return getattr(processor, "audio_token", "<|pad|>") |
| 1088 | |
| 1089 | @classmethod |
| 1090 | def get_speech_to_text_config( |
| 1091 | cls, model_config: ModelConfig, task_type: str |
| 1092 | ) -> SpeechToTextConfig: |
| 1093 | processor = cached_processor_from_config(model_config) |
| 1094 | feature_extractor = processor.feature_extractor |
| 1095 | max_audio_clip_s = getattr(processor, "max_audio_len", DEFAULT_MAX_AUDIO_LEN_S) |
| 1096 | return SpeechToTextConfig( |
| 1097 | max_audio_clip_s=max_audio_clip_s, |
| 1098 | sample_rate=feature_extractor.sampling_rate, |
| 1099 | ) |
| 1100 | |
| 1101 | @classmethod |
| 1102 | def get_generation_prompt(cls, stt_params: SpeechToTextParams) -> PromptType: |
| 1103 | """Get the generation prompt to be used for transcription requests.""" |
| 1104 | audio = stt_params.audio |
| 1105 | model_config = stt_params.model_config |
| 1106 | task_type = stt_params.task_type |
| 1107 | to_language = stt_params.to_language |
| 1108 | tokenizer = cached_tokenizer_from_config(model_config) |
| 1109 | audio_token = cls._get_audio_token(model_config) |
| 1110 | |
| 1111 | if task_type == "translate": |
| 1112 | full_lang_name_to = ( |
| 1113 | cls.supported_languages.get(to_language, to_language) |
| 1114 | if to_language is not None |
| 1115 | else "" |
| 1116 | ) |
| 1117 | user_content = f"{audio_token}translate the speech to {full_lang_name_to}" |
| 1118 | elif task_type == "transcribe": |
| 1119 | user_content = ( |
| 1120 | f"{audio_token}can you transcribe the speech into a written format?" |
| 1121 | ) |
| 1122 | else: |
| 1123 | raise ValueError(f"Unsupported task type {task_type}") |
| 1124 | |
| 1125 | messages = [{"role": "user", "content": user_content}] |
| 1126 | prompt = tokenizer.apply_chat_template( |
| 1127 | messages, tokenize=False, add_generation_prompt=True |
| 1128 | ) |
| 1129 | |
| 1130 | prompt_token_ids = tokenizer.encode(prompt) |
| 1131 | |
| 1132 | return TokensPrompt( |
| 1133 | prompt_token_ids=prompt_token_ids, |
| 1134 | multi_modal_data={"audio": audio}, |
| 1135 | ) |