| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Adapted from |
| 5 | # https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py |
| 6 | # Copyright 2025 The vLLM team. |
| 7 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 8 | # |
| 9 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 10 | # and OPT implementations in this library. It has been modified from its |
| 11 | # original forms to accommodate minor architectural differences compared |
| 12 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 13 | # |
| 14 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 15 | # you may not use this file except in compliance with the License. |
| 16 | # You may obtain a copy of the License at |
| 17 | # |
| 18 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 19 | # |
| 20 | # Unless required by applicable law or agreed to in writing, software |
| 21 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 22 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 23 | # See the License for the specific language governing permissions and |
| 24 | # limitations under the License. |
| 25 | """Inference-only IBM Granite speech model.""" |
| 26 | |
| 27 | import math |
| 28 | from collections.abc import Iterable, Mapping |
| 29 | from typing import Annotated |
| 30 | |
| 31 | import torch |
| 32 | import torch.nn.functional as F |
| 33 | from torch import nn |
| 34 | from transformers import BatchFeature, PreTrainedConfig |
| 35 | |
| 36 | from vllm.config import CacheConfig, ModelConfig, SpeechToTextConfig, VllmConfig |
| 37 | from vllm.config.multimodal import MultiModalDummyOptions |
| 38 | from vllm.config.speech_to_text import SpeechToTextParams |
| 39 | from vllm.inputs import MultiModalDataDict, PromptType, TokensPrompt |
| 40 | from vllm.model_executor.layers.linear import ColumnParallelLinear, RowParallelLinear |
| 41 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 42 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 43 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 44 | from vllm.multimodal.inputs import ( |
| 45 | MultiModalFieldConfig, |
| 46 | MultiModalKwargsItems, |
| 47 | ) |
| 48 | from vllm.multimodal.parse import ( |
| 49 | AudioProcessorItems, |
| 50 | MultiModalDataItems, |
| 51 | MultiModalDataParser, |
| 52 | ) |
| 53 | from vllm.multimodal.processing import ( |
| 54 | BaseDummyInputsBuilder, |
| 55 | BaseMultiModalProcessor, |
| 56 | BaseProcessingInfo, |
| 57 | PromptReplacement, |
| 58 | PromptUpdate, |
| 59 | ) |
| 60 | from vllm.sequence import IntermediateTensors |
| 61 | from vllm.tokenizers import cached_tokenizer_from_config |
| 62 | from vllm.transformers_utils.processor import cached_processor_from_config |
| 63 | from vllm.utils.gpu_sync_debug import gpu_sync_allowed |
| 64 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 65 | from vllm.utils.torch_utils import PIN_MEMORY |
| 66 | |
| 67 | from .blip2 import Blip2QFormerModel |
| 68 | from .interfaces import ( |
| 69 | MultiModalEmbeddings, |
| 70 | SupportsLoRA, |
| 71 | SupportsMultiModal, |
| 72 | SupportsPP, |
| 73 | SupportsTranscription, |
| 74 | ) |
| 75 | from .utils import AutoWeightsLoader, init_vllm_registered_model, maybe_prefix |
| 76 | |
| 77 | # NOTE lang support is based on what is written here: |
| 78 | # https://huggingface.co/ibm-granite/granite-speech-3.3-2b |
| 79 | # https://huggingface.co/ibm-granite/granite-4.0-1b-speech |
| 80 | # Though this may vary from model to model, and also many langs |
| 81 | # work pretty well with zero shot. |
| 82 | ISO639_1_SUPPORTED_LANGS = { |
| 83 | "en": "English", |
| 84 | "fr": "French", |
| 85 | "de": "German", |
| 86 | "ja": "Japanese", |
| 87 | "pt": "Portuguese", |
| 88 | "es": "Spanish", |
| 89 | } |
| 90 | |
| 91 | |
| 92 | ### Audio Input |
| 93 | class GraniteSpeechAudioInputs(TensorSchema): |
| 94 | """Audio input features for Granite Speech model. |
| 95 | |
| 96 | Dimensions: |
| 97 | - b: Batch size |
| 98 | - fi: Number of input features from the Mel spectrogram. |
| 99 | - fo: Number of output features, i.e. the embedding size. |
| 100 | - 160: Fixed feature dimension for Mel spectrogram features |
| 101 | """ |
| 102 | |
| 103 | input_features: Annotated[torch.Tensor, TensorShape("b", "fi", 160)] |
| 104 | """Audio input features.""" |
| 105 | |
| 106 | input_features_mask: Annotated[torch.Tensor, TensorShape("b", "fo")] |
| 107 | """Mask for variable length audio features.""" |
| 108 | |
| 109 | audio_embed_sizes: Annotated[list[int], TensorShape("b")] |
| 110 | """List of audio embedding sizes for each item in batch.""" |
| 111 | |
| 112 | |
| 113 | class GraniteSpeechProcessingInfo(BaseProcessingInfo): |
| 114 | def get_data_parser(self): |
| 115 | feature_extractor = self.get_hf_processor().audio_processor |
| 116 | |
| 117 | return MultiModalDataParser( |
| 118 | target_sr=feature_extractor.melspec_kwargs["sample_rate"], |
| 119 | expected_hidden_size=self._get_expected_hidden_size(), |
| 120 | ) |
| 121 | |
| 122 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 123 | return {"audio": 1} |
| 124 | |
| 125 | # There is no limit to the maximum number of audio tokens that can be |
| 126 | # encoded as features; we pick ~5000 as a number that is probably higher |
| 127 | # than we would expect to encounter. The sequence of length |
| 128 | # get_max_audio_len() produces get_max_audio_tokens(). |
| 129 | def get_max_audio_tokens(self): |
| 130 | return 5001 |
| 131 | |
| 132 | def get_max_audio_len(self): |
| 133 | return 8000000 |
| 134 | |
| 135 | |
| 136 | ### Input Processing & Multimodal utils |
| 137 | class GraniteSpeechMultiModalProcessor( |
| 138 | BaseMultiModalProcessor[GraniteSpeechProcessingInfo] |
| 139 | ): |
| 140 | def _get_mm_fields_config( |
| 141 | self, |
| 142 | hf_inputs: BatchFeature, |
| 143 | hf_processor_mm_kwargs: Mapping[str, object], |
| 144 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 145 | return dict( |
| 146 | input_features=MultiModalFieldConfig.batched("audio"), |
| 147 | audio_embed_sizes=MultiModalFieldConfig.batched("audio", keep_on_cpu=True), |
| 148 | ) |
| 149 | |
| 150 | def _get_prompt_updates( |
| 151 | self, |
| 152 | mm_items: MultiModalDataItems, |
| 153 | hf_processor_mm_kwargs: Mapping[str, object], |
| 154 | out_mm_kwargs: MultiModalKwargsItems, |
| 155 | ) -> list[PromptUpdate]: |
| 156 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 157 | tokenizer = self.info.get_tokenizer() |
| 158 | feature_extractor = processor.audio_processor |
| 159 | vocab = tokenizer.get_vocab() |
| 160 | |
| 161 | # Use getattr with default to be compatible with transformers<4.48 |
| 162 | audio_token = getattr(processor, "audio_token", "<|audio|>") |
| 163 | audio_token_id = vocab[audio_token] |
| 164 | |
| 165 | def get_replacement(item_idx: int): |
| 166 | audios = mm_items.get_items("audio", AudioProcessorItems) |
| 167 | audio = audios.get(item_idx) |
| 168 | if audio is None: |
| 169 | raise ValueError(f"Missing audio item {item_idx}") |
| 170 | audio_length = len(audio) if isinstance(audio, list) else audio.shape[-1] |
| 171 | num_projector_features = feature_extractor._get_num_audio_features( |
| 172 | [audio_length] |
| 173 | )[0] |
| 174 | return [audio_token_id] * num_projector_features |
| 175 | |
| 176 | return [ |
| 177 | PromptReplacement( |
| 178 | modality="audio", |
| 179 | target=[audio_token_id], |
| 180 | replacement=get_replacement, |
| 181 | ) |
| 182 | ] |
| 183 | |
| 184 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 185 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 186 | |
| 187 | def _postprocess_hf_mm_data( |
| 188 | self, |
| 189 | mm_data: Mapping[str, object], |
| 190 | hf_processor_mm_kwargs: Mapping[str, object], |
| 191 | processed_data: BatchFeature, |
| 192 | ) -> BatchFeature: |
| 193 | if "audio" in mm_data: |
| 194 | # Calculate the number of audio tokens per entry in the batch; |
| 195 | # This is used to split the batch back out after padding. |
| 196 | audio_token_index = self.info.get_hf_config().audio_token_index |
| 197 | processed_data["audio_embed_sizes"] = ( |
| 198 | processed_data["input_ids"] == audio_token_index |
| 199 | ).sum(-1) |
| 200 | |
| 201 | return processed_data |
| 202 | |
| 203 | |
| 204 | class GraniteSpeechDummyInputsBuilder( |
| 205 | BaseDummyInputsBuilder[GraniteSpeechProcessingInfo] |
| 206 | ): |
| 207 | def get_dummy_mm_data( |
| 208 | self, |
| 209 | seq_len: int, |
| 210 | mm_counts: Mapping[str, int], |
| 211 | mm_options: MultiModalDummyOptions, |
| 212 | ) -> MultiModalDataDict: |
| 213 | return { |
| 214 | "audio": self._get_dummy_audios( |
| 215 | length=self.info.get_max_audio_len(), |
| 216 | num_audios=mm_counts.get("audio", 0), |
| 217 | overrides=mm_options.get("audio"), |
| 218 | ) |
| 219 | } |
| 220 | |
| 221 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 222 | num_audios = mm_counts.get("audio", 0) |
| 223 | hf_processor = self.info.get_hf_processor() |
| 224 | audio_token = getattr(hf_processor, "audio_token", "<|audio|>") |
| 225 | return audio_token * num_audios |
| 226 | |
| 227 | |
| 228 | ### QFormer Projector |
| 229 | class GraniteSpeechEncoderProjector(nn.Module): |
| 230 | def __init__( |
| 231 | self, |
| 232 | config: PreTrainedConfig, |
| 233 | cache_config: CacheConfig, |
| 234 | quant_config: QuantizationConfig | None = None, |
| 235 | prefix: str = "", |
| 236 | ): |
| 237 | super().__init__() |
| 238 | self.hidden_size = config.projector_config.hidden_size |
| 239 | self.downsample_rate = config.downsample_rate |
| 240 | self.window_size = config.window_size |
| 241 | self.num_queries = config.window_size // config.downsample_rate |
| 242 | |
| 243 | self.query = nn.Parameter( |
| 244 | torch.zeros(1, self.num_queries, config.projector_config.hidden_size) |
| 245 | ) |
| 246 | |
| 247 | # NOTE - this is implemented generically in transformers, |
| 248 | # but for now we create the QFormer model directly since |
| 249 | # all existing models use this for the projector. |
| 250 | self.qformer = Blip2QFormerModel( |
| 251 | config.projector_config, |
| 252 | quant_config=quant_config, |
| 253 | cache_config=cache_config, |
| 254 | prefix=f"{prefix}.qformer", |
| 255 | ) |
| 256 | self.linear = nn.Linear( |
| 257 | config.projector_config.hidden_size, config.text_config.hidden_size |
| 258 | ) |
| 259 | |
| 260 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 261 | batch_size, seq_len, dim = hidden_states.size() |
| 262 | nblocks = math.ceil(seq_len / self.window_size) |
| 263 | pad = nblocks * self.window_size - seq_len |
| 264 | hidden_states = nn.functional.pad(hidden_states, (0, 0, 0, pad), "constant", 0) |
| 265 | hidden_states = hidden_states.view(batch_size * nblocks, self.window_size, dim) |
| 266 | |
| 267 | last_hidden_state = self.qformer( |
| 268 | query_embeds=self.query.data, |
| 269 | encoder_hidden_states=hidden_states, |
| 270 | ) |
| 271 | |
| 272 | query_proj = self.linear( |
| 273 | last_hidden_state.view( |
| 274 | batch_size, |
| 275 | nblocks * self.window_size // self.downsample_rate, |
| 276 | -1, |
| 277 | ) |
| 278 | ) |
| 279 | return query_proj |
| 280 | |
| 281 | |
| 282 | # Encoder - conformer is adapted from: https://github.com/lucidrains/conformer.git |
| 283 | # NOTE - it would be nice to see if we can align this with other models using |
| 284 | # conformer in vLLM, e.g., phi4mm audio. |
| 285 | class GraniteSpeechConformerFeedForward(nn.Module): |
| 286 | """Feedforward module for conformer encoder blocks.""" |
| 287 | |
| 288 | def __init__( |
| 289 | self, |
| 290 | config: PreTrainedConfig, |
| 291 | quant_config: QuantizationConfig | None = None, |
| 292 | prefix: str = "", |
| 293 | ): |
| 294 | super().__init__() |
| 295 | self.pre_norm = nn.LayerNorm(config.hidden_dim) |
| 296 | |
| 297 | self.up_proj = ColumnParallelLinear( |
| 298 | input_size=config.hidden_dim, |
| 299 | output_size=config.hidden_dim * config.feedforward_mult, |
| 300 | quant_config=quant_config, |
| 301 | prefix=f"{prefix}.up_proj", |
| 302 | ) |
| 303 | self.silu = nn.SiLU() |
| 304 | |
| 305 | self.down_proj = RowParallelLinear( |
| 306 | input_size=config.hidden_dim * config.feedforward_mult, |
| 307 | output_size=config.hidden_dim, |
| 308 | quant_config=quant_config, |
| 309 | prefix=f"{prefix}.down_proj", |
| 310 | ) |
| 311 | |
| 312 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 313 | hidden_states = self.pre_norm(hidden_states) |
| 314 | hidden_states, _ = self.up_proj(hidden_states) |
| 315 | hidden_states = self.silu(hidden_states) |
| 316 | hidden_states, _ = self.down_proj(hidden_states) |
| 317 | return hidden_states |
| 318 | |
| 319 | |
| 320 | class GraniteSpeechConformerAttention(nn.Module): |
| 321 | """Attention for conformer blocks using Shaw's relative positional |
| 322 | embeddings. See the following [paper](https://arxiv.org/pdf/1803.02155) |
| 323 | for more details. |
| 324 | """ |
| 325 | |
| 326 | def __init__(self, config: PreTrainedConfig, prefix: str = ""): |
| 327 | super().__init__() |
| 328 | |
| 329 | inner_dim = config.dim_head * config.num_heads |
| 330 | self.max_pos_emb = config.max_pos_emb |
| 331 | self.context_size = config.context_size |
| 332 | self.num_heads = config.num_heads |
| 333 | self.dim_head = config.dim_head |
| 334 | self.scale = self.dim_head**-0.5 |
| 335 | self.pre_norm = nn.LayerNorm(config.hidden_dim) |
| 336 | self.to_q = nn.Linear(config.hidden_dim, inner_dim, bias=False) |
| 337 | self.to_kv = nn.Linear(config.hidden_dim, inner_dim * 2, bias=False) |
| 338 | self.to_out = nn.Linear(inner_dim, config.hidden_dim) |
| 339 | self.rel_pos_emb = nn.Embedding(2 * self.max_pos_emb + 1, self.dim_head) |
| 340 | |
| 341 | if self.context_size <= 0 or self.context_size > self.max_pos_emb: |
| 342 | raise ValueError( |
| 343 | f"Context size should be > 0 and " |
| 344 | f"<= max_pos_emb ({self.max_pos_emb}), " |
| 345 | f"got {self.context_size}." |
| 346 | ) |
| 347 | |
| 348 | def forward( |
| 349 | self, hidden_states: torch.Tensor, attention_dists: torch.Tensor |
| 350 | ) -> torch.Tensor: |
| 351 | hidden_states = self.pre_norm(hidden_states) |
| 352 | bsz, num_features, _ = hidden_states.shape |
| 353 | |
| 354 | num_blocks = math.ceil(num_features / self.context_size) |
| 355 | remainder = num_features % self.context_size |
| 356 | if remainder > 0: |
| 357 | # right padding to reach block size |
| 358 | hidden_states = torch.nn.functional.pad( |
| 359 | hidden_states, (0, 0, 0, self.context_size - remainder) |
| 360 | ) |
| 361 | |
| 362 | # NOTE: would be nice to try to use qkvparallellinear |
| 363 | # here for this block attention implementation if possible |
| 364 | query_states = self.to_q(hidden_states) |
| 365 | key_states, value_states = self.to_kv(hidden_states).chunk(2, dim=-1) |
| 366 | |
| 367 | query_states = query_states.reshape( |
| 368 | bsz, num_blocks, self.context_size, self.num_heads, -1 |
| 369 | ).transpose(2, 3) |
| 370 | key_states = key_states.reshape( |
| 371 | bsz, num_blocks, self.context_size, self.num_heads, -1 |
| 372 | ).transpose(2, 3) |
| 373 | value_states = value_states.reshape( |
| 374 | bsz, num_blocks, self.context_size, self.num_heads, -1 |
| 375 | ).transpose(2, 3) |
| 376 | |
| 377 | # shaw's relative positional embedding |
| 378 | dist = attention_dists.to(hidden_states.device) |
| 379 | rel_pos_emb = self.rel_pos_emb(dist) |
| 380 | pos_attn = ( |
| 381 | torch.einsum("bnhid,ijd->bnhij", query_states, rel_pos_emb) * self.scale |
| 382 | ) |
| 383 | |
| 384 | if remainder > 0: |
| 385 | # masked attention in the extended block |
| 386 | mask = torch.ones( |
| 387 | self.context_size, |
| 388 | self.context_size, |
| 389 | dtype=bool, |
| 390 | device=hidden_states.device, |
| 391 | ) |
| 392 | mask[:remainder, :remainder] = 0 |
| 393 | mask_value = -torch.finfo(pos_attn.dtype).max |
| 394 | pos_attn[:, -1, :].masked_fill_(mask, mask_value) |
| 395 | |
| 396 | with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH): |
| 397 | out = F.scaled_dot_product_attention( |
| 398 | query_states, |
| 399 | key_states, |
| 400 | value_states, |
| 401 | attn_mask=pos_attn, |
| 402 | scale=self.scale, |
| 403 | ) |
| 404 | out = out.transpose(2, 3).reshape(bsz, hidden_states.shape[1], -1) |
| 405 | return self.to_out(out[:, :num_features, :]) |
| 406 | |
| 407 | |
| 408 | class GraniteSpeechConformerDepthWiseConv1d(nn.Module): |
| 409 | """Wrapper for padded 1D pointwise convolution.""" |
| 410 | |
| 411 | def __init__(self, chan_in: int, chan_out: int, kernel_size: int, prefix: str = ""): |
| 412 | super().__init__() |
| 413 | # Padding for the 1D conv is symmetric or close (i.e., offset by one). |
| 414 | pad = kernel_size // 2 |
| 415 | pad_offset = (kernel_size + 1) % 2 |
| 416 | self.padding = (pad, pad - pad_offset) |
| 417 | |
| 418 | self.conv = nn.Conv1d( |
| 419 | chan_in, chan_out, kernel_size, groups=chan_in, bias=False |
| 420 | ) |
| 421 | |
| 422 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 423 | hidden_states = F.pad(hidden_states, self.padding) |
| 424 | return self.conv(hidden_states) |
| 425 | |
| 426 | |
| 427 | class GraniteSpeechConformerConvModule(nn.Module): |
| 428 | """Conformer conv module consisting of several 1D/depthwise 1D |
| 429 | convolutional layers. |
| 430 | """ |
| 431 | |
| 432 | def __init__(self, config: PreTrainedConfig, prefix: str = ""): |
| 433 | super().__init__() |
| 434 | inner_dim = config.hidden_dim * config.conv_expansion_factor |
| 435 | |
| 436 | self.norm = nn.LayerNorm(config.hidden_dim) |
| 437 | self.up_conv = nn.Conv1d(config.hidden_dim, inner_dim * 2, 1) |
| 438 | self.glu = nn.GLU(dim=1) |
| 439 | self.depth_conv = GraniteSpeechConformerDepthWiseConv1d( |
| 440 | inner_dim, |
| 441 | inner_dim, |
| 442 | kernel_size=config.conv_kernel_size, |
| 443 | prefix=f"{prefix}.depth_conv", |
| 444 | ) |
| 445 | self.silu = nn.SiLU() |
| 446 | self.batch_norm = nn.BatchNorm1d(inner_dim) |
| 447 | self.down_conv = nn.Conv1d(inner_dim, config.hidden_dim, 1) |
| 448 | |
| 449 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 450 | hidden_states = self.norm(hidden_states) |
| 451 | hidden_states = self.up_conv(hidden_states.permute(0, 2, 1)) |
| 452 | hidden_states = self.glu(hidden_states) |
| 453 | hidden_states = self.depth_conv(hidden_states) |
| 454 | hidden_states = self.silu(self.batch_norm(hidden_states)) |
| 455 | hidden_states = self.down_conv(hidden_states).permute(0, 2, 1) |
| 456 | return hidden_states |
| 457 | |
| 458 | |
| 459 | class GraniteSpeechConformerBlock(nn.Module): |
| 460 | """Conformer block, consisting largely of linear layers, |
| 461 | attention, and convolutional layers.""" |
| 462 | |
| 463 | def __init__(self, config: PreTrainedConfig, prefix: str = ""): |
| 464 | super().__init__() |
| 465 | self.ff1 = GraniteSpeechConformerFeedForward(config, prefix=f"{prefix}.ff1") |
| 466 | self.attn = GraniteSpeechConformerAttention(config, prefix=f"{prefix}.attn") |
| 467 | self.conv = GraniteSpeechConformerConvModule(config, prefix=f"{prefix}.conv") |
| 468 | self.ff2 = GraniteSpeechConformerFeedForward(config, prefix=f"{prefix}.ff2") |
| 469 | self.post_norm = nn.LayerNorm(config.hidden_dim) |
| 470 | |
| 471 | def forward( |
| 472 | self, hidden_states: torch.Tensor, attention_dists: torch.Tensor |
| 473 | ) -> torch.Tensor: |
| 474 | hidden_states = 0.5 * self.ff1(hidden_states) + hidden_states |
| 475 | hidden_states = ( |
| 476 | self.attn(hidden_states, attention_dists=attention_dists) + hidden_states |
| 477 | ) |
| 478 | hidden_states = self.conv(hidden_states) + hidden_states |
| 479 | hidden_states = 0.5 * self.ff2(hidden_states) + hidden_states |
| 480 | hidden_states = self.post_norm(hidden_states) |
| 481 | return hidden_states |
| 482 | |
| 483 | |
| 484 | class GraniteSpeechCTCEncoder(nn.Module): |
| 485 | """CTC Encoder comprising conformer blocks and additional linear layers.""" |
| 486 | |
| 487 | def __init__( |
| 488 | self, |
| 489 | config: PreTrainedConfig, |
| 490 | prefix: str, |
| 491 | quant_config: QuantizationConfig | None = None, |
| 492 | ): |
| 493 | super().__init__() |
| 494 | self.config = config |
| 495 | |
| 496 | # Precompute clamped relative positional encoding distances |
| 497 | seq = torch.arange(config.context_size) |
| 498 | relpos_dist = seq.view(-1, 1) - seq.view(1, -1) |
| 499 | self.attention_dists = ( |
| 500 | torch.clamp(relpos_dist, -config.context_size, config.context_size) |
| 501 | + config.max_pos_emb |
| 502 | ) |
| 503 | |
| 504 | self.input_linear = nn.Linear(config.input_dim, config.hidden_dim, bias=True) |
| 505 | self.layers = nn.ModuleList( |
| 506 | [ |
| 507 | GraniteSpeechConformerBlock( |
| 508 | config, |
| 509 | prefix=f"{prefix}.layers.{idx}", |
| 510 | ) |
| 511 | for idx in range(config.num_layers) |
| 512 | ] |
| 513 | ) |
| 514 | |
| 515 | self.out = ColumnParallelLinear( |
| 516 | input_size=config.hidden_dim, |
| 517 | output_size=config.output_dim, |
| 518 | bias=True, |
| 519 | quant_config=quant_config, |
| 520 | prefix=f"{prefix}.out", |
| 521 | ) |
| 522 | |
| 523 | self.out_mid = RowParallelLinear( |
| 524 | input_size=config.output_dim, |
| 525 | output_size=config.hidden_dim, |
| 526 | bias=True, |
| 527 | quant_config=quant_config, |
| 528 | prefix=f"{prefix}.out_mid", |
| 529 | ) |
| 530 | self.softmax = nn.Softmax(dim=-1) |
| 531 | self.num_layers = config.num_layers |
| 532 | |
| 533 | def forward(self, hidden_states: torch.Tensor): |
| 534 | hidden_states = self.input_linear(hidden_states) |
| 535 | for idx, layer in enumerate(self.layers, start=1): |
| 536 | hidden_states = layer(hidden_states, attention_dists=self.attention_dists) |
| 537 | |
| 538 | if idx == self.num_layers // 2: |
| 539 | hidden_states_mid = hidden_states.clone() |
| 540 | hidden_states_mid, _ = self.out(hidden_states_mid) |
| 541 | hidden_states_mid = self.softmax(hidden_states_mid) |
| 542 | hidden_states_mid, _ = self.out_mid(hidden_states_mid) |
| 543 | hidden_states += hidden_states_mid |
| 544 | return hidden_states |
| 545 | |
| 546 | |
| 547 | @MULTIMODAL_REGISTRY.register_processor( |
| 548 | GraniteSpeechMultiModalProcessor, |
| 549 | info=GraniteSpeechProcessingInfo, |
| 550 | dummy_inputs=GraniteSpeechDummyInputsBuilder, |
| 551 | ) |
| 552 | class GraniteSpeechForConditionalGeneration( |
| 553 | nn.Module, |
| 554 | SupportsMultiModal, |
| 555 | SupportsPP, |
| 556 | SupportsLoRA, |
| 557 | SupportsTranscription, |
| 558 | ): |
| 559 | supported_languages = ISO639_1_SUPPORTED_LANGS |
| 560 | |
| 561 | packed_modules_mapping = { |
| 562 | "qkv_proj": [ |
| 563 | "q_proj", |
| 564 | "k_proj", |
| 565 | "v_proj", |
| 566 | ], |
| 567 | "gate_up_proj": [ |
| 568 | "gate_proj", |
| 569 | "up_proj", |
| 570 | ], |
| 571 | } |
| 572 | |
| 573 | @classmethod |
| 574 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 575 | if modality.startswith("audio"): |
| 576 | return "<|audio|>" |
| 577 | |
| 578 | raise ValueError("Only audio modality is supported") |
| 579 | |
| 580 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 581 | super().__init__() |
| 582 | config = vllm_config.model_config.hf_config |
| 583 | quant_config = vllm_config.quant_config |
| 584 | cache_config = vllm_config.cache_config |
| 585 | |
| 586 | self.config = config |
| 587 | self.quant_config = quant_config |
| 588 | self.cache_config = cache_config |
| 589 | |
| 590 | # Check for OOV tokens to see if offsets need to be preserved |
| 591 | self.configure_mm_token_handling( |
| 592 | vocab_size=config.text_config.vocab_size, |
| 593 | mm_token_ids=[config.audio_token_index], |
| 594 | ) |
| 595 | |
| 596 | with self._mark_language_model(vllm_config): |
| 597 | # The language model is typically a Granite LLM |
| 598 | self.language_model = init_vllm_registered_model( |
| 599 | vllm_config=vllm_config, |
| 600 | hf_config=config.text_config, |
| 601 | prefix=maybe_prefix(prefix, "language_model"), |
| 602 | ) |
| 603 | |
| 604 | with self._mark_tower_model(vllm_config, "audio"): |
| 605 | self.encoder = self._build_encoder( |
| 606 | config=config.encoder_config, |
| 607 | quant_config=quant_config, |
| 608 | prefix=maybe_prefix(prefix, "encoder"), |
| 609 | ) |
| 610 | |
| 611 | # Blip2 QFormer |
| 612 | self.projector = GraniteSpeechEncoderProjector( |
| 613 | config=config, |
| 614 | quant_config=quant_config, |
| 615 | cache_config=cache_config, |
| 616 | prefix=maybe_prefix(prefix, "projector"), |
| 617 | ) |
| 618 | |
| 619 | self.make_empty_intermediate_tensors = ( |
| 620 | self.language_model.make_empty_intermediate_tensors |
| 621 | ) |
| 622 | |
| 623 | def _build_encoder( |
| 624 | self, |
| 625 | config: PreTrainedConfig, |
| 626 | quant_config: QuantizationConfig | None, |
| 627 | prefix: str, |
| 628 | ) -> "GraniteSpeechCTCEncoder": |
| 629 | return GraniteSpeechCTCEncoder( |
| 630 | config=config, |
| 631 | quant_config=quant_config, |
| 632 | prefix=prefix, |
| 633 | ) |
| 634 | |
| 635 | def _parse_and_validate_audio_input( |
| 636 | self, |
| 637 | **kwargs: object, |
| 638 | ) -> GraniteSpeechAudioInputs | None: |
| 639 | input_features = kwargs.pop("input_features", None) |
| 640 | input_features_mask = kwargs.pop("input_features_mask", None) |
| 641 | audio_embed_sizes = kwargs.pop("audio_embed_sizes", None) |
| 642 | |
| 643 | if input_features is None: |
| 644 | return None |
| 645 | |
| 646 | if not isinstance(audio_embed_sizes, torch.Tensor): |
| 647 | raise ValueError("audio_embed_sizes must be a tensor") |
| 648 | |
| 649 | # If we have a batch of variable feature length audio clips, we need |
| 650 | # to mask the features; usually we would get an input_features_mask |
| 651 | # from the processor, but we handle rebuilding it here since |
| 652 | # vLLM generally processes everything independently + batches. |
| 653 | if input_features_mask is None: |
| 654 | input_features_mask = self._build_input_features_mask(audio_embed_sizes) |
| 655 | |
| 656 | if not isinstance(input_features, (torch.Tensor, list)): |
| 657 | raise ValueError( |
| 658 | "Incorrect type of audio input features. " |
| 659 | f"Got type: {type(input_features)}" |
| 660 | ) |
| 661 | |
| 662 | if input_features_mask is not None and not isinstance( |
| 663 | input_features_mask, torch.Tensor |
| 664 | ): |
| 665 | raise ValueError( |
| 666 | "Incorrect type of audio input features mask. " |
| 667 | f"Got type: {type(input_features_mask)}" |
| 668 | ) |
| 669 | |
| 670 | if isinstance(input_features, torch.Tensor): |
| 671 | # Granite speech currently only allows one audio token per instance |
| 672 | # and features are already unsqueezed in the processor, so one |
| 673 | # instance will have shape [1, {num_features}, 160]. As such, |
| 674 | # input features will usually be of shape |
| 675 | # [bsz, 1, num_features, 160], which we squeeze to be 3D here. |
| 676 | if len(input_features.shape) == 4: |
| 677 | input_features = input_features.squeeze(1) |
| 678 | if len(input_features.shape) != 3: |
| 679 | raise ValueError( |
| 680 | "Squeezed input features should be 3D but are of shape " |
| 681 | f"{input_features.shape}" |
| 682 | ) |
| 683 | input_features = input_features.to(self.encoder.input_linear.weight.dtype) |
| 684 | |
| 685 | else: |
| 686 | # Otherwise we have a list of tensors, which are almost certainly |
| 687 | # differing in their respective numbers of audio features; when |
| 688 | # passed as a batch, we expect a list of 2D var len input features |
| 689 | # so unsqueeze them. |
| 690 | input_features = [ |
| 691 | feat.unsqueeze(dim=0) for feat in input_features if feat.ndim == 2 |
| 692 | ] |
| 693 | |
| 694 | # stack them into a 3D tensor of size [bsz, most_num_features, 160]. |
| 695 | input_features = self._pad_and_stack_input_features( |
| 696 | input_features, |
| 697 | ).to(self.encoder.input_linear.weight.dtype) |
| 698 | |
| 699 | return GraniteSpeechAudioInputs( |
| 700 | input_features=input_features, |
| 701 | input_features_mask=input_features_mask, |
| 702 | audio_embed_sizes=audio_embed_sizes.flatten().tolist(), |
| 703 | ) |
| 704 | |
| 705 | def _build_input_features_mask( |
| 706 | self, |
| 707 | audio_embed_sizes: torch.Tensor, |
| 708 | ) -> torch.Tensor: |
| 709 | """Calculate the input features mask, which will generally be used |
| 710 | to mask the padded features for all entries in the batch except |
| 711 | for those with the most audio features. |
| 712 | |
| 713 | Args: |
| 714 | audio_embed_sizes: torch.Tensor |
| 715 | Tensor of num features in each seq in the batch. |
| 716 | |
| 717 | Returns: |
| 718 | torch.Tensor: Mask of shape (bsz, num_features) to be applied to |
| 719 | the audio features prior to splitting the audio embeddings. |
| 720 | |
| 721 | """ |
| 722 | most_audio_features = int(torch.max(audio_embed_sizes)) |
| 723 | mask_indices = torch.arange(most_audio_features).view(1, -1) |
| 724 | input_features_mask = mask_indices < audio_embed_sizes.view(-1, 1) |
| 725 | target_device = self.encoder.input_linear.weight.device |
| 726 | if target_device == input_features_mask.device: |
| 727 | return input_features_mask |
| 728 | masked = input_features_mask.pin_memory() if PIN_MEMORY else input_features_mask |
| 729 | return masked.to(target_device, non_blocking=True) |
| 730 | |
| 731 | def _pad_and_stack_input_features( |
| 732 | self, |
| 733 | input_features: list[torch.Tensor], |
| 734 | ) -> torch.Tensor: |
| 735 | """Given a list of input features of varying length, pad them to the |
| 736 | same length and stack them into a torch.Tensor. |
| 737 | |
| 738 | NOTE: Usually, padding is done in the input processor/feature extractor |
| 739 | and zero padded prior to the computation of the Mel features; the |
| 740 | resulting values are only constant within a batch and generally nonzero |
| 741 | (i.e., slightly negative nums); we should validate that this is okay |
| 742 | since we don't use a feature attention mask, but the more important |
| 743 | thing is that we apply the input_features_mask with variable len |
| 744 | batches. |
| 745 | |
| 746 | Args: |
| 747 | input_features: list[torch.Tensor] |
| 748 | 3D Input features to be coerced into a tensor. |
| 749 | |
| 750 | Returns: |
| 751 | torch.Tensor: Tensor of shape [bsz, num_features, 160], where |
| 752 | num_features is the max number of features of any entry in the |
| 753 | batch. |
| 754 | |
| 755 | """ |
| 756 | feat_lens = [feats.shape[1] for feats in input_features] |
| 757 | padding = [max(feat_lens) - length for length in feat_lens] |
| 758 | # TODO (Alex) - Validate that it's okay to zero pad like this; |
| 759 | # in transformers we zero pad prior to calculating the speech features, |
| 760 | # so the value is not zero and is dependent on the batched features. |
| 761 | padded = [ |
| 762 | torch.nn.functional.pad(feats, (0, 0, 0, pad, 0, 0)) |
| 763 | for feats, pad in zip(input_features, padding) |
| 764 | ] |
| 765 | stacked_features = torch.cat(padded, dim=0).to(input_features[0]) |
| 766 | return stacked_features |
| 767 | |
| 768 | def _process_audio_input( |
| 769 | self, |
| 770 | audio_input: GraniteSpeechAudioInputs, |
| 771 | ) -> tuple[torch.Tensor]: |
| 772 | """Compute the audio features to be merged into the LLM embeddings. |
| 773 | |
| 774 | Args: |
| 775 | audio_input: GraniteSpeechAudioInputs |
| 776 | Audio inputs object containing Mel features, an input features |
| 777 | mask, and the (flattened) number of audio tokens per instance. |
| 778 | |
| 779 | Returns: |
| 780 | tuple[torch.Tensor]: List of length bsz. |
| 781 | |
| 782 | """ |
| 783 | # TODO (Alex) - support embedding inputs |
| 784 | encoder_embeds = self.encoder(audio_input["input_features"]) |
| 785 | # [bsz, <max feature size>, 4096] |
| 786 | projected_embeds = self.projector(encoder_embeds) |
| 787 | # Apply mask on variable length audio features. |
| 788 | with gpu_sync_allowed(): |
| 789 | masked_embeds = projected_embeds[audio_input["input_features_mask"]] |
| 790 | # Split variable length features into a tuple |
| 791 | return torch.split(masked_embeds, audio_input["audio_embed_sizes"]) |
| 792 | |
| 793 | def embed_multimodal( |
| 794 | self, |
| 795 | **kwargs: object, |
| 796 | ) -> MultiModalEmbeddings: |
| 797 | """Compute the audio embeddings if audio inputs are present.""" |
| 798 | audio_input = self._parse_and_validate_audio_input(**kwargs) |
| 799 | if audio_input is None: |
| 800 | return [] |
| 801 | |
| 802 | audio_features = self._process_audio_input(audio_input) |
| 803 | return audio_features |
| 804 | |
| 805 | def embed_input_ids( |
| 806 | self, |
| 807 | input_ids: torch.Tensor, |
| 808 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 809 | *, |
| 810 | is_multimodal: torch.Tensor | None = None, |
| 811 | ) -> torch.Tensor: |
| 812 | # This is to satisfy the type checker for each overload |
| 813 | if multimodal_embeddings is None or is_multimodal is None: |
| 814 | return super().embed_input_ids(input_ids) |
| 815 | |
| 816 | return super().embed_input_ids( |
| 817 | input_ids, |
| 818 | multimodal_embeddings=multimodal_embeddings, |
| 819 | is_multimodal=is_multimodal, |
| 820 | ) |
| 821 | |
| 822 | def forward( |
| 823 | self, |
| 824 | input_ids: torch.Tensor | None, |
| 825 | positions: torch.Tensor, |
| 826 | intermediate_tensors: IntermediateTensors | None = None, |
| 827 | inputs_embeds: torch.Tensor | None = None, |
| 828 | **kwargs: object, |
| 829 | ) -> torch.Tensor | IntermediateTensors: |
| 830 | if intermediate_tensors is not None: |
| 831 | inputs_embeds = None |
| 832 | |
| 833 | model_output = self.language_model( |
| 834 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 835 | ) |
| 836 | return model_output |
| 837 | |
| 838 | def compute_logits( |
| 839 | self, |
| 840 | hidden_states: torch.Tensor, |
| 841 | ) -> torch.Tensor | None: |
| 842 | return self.language_model.compute_logits(hidden_states) |
| 843 | |
| 844 | def load_weights( |
| 845 | self, |
| 846 | weights: Iterable[tuple[str, torch.Tensor]], |
| 847 | ) -> set[str]: |
| 848 | loader = AutoWeightsLoader(self) |
| 849 | return loader.load_weights(weights) |
| 850 | |
| 851 | def get_mm_mapping(self) -> MultiModelKeys: |
| 852 | """Get the module prefix in multimodal models.""" |
| 853 | return MultiModelKeys.from_string_field( |
| 854 | language_model="language_model", |
| 855 | connector="projector", |
| 856 | tower_model="encoder", |
| 857 | ) |
| 858 | |
| 859 | ### Support for speech-to-text Transcription |
| 860 | @classmethod |
| 861 | def get_generation_prompt( |
| 862 | cls, |
| 863 | stt_params: SpeechToTextParams, |
| 864 | ) -> PromptType: |
| 865 | """Get the generation prompt to be used for transcription requests.""" |
| 866 | audio = stt_params.audio |
| 867 | model_config = stt_params.model_config |
| 868 | task_type = stt_params.task_type |
| 869 | to_language = stt_params.to_language |
| 870 | |
| 871 | # Audio placeholders don't use an index, so value doesn't matter |
| 872 | audio_tok = cls.get_placeholder_str("audio", 0) |
| 873 | |
| 874 | if task_type == "translate": |
| 875 | full_lang_name_to = ( |
| 876 | cls.supported_languages.get(to_language, to_language) |
| 877 | if to_language is not None |
| 878 | else "" |
| 879 | ) |
| 880 | user_prompt = f"{audio_tok}translate the speech to {full_lang_name_to}" # noqa: E501 |
| 881 | elif task_type == "transcribe": |
| 882 | user_prompt = ( |
| 883 | f"{audio_tok}can you transcribe the speech into a written format?" # noqa: E501 |
| 884 | ) |
| 885 | else: |
| 886 | raise ValueError(f"Unsupported task type {task_type}") |
| 887 | |
| 888 | tokenizer = cached_tokenizer_from_config(model_config) |
| 889 | chat = [dict(role="user", content=user_prompt)] |
| 890 | prompt = tokenizer.apply_chat_template( |
| 891 | chat, |
| 892 | tokenize=False, |
| 893 | add_generation_prompt=True, |
| 894 | ) |
| 895 | |
| 896 | prompt_token_ids = tokenizer.encode(prompt) |
| 897 | |
| 898 | return TokensPrompt( |
| 899 | prompt_token_ids=prompt_token_ids, |
| 900 | multi_modal_data={"audio": audio}, |
| 901 | ) |
| 902 | |
| 903 | # Adapted from https://github.com/huggingface/transformers/blob/v4.56.0/src/transformers/models/granite_speech/feature_extraction_granite_speech.py#L122 # noqa: E501 |
| 904 | @classmethod |
| 905 | def get_num_audio_tokens( |
| 906 | cls, |
| 907 | audio_duration_s: float, |
| 908 | stt_config: SpeechToTextConfig, |
| 909 | model_config: ModelConfig, |
| 910 | ) -> int | None: |
| 911 | """Get the number of audio tokens for an audio duration in sec.""" |
| 912 | processor = cached_processor_from_config(model_config) |
| 913 | hop_length = processor.audio_processor.melspec_kwargs["hop_length"] |
| 914 | proj_win_size = processor.audio_processor.projector_window_size |
| 915 | ds_rate = processor.audio_processor.projector_downsample_rate |
| 916 | effective_window_size = proj_win_size // ds_rate |
| 917 | |
| 918 | raw_length = audio_duration_s * stt_config.sample_rate |
| 919 | |
| 920 | # mel sequence length computation |
| 921 | mel_length = raw_length // hop_length + 1 |
| 922 | # encoder frame takes two mel features |
| 923 | encoder_length = mel_length // 2 |
| 924 | nblocks = math.ceil(encoder_length / proj_win_size) |
| 925 | # projector output length |
| 926 | return nblocks * effective_window_size |
| 927 | |
| 928 | @classmethod |
| 929 | def get_speech_to_text_config( |
| 930 | cls, model_config: ModelConfig, task_type: str |
| 931 | ) -> SpeechToTextConfig: |
| 932 | """Get the stt config for this model.""" |
| 933 | # Default settings are reasonable for this model and we don't currently |
| 934 | # expose this information in the model configs, but this may change in |
| 935 | # the future |
| 936 | return SpeechToTextConfig() |