| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Copyright 2025 The Qwen team. |
| 4 | # Copyright 2023 The vLLM team. |
| 5 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 6 | # |
| 7 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 8 | # and OPT implementations in this library. It has been modified from its |
| 9 | # original forms to accommodate minor architectural differences compared |
| 10 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 11 | # |
| 12 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 13 | # you may not use this file except in compliance with the License. |
| 14 | # You may obtain a copy of the License at |
| 15 | # |
| 16 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 17 | # |
| 18 | # Unless required by applicable law or agreed to in writing, software |
| 19 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 20 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 21 | # See the License for the specific language governing permissions and |
| 22 | # limitations under the License. |
| 23 | """Inference-only Qwen3-Omni-Moe model (thinker part).""" |
| 24 | |
| 25 | from collections.abc import Callable, Iterable, Iterator, Mapping, Sequence |
| 26 | from functools import partial |
| 27 | from typing import Any, ClassVar |
| 28 | |
| 29 | import numpy as np |
| 30 | import torch |
| 31 | import torch.nn as nn |
| 32 | import torch.nn.functional as F |
| 33 | from transformers import PreTrainedConfig |
| 34 | from transformers.feature_extraction_utils import BatchFeature |
| 35 | from transformers.models.qwen3_omni_moe.configuration_qwen3_omni_moe import ( |
| 36 | Qwen3OmniMoeAudioEncoderConfig, |
| 37 | Qwen3OmniMoeConfig, |
| 38 | Qwen3OmniMoeThinkerConfig, |
| 39 | ) |
| 40 | from transformers.models.qwen3_omni_moe.processing_qwen3_omni_moe import ( |
| 41 | Qwen3OmniMoeProcessor, |
| 42 | ) |
| 43 | from transformers.models.whisper import WhisperFeatureExtractor |
| 44 | from typing_extensions import TypeVar |
| 45 | |
| 46 | from vllm.compilation.decorators import support_torch_compile |
| 47 | from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig |
| 48 | from vllm.config.speech_to_text import SpeechToTextParams |
| 49 | from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size |
| 50 | from vllm.inputs import PromptType, TextPrompt |
| 51 | from vllm.logger import init_logger |
| 52 | from vllm.model_executor.layers.activation import _ACTIVATION_REGISTRY |
| 53 | from vllm.model_executor.layers.attention.mm_encoder_attention import ( |
| 54 | MMEncoderAttention, |
| 55 | ) |
| 56 | from vllm.model_executor.layers.conv import Conv3dLayer |
| 57 | from vllm.model_executor.layers.linear import ( |
| 58 | ColumnParallelLinear, |
| 59 | QKVParallelLinear, |
| 60 | ReplicatedLinear, |
| 61 | RowParallelLinear, |
| 62 | ) |
| 63 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 64 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 65 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 66 | from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 67 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 68 | from vllm.model_executor.models.qwen2_audio import Qwen2AudioProcessingInfo |
| 69 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 70 | from vllm.multimodal.inputs import ( |
| 71 | MultiModalFeatureSpec, |
| 72 | MultiModalKwargsItem, |
| 73 | MultiModalKwargsItems, |
| 74 | ) |
| 75 | from vllm.multimodal.parse import AudioProcessorItems, MultiModalDataItems |
| 76 | from vllm.multimodal.processing.processor import ( |
| 77 | BaseProcessingInfo, |
| 78 | HFMultiModalInputs, |
| 79 | MultiModalProcessingResult, |
| 80 | MultiModalPromptUpdates, |
| 81 | PlaceholderFeaturesInfo, |
| 82 | PromptReplacement, |
| 83 | PromptUpdate, |
| 84 | PromptUpdateDetails, |
| 85 | ) |
| 86 | from vllm.sequence import IntermediateTensors |
| 87 | from vllm.transformers_utils.processor import cached_processor_from_config |
| 88 | from vllm.utils.gpu_sync_debug import gpu_sync_allowed |
| 89 | from vllm.utils.torch_utils import async_tensor_h2d |
| 90 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 91 | |
| 92 | from .interfaces import ( |
| 93 | MultiModalEmbeddings, |
| 94 | SupportsEagle3, |
| 95 | SupportsLoRA, |
| 96 | SupportsMRoPE, |
| 97 | SupportsMultiModal, |
| 98 | SupportsPP, |
| 99 | SupportsTranscription, |
| 100 | ) |
| 101 | from .qwen2_5_omni_thinker import ( |
| 102 | Qwen2_5OmniAudioFeatureInputs, |
| 103 | Qwen2_5OmniConditionalGenerationMixin, |
| 104 | Qwen2_5OmniThinkerDummyInputsBuilder, |
| 105 | Qwen2_5OmniThinkerMultiModalDataParser, |
| 106 | Qwen2_5OmniThinkerMultiModalProcessor, |
| 107 | _get_feature_tensor, |
| 108 | check_interleaved_audio_video, |
| 109 | merge_interleaved_embeddings, |
| 110 | ) |
| 111 | from .qwen2_5_vl import ( |
| 112 | Qwen2_5_VisionAttention, |
| 113 | Qwen2_5_VLImageInputs, |
| 114 | Qwen2_5_VLProcessingInfo, |
| 115 | Qwen2_5_VLVideoInputs, |
| 116 | ) |
| 117 | from .qwen3_moe import Qwen3MoeForCausalLM, Qwen3MoeModel |
| 118 | from .utils import ( |
| 119 | AutoWeightsLoader, |
| 120 | StageMissingLayer, |
| 121 | WeightsMapper, |
| 122 | _merge_multimodal_embeddings, |
| 123 | maybe_prefix, |
| 124 | ) |
| 125 | from .vision import get_vit_attn_backend |
| 126 | |
| 127 | logger = init_logger(__name__) |
| 128 | |
| 129 | # Speech input languages supported by Qwen3-Omni |
| 130 | # From: https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct |
| 131 | ISO639_1_SUPPORTED_LANGS = { |
| 132 | "en": "English", |
| 133 | "zh": "Chinese", |
| 134 | "ko": "Korean", |
| 135 | "ja": "Japanese", |
| 136 | "de": "German", |
| 137 | "ru": "Russian", |
| 138 | "it": "Italian", |
| 139 | "fr": "French", |
| 140 | "es": "Spanish", |
| 141 | "pt": "Portuguese", |
| 142 | "ms": "Malay", |
| 143 | "nl": "Dutch", |
| 144 | "id": "Indonesian", |
| 145 | "tr": "Turkish", |
| 146 | "vi": "Vietnamese", |
| 147 | "yue": "Cantonese", |
| 148 | "ar": "Arabic", |
| 149 | "ur": "Urdu", |
| 150 | } |
| 151 | |
| 152 | |
| 153 | def _get_feat_extract_output_lengths(input_lengths: torch.Tensor): |
| 154 | input_lengths_leave = input_lengths % 100 |
| 155 | feat_lengths = (input_lengths_leave - 1) // 2 + 1 |
| 156 | output_lengths = ( |
| 157 | ((feat_lengths - 1) // 2 + 1 - 1) // 2 + 1 + (input_lengths // 100) * 13 |
| 158 | ) |
| 159 | return output_lengths |
| 160 | |
| 161 | |
| 162 | # ============= Audio Encoder Components ============= |
| 163 | |
| 164 | |
| 165 | class SinusoidsPositionEmbedding(nn.Module): |
| 166 | """Sinusoidal position embedding for audio encoder.""" |
| 167 | |
| 168 | def __init__(self, length: int, channels: int, max_timescale: int = 10000): |
| 169 | super().__init__() |
| 170 | self.length = length |
| 171 | self.channels = channels |
| 172 | self.max_timescale = max_timescale |
| 173 | |
| 174 | if channels % 2 != 0: |
| 175 | raise ValueError("SinusoidsPositionEmbedding needs even channels input") |
| 176 | |
| 177 | log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1) |
| 178 | inv_timescales = torch.exp( |
| 179 | -log_timescale_increment * torch.arange(channels // 2).float() |
| 180 | ) |
| 181 | scaled_time = ( |
| 182 | torch.arange(length)[:, np.newaxis] * inv_timescales[np.newaxis, :] |
| 183 | ) |
| 184 | positional_embedding = torch.cat( |
| 185 | [torch.sin(scaled_time), torch.cos(scaled_time)], dim=1 |
| 186 | ) |
| 187 | self.register_buffer( |
| 188 | "positional_embedding", positional_embedding, persistent=False |
| 189 | ) |
| 190 | |
| 191 | def forward(self, seqlen: int) -> torch.Tensor: |
| 192 | return self.positional_embedding[:seqlen, :] |
| 193 | |
| 194 | |
| 195 | class Qwen3OmniMoeAudioAttention(nn.Module): |
| 196 | """Multi-headed attention for Qwen3-Omni Audio Encoder using MMEncoderAttention.""" |
| 197 | |
| 198 | def __init__( |
| 199 | self, |
| 200 | config: Qwen3OmniMoeAudioEncoderConfig, |
| 201 | prefix: str = "", |
| 202 | ): |
| 203 | super().__init__() |
| 204 | self.embed_dim = config.d_model |
| 205 | self.num_heads = config.encoder_attention_heads |
| 206 | self.head_dim = self.embed_dim // self.num_heads |
| 207 | # Audio encoder uses 20 heads. Shard across TP when divisible |
| 208 | # (e.g. TP=2/4); otherwise keep unreplicated (e.g. TP=8). |
| 209 | tp_size = get_tensor_model_parallel_world_size() |
| 210 | disable_tp = self.num_heads % tp_size != 0 |
| 211 | effective_tp = 1 if disable_tp else tp_size |
| 212 | self.num_local_heads = self.num_heads // effective_tp |
| 213 | |
| 214 | if (self.head_dim * self.num_heads) != self.embed_dim: |
| 215 | raise ValueError( |
| 216 | f"embed_dim must be divisible by num_heads (got `embed_dim`: " |
| 217 | f"{self.embed_dim} and `num_heads`: {self.num_heads})." |
| 218 | ) |
| 219 | |
| 220 | self.scaling = self.head_dim**-0.5 |
| 221 | |
| 222 | self.qkv_proj = QKVParallelLinear( |
| 223 | hidden_size=self.embed_dim, |
| 224 | head_size=self.head_dim, |
| 225 | total_num_heads=self.num_heads, |
| 226 | total_num_kv_heads=self.num_heads, |
| 227 | bias=True, |
| 228 | disable_tp=disable_tp, |
| 229 | prefix=f"{prefix}.qkv_proj", |
| 230 | ) |
| 231 | |
| 232 | self.out_proj = RowParallelLinear( |
| 233 | input_size=self.embed_dim, |
| 234 | output_size=self.embed_dim, |
| 235 | bias=True, |
| 236 | disable_tp=disable_tp, |
| 237 | prefix=f"{prefix}.out_proj", |
| 238 | ) |
| 239 | |
| 240 | self.attn = MMEncoderAttention( |
| 241 | num_heads=self.num_local_heads, |
| 242 | head_size=self.head_dim, |
| 243 | scale=self.scaling, |
| 244 | prefix=f"{prefix}.attn", |
| 245 | ) |
| 246 | |
| 247 | def forward( |
| 248 | self, |
| 249 | hidden_states: torch.Tensor, |
| 250 | cu_seqlens: torch.Tensor, |
| 251 | max_seqlen: torch.Tensor | None, |
| 252 | ) -> torch.Tensor: |
| 253 | seq_length, _ = hidden_states.size() |
| 254 | qkv, _ = self.qkv_proj(hidden_states) |
| 255 | q, k, v = qkv.chunk(3, dim=-1) |
| 256 | q = q.view(1, seq_length, -1, self.head_dim) |
| 257 | k = k.view(1, seq_length, -1, self.head_dim) |
| 258 | v = v.view(1, seq_length, -1, self.head_dim) |
| 259 | |
| 260 | attn_output = self.attn( |
| 261 | query=q, |
| 262 | key=k, |
| 263 | value=v, |
| 264 | cu_seqlens=cu_seqlens, |
| 265 | max_seqlen=max_seqlen, |
| 266 | ) |
| 267 | |
| 268 | attn_output = attn_output.view(seq_length, -1) |
| 269 | output, _ = self.out_proj(attn_output) |
| 270 | return output |
| 271 | |
| 272 | |
| 273 | class Qwen3OmniMoeAudioEncoderLayer(nn.Module): |
| 274 | """Transformer encoder layer for Qwen3-Omni Audio Encoder.""" |
| 275 | |
| 276 | def __init__( |
| 277 | self, |
| 278 | config: Qwen3OmniMoeAudioEncoderConfig, |
| 279 | prefix: str = "", |
| 280 | ): |
| 281 | super().__init__() |
| 282 | self.embed_dim = config.d_model |
| 283 | self.self_attn = Qwen3OmniMoeAudioAttention( |
| 284 | config, prefix=f"{prefix}.self_attn" |
| 285 | ) |
| 286 | self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) |
| 287 | self.activation_fn = _ACTIVATION_REGISTRY[config.activation_function] |
| 288 | tp_size = get_tensor_model_parallel_world_size() |
| 289 | disable_tp = config.encoder_attention_heads % tp_size != 0 |
| 290 | self.fc1 = ColumnParallelLinear( |
| 291 | self.embed_dim, |
| 292 | config.encoder_ffn_dim, |
| 293 | bias=True, |
| 294 | disable_tp=disable_tp, |
| 295 | prefix=f"{prefix}.fc1", |
| 296 | ) |
| 297 | self.fc2 = RowParallelLinear( |
| 298 | config.encoder_ffn_dim, |
| 299 | self.embed_dim, |
| 300 | bias=True, |
| 301 | disable_tp=disable_tp, |
| 302 | prefix=f"{prefix}.fc2", |
| 303 | ) |
| 304 | self.final_layer_norm = nn.LayerNorm(self.embed_dim) |
| 305 | |
| 306 | def forward( |
| 307 | self, |
| 308 | hidden_states: torch.Tensor, |
| 309 | cu_seqlens: torch.Tensor, |
| 310 | max_seqlen: torch.Tensor | None, |
| 311 | ) -> torch.Tensor: |
| 312 | """Args: |
| 313 | hidden_states: Input tensor of shape (seq_len, hidden_size) |
| 314 | cu_seqlens: Cumulative sequence lengths |
| 315 | max_seqlen: Maximum sequence length in the batch |
| 316 | |
| 317 | """ |
| 318 | residual = hidden_states |
| 319 | hidden_states = self.self_attn_layer_norm(hidden_states) |
| 320 | hidden_states = self.self_attn( |
| 321 | hidden_states=hidden_states, |
| 322 | cu_seqlens=cu_seqlens, |
| 323 | max_seqlen=max_seqlen, |
| 324 | ) |
| 325 | hidden_states = residual + hidden_states |
| 326 | |
| 327 | residual = hidden_states |
| 328 | hidden_states = self.final_layer_norm(hidden_states) |
| 329 | hidden_states, _ = self.fc1(hidden_states) |
| 330 | hidden_states = self.activation_fn(hidden_states) |
| 331 | hidden_states, _ = self.fc2(hidden_states) |
| 332 | hidden_states = residual + hidden_states |
| 333 | |
| 334 | # Clamp for numerical stability with fp16 |
| 335 | if hidden_states.dtype == torch.float16: |
| 336 | clamp_value = torch.finfo(hidden_states.dtype).max - 1000 |
| 337 | hidden_states = torch.clamp( |
| 338 | hidden_states, min=-clamp_value, max=clamp_value |
| 339 | ) |
| 340 | |
| 341 | return hidden_states |
| 342 | |
| 343 | |
| 344 | class Qwen3OmniMoeAudioEncoder(nn.Module): |
| 345 | """vLLM-native Qwen3-Omni Audio Encoder.""" |
| 346 | |
| 347 | hf_to_vllm_mapper = WeightsMapper( |
| 348 | orig_to_new_stacked={ |
| 349 | ".self_attn.q_proj.": (".self_attn.qkv_proj.", "q"), |
| 350 | ".self_attn.k_proj.": (".self_attn.qkv_proj.", "k"), |
| 351 | ".self_attn.v_proj.": (".self_attn.qkv_proj.", "v"), |
| 352 | } |
| 353 | ) |
| 354 | |
| 355 | def __init__( |
| 356 | self, |
| 357 | config: Qwen3OmniMoeAudioEncoderConfig, |
| 358 | prefix: str = "", |
| 359 | ): |
| 360 | super().__init__() |
| 361 | |
| 362 | embed_dim = config.d_model |
| 363 | self.num_mel_bins = config.num_mel_bins |
| 364 | self.max_source_positions = config.max_source_positions |
| 365 | self.n_window = config.n_window |
| 366 | self.n_window_infer = config.n_window_infer |
| 367 | self.conv_chunksize = config.conv_chunksize |
| 368 | |
| 369 | # Position embedding |
| 370 | self.positional_embedding = SinusoidsPositionEmbedding( |
| 371 | self.max_source_positions, embed_dim |
| 372 | ) |
| 373 | |
| 374 | # Convolutional layers for mel-spectrogram processing |
| 375 | self.conv2d1 = nn.Conv2d(1, config.downsample_hidden_size, 3, 2, padding=1) |
| 376 | self.conv2d2 = nn.Conv2d( |
| 377 | config.downsample_hidden_size, |
| 378 | config.downsample_hidden_size, |
| 379 | 3, |
| 380 | 2, |
| 381 | padding=1, |
| 382 | ) |
| 383 | self.conv2d3 = nn.Conv2d( |
| 384 | config.downsample_hidden_size, |
| 385 | config.downsample_hidden_size, |
| 386 | 3, |
| 387 | 2, |
| 388 | padding=1, |
| 389 | ) |
| 390 | |
| 391 | conv_out_dim = config.downsample_hidden_size * ( |
| 392 | (((config.num_mel_bins + 1) // 2 + 1) // 2 + 1) // 2 |
| 393 | ) |
| 394 | self.conv_out = ReplicatedLinear( |
| 395 | conv_out_dim, |
| 396 | config.d_model, |
| 397 | bias=False, |
| 398 | return_bias=False, |
| 399 | prefix=f"{prefix}.conv_out", |
| 400 | ) |
| 401 | |
| 402 | # Transformer encoder layers |
| 403 | self.layers = nn.ModuleList( |
| 404 | [ |
| 405 | Qwen3OmniMoeAudioEncoderLayer( |
| 406 | config, |
| 407 | prefix=f"{prefix}.layers.{i}", |
| 408 | ) |
| 409 | for i in range(config.encoder_layers) |
| 410 | ] |
| 411 | ) |
| 412 | |
| 413 | # Output layers |
| 414 | self.ln_post = nn.LayerNorm(config.d_model) |
| 415 | self.proj1 = ReplicatedLinear( |
| 416 | config.d_model, |
| 417 | config.d_model, |
| 418 | bias=True, |
| 419 | return_bias=False, |
| 420 | prefix=f"{prefix}.proj1", |
| 421 | ) |
| 422 | self.act = _ACTIVATION_REGISTRY[config.activation_function] |
| 423 | self.proj2 = ReplicatedLinear( |
| 424 | config.d_model, |
| 425 | config.output_dim, |
| 426 | bias=True, |
| 427 | return_bias=False, |
| 428 | prefix=f"{prefix}.proj2", |
| 429 | ) |
| 430 | |
| 431 | # Get attention backend |
| 432 | self.attn_backend = get_vit_attn_backend( |
| 433 | head_size=config.d_model // config.encoder_attention_heads, |
| 434 | dtype=torch.get_default_dtype(), |
| 435 | ) |
| 436 | |
| 437 | def compute_attn_mask_seqlen(self, cu_seqlens: torch.Tensor) -> torch.Tensor | None: |
| 438 | """Compute max_seqlen only for flash attention backends.""" |
| 439 | max_seqlen = None |
| 440 | if self.attn_backend in { |
| 441 | AttentionBackendEnum.FLASH_ATTN, |
| 442 | AttentionBackendEnum.ROCM_AITER_FA, |
| 443 | AttentionBackendEnum.TRITON_ATTN, |
| 444 | }: |
| 445 | max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max() |
| 446 | return max_seqlen |
| 447 | |
| 448 | @property |
| 449 | def dtype(self) -> torch.dtype: |
| 450 | return self.conv2d1.weight.dtype |
| 451 | |
| 452 | @property |
| 453 | def device(self) -> torch.device: |
| 454 | return self.conv2d1.weight.device |
| 455 | |
| 456 | def forward( |
| 457 | self, |
| 458 | input_features: torch.Tensor, |
| 459 | feature_lens: torch.Tensor, |
| 460 | aftercnn_lens: torch.Tensor, |
| 461 | ): |
| 462 | # Compute chunk information. |
| 463 | chunk_num = torch.ceil(feature_lens / (self.n_window * 2)).long() |
| 464 | |
| 465 | with gpu_sync_allowed(): |
| 466 | total_chunks = int(chunk_num.sum()) |
| 467 | |
| 468 | chunk_lengths = async_tensor_h2d( |
| 469 | [self.n_window * 2] * total_chunks, |
| 470 | dtype=torch.long, |
| 471 | device=feature_lens.device, |
| 472 | ) |
| 473 | tail_chunk_index = F.pad(chunk_num, (1, 0), value=-1).cumsum(0)[1:] |
| 474 | chunk_lengths[tail_chunk_index] = feature_lens % (self.n_window * 2) |
| 475 | chunk_lengths[chunk_lengths == 0] = self.n_window * 2 |
| 476 | |
| 477 | # Split input features into chunks and pad |
| 478 | with gpu_sync_allowed(): |
| 479 | chunk_lengths_list = chunk_lengths.tolist() |
| 480 | chunk_list = input_features.T.split(chunk_lengths_list, dim=0) |
| 481 | padded_feature = nn.utils.rnn.pad_sequence( |
| 482 | chunk_list, batch_first=True |
| 483 | ).transpose(1, 2) |
| 484 | |
| 485 | # Compute feature lengths after CNN |
| 486 | feature_lens_after_cnn = self._get_cnn_output_lengths(chunk_lengths) |
| 487 | # Vectorized mask creation: avoid creating many small tensors. |
| 488 | with gpu_sync_allowed(): |
| 489 | max_len_after_cnn = feature_lens_after_cnn.max().item() |
| 490 | indices = torch.arange(max_len_after_cnn, device=padded_feature.device) |
| 491 | padded_mask_after_cnn = indices.unsqueeze(0) < feature_lens_after_cnn.unsqueeze( |
| 492 | 1 |
| 493 | ) |
| 494 | |
| 495 | # Add channel dimension for conv2d |
| 496 | padded_feature = padded_feature.unsqueeze(1) |
| 497 | |
| 498 | # Apply convolutional layers (chunk if needed to avoid OOM) |
| 499 | if padded_feature.size(0) <= self.conv_chunksize: |
| 500 | # Fast path: no chunking needed |
| 501 | padded_embed = F.gelu(self.conv2d1(padded_feature)) |
| 502 | padded_embed = F.gelu(self.conv2d2(padded_embed)) |
| 503 | padded_embed = F.gelu(self.conv2d3(padded_embed)) |
| 504 | else: |
| 505 | # Chunked processing to avoid OOM |
| 506 | padded_embeds = [] |
| 507 | for chunk in padded_feature.split(self.conv_chunksize, dim=0): |
| 508 | padded_embed = F.gelu(self.conv2d1(chunk)) |
| 509 | padded_embed = F.gelu(self.conv2d2(padded_embed)) |
| 510 | padded_embed = F.gelu(self.conv2d3(padded_embed)) |
| 511 | padded_embeds.append(padded_embed) |
| 512 | padded_embed = torch.cat(padded_embeds, dim=0) |
| 513 | |
| 514 | # (batch, channels, freq, time) -> (batch, time, channels*freq) |
| 515 | b, c, f, t = padded_embed.size() |
| 516 | padded_embed = self.conv_out( |
| 517 | padded_embed.permute(0, 3, 1, 2).contiguous().view(b, t, c * f) |
| 518 | ) |
| 519 | |
| 520 | # Add positional embedding |
| 521 | positional_embedding = ( |
| 522 | self.positional_embedding.positional_embedding[: padded_embed.shape[1], :] |
| 523 | .unsqueeze(0) |
| 524 | .to(padded_embed.dtype) |
| 525 | ) |
| 526 | padded_embed = padded_embed + positional_embedding |
| 527 | |
| 528 | with gpu_sync_allowed(): |
| 529 | hidden_states = padded_embed[padded_mask_after_cnn] |
| 530 | # Use tolist() for efficient batch conversion from tensor to Python. |
| 531 | aftercnn_lens_list = aftercnn_lens.tolist() |
| 532 | |
| 533 | # Compute cumulative sequence lengths for chunked attention |
| 534 | cu_chunk_lens = [0] |
| 535 | window_aftercnn = padded_mask_after_cnn.shape[-1] * ( |
| 536 | self.n_window_infer // (self.n_window * 2) |
| 537 | ) |
| 538 | for cnn_len in aftercnn_lens_list: |
| 539 | num_full_chunks = cnn_len // window_aftercnn |
| 540 | remainder = cnn_len % window_aftercnn |
| 541 | cu_chunk_lens.extend([window_aftercnn] * num_full_chunks) |
| 542 | if remainder: |
| 543 | cu_chunk_lens.append(remainder) |
| 544 | cu_seqlens = async_tensor_h2d( |
| 545 | cu_chunk_lens, dtype=torch.int32, device=aftercnn_lens.device |
| 546 | ).cumsum(-1, dtype=torch.int32) |
| 547 | |
| 548 | max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens) |
| 549 | |
| 550 | # Apply transformer layers |
| 551 | for encoder_layer in self.layers: |
| 552 | hidden_states = encoder_layer( |
| 553 | hidden_states, |
| 554 | cu_seqlens, |
| 555 | max_seqlen, |
| 556 | ) |
| 557 | |
| 558 | # Apply output layers |
| 559 | hidden_states = self.ln_post(hidden_states) |
| 560 | hidden_states = self.proj1(hidden_states) |
| 561 | hidden_states = self.act(hidden_states) |
| 562 | hidden_states = self.proj2(hidden_states) |
| 563 | |
| 564 | return hidden_states |
| 565 | |
| 566 | def _get_cnn_output_lengths(self, input_lengths: torch.Tensor) -> torch.Tensor: |
| 567 | """Compute output lengths after the three conv2d layers.""" |
| 568 | lengths = input_lengths |
| 569 | for _ in range(3): |
| 570 | lengths = (lengths - 1) // 2 + 1 |
| 571 | return lengths |
| 572 | |
| 573 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 574 | loader = AutoWeightsLoader(self) |
| 575 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 576 | |
| 577 | |
| 578 | class Qwen3_VisionPatchEmbed(nn.Module): |
| 579 | def __init__( |
| 580 | self, |
| 581 | patch_size: int = 14, |
| 582 | temporal_patch_size: int = 2, |
| 583 | in_channels: int = 3, |
| 584 | hidden_size: int = 1152, |
| 585 | ) -> None: |
| 586 | super().__init__() |
| 587 | self.patch_size = patch_size |
| 588 | self.temporal_patch_size = temporal_patch_size |
| 589 | self.hidden_size = hidden_size |
| 590 | |
| 591 | kernel_size = (temporal_patch_size, patch_size, patch_size) |
| 592 | self.proj = Conv3dLayer( |
| 593 | in_channels, |
| 594 | hidden_size, |
| 595 | kernel_size=kernel_size, |
| 596 | stride=kernel_size, |
| 597 | bias=True, |
| 598 | ) |
| 599 | |
| 600 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 601 | L, _ = x.shape |
| 602 | x = x.view(L, -1, self.temporal_patch_size, self.patch_size, self.patch_size) |
| 603 | x = self.proj(x).view(L, self.hidden_size) |
| 604 | return x |
| 605 | |
| 606 | |
| 607 | class Qwen3_VisionMLP(nn.Module): |
| 608 | def __init__( |
| 609 | self, |
| 610 | in_features: int, |
| 611 | hidden_features: int, |
| 612 | bias: bool = False, |
| 613 | act_fn: Callable[[torch.Tensor], torch.Tensor] = F.silu, |
| 614 | quant_config: QuantizationConfig | None = None, |
| 615 | prefix: str = "", |
| 616 | ): |
| 617 | super().__init__() |
| 618 | self.linear_fc1 = ColumnParallelLinear( |
| 619 | in_features, |
| 620 | hidden_features, |
| 621 | bias=bias, |
| 622 | quant_config=quant_config, |
| 623 | return_bias=False, |
| 624 | prefix=f"{prefix}.linear_fc1", |
| 625 | ) |
| 626 | self.linear_fc2 = RowParallelLinear( |
| 627 | hidden_features, |
| 628 | in_features, |
| 629 | bias=bias, |
| 630 | quant_config=quant_config, |
| 631 | return_bias=False, |
| 632 | prefix=f"{prefix}.linear_fc2", |
| 633 | ) |
| 634 | self.act_fn = act_fn |
| 635 | |
| 636 | def forward(self, x: torch.Tensor): |
| 637 | mlp_output = self.linear_fc2(self.act_fn(self.linear_fc1(x))) |
| 638 | return mlp_output |
| 639 | |
| 640 | |
| 641 | class Qwen3_VisionBlock(nn.Module): |
| 642 | def __init__( |
| 643 | self, |
| 644 | dim: int, |
| 645 | num_heads: int, |
| 646 | mlp_hidden_dim: int, |
| 647 | act_fn: Callable[[torch.Tensor], torch.Tensor] = F.silu, |
| 648 | norm_layer: Callable[[int], nn.Module] | None = None, |
| 649 | quant_config: QuantizationConfig | None = None, |
| 650 | prefix: str = "", |
| 651 | ) -> None: |
| 652 | super().__init__() |
| 653 | if norm_layer is None: |
| 654 | norm_layer = partial(nn.LayerNorm, eps=1e-6) |
| 655 | self.norm1 = norm_layer(dim) |
| 656 | self.norm2 = norm_layer(dim) |
| 657 | self.attn = Qwen2_5_VisionAttention( |
| 658 | embed_dim=dim, |
| 659 | num_heads=num_heads, |
| 660 | projection_size=dim, |
| 661 | quant_config=quant_config, |
| 662 | prefix=f"{prefix}.attn", |
| 663 | ) |
| 664 | self.mlp = Qwen3_VisionMLP( |
| 665 | dim, |
| 666 | mlp_hidden_dim, |
| 667 | act_fn=act_fn, |
| 668 | bias=True, |
| 669 | quant_config=quant_config, |
| 670 | prefix=f"{prefix}.mlp", |
| 671 | ) |
| 672 | |
| 673 | def forward( |
| 674 | self, |
| 675 | x: torch.Tensor, |
| 676 | cu_seqlens: torch.Tensor, |
| 677 | rotary_pos_emb_cos: torch.Tensor, |
| 678 | rotary_pos_emb_sin: torch.Tensor, |
| 679 | max_seqlen: torch.Tensor | None, # Only used for Flash Attention |
| 680 | sequence_lengths: torch.Tensor | None, # Only used for FlashInfer CuDNN backend |
| 681 | ) -> torch.Tensor: |
| 682 | x = x + self.attn( |
| 683 | self.norm1(x), |
| 684 | cu_seqlens=cu_seqlens, |
| 685 | rotary_pos_emb_cos=rotary_pos_emb_cos, |
| 686 | rotary_pos_emb_sin=rotary_pos_emb_sin, |
| 687 | max_seqlen=max_seqlen, |
| 688 | sequence_lengths=sequence_lengths, |
| 689 | ) |
| 690 | |
| 691 | x = x + self.mlp(self.norm2(x)) |
| 692 | return x |
| 693 | |
| 694 | |
| 695 | class Qwen3_VisionPatchMerger(nn.Module): |
| 696 | def __init__( |
| 697 | self, |
| 698 | d_model: int, |
| 699 | context_dim: int, |
| 700 | norm_layer: Callable[[int], nn.Module] | None = None, |
| 701 | spatial_merge_size: int = 2, |
| 702 | use_postshuffle_norm: bool = False, |
| 703 | quant_config: QuantizationConfig | None = None, |
| 704 | prefix: str = "", |
| 705 | ) -> None: |
| 706 | super().__init__() |
| 707 | self.hidden_size = context_dim * (spatial_merge_size**2) |
| 708 | |
| 709 | self.use_postshuffle_norm = use_postshuffle_norm |
| 710 | if self.use_postshuffle_norm: |
| 711 | context_dim = self.hidden_size |
| 712 | |
| 713 | if norm_layer is None: |
| 714 | norm_layer = partial(nn.LayerNorm, eps=1e-6) |
| 715 | self.use_postshuffle_norm = use_postshuffle_norm |
| 716 | self.ln_q = norm_layer( |
| 717 | self.hidden_size if use_postshuffle_norm else context_dim |
| 718 | ) |
| 719 | self.mlp = nn.ModuleList( |
| 720 | [ |
| 721 | ColumnParallelLinear( |
| 722 | self.hidden_size, |
| 723 | self.hidden_size, |
| 724 | bias=True, |
| 725 | quant_config=quant_config, |
| 726 | prefix=f"{prefix}.mlp.0", |
| 727 | ), |
| 728 | nn.GELU(), |
| 729 | RowParallelLinear( |
| 730 | self.hidden_size, |
| 731 | d_model, |
| 732 | bias=True, |
| 733 | quant_config=quant_config, |
| 734 | prefix=f"{prefix}.mlp.2", |
| 735 | ), |
| 736 | ] |
| 737 | ) |
| 738 | |
| 739 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 740 | if self.use_postshuffle_norm: |
| 741 | x = self.ln_q(x.view(-1, self.hidden_size)) |
| 742 | else: |
| 743 | x = self.ln_q(x).view(-1, self.hidden_size) |
| 744 | |
| 745 | mlp_fc1, mlp_act, mlp_fc2 = self.mlp |
| 746 | x_parallel, _ = mlp_fc1(x) |
| 747 | x_parallel = mlp_act(x_parallel) |
| 748 | out, _ = mlp_fc2(x_parallel) |
| 749 | return out |
| 750 | |
| 751 | |
| 752 | class Qwen3Omni_VisionTransformer(nn.Module): |
| 753 | hf_to_vllm_mapper = WeightsMapper( |
| 754 | orig_to_new_stacked={ |
| 755 | ".attn.q.": (".attn.qkv.", "q"), |
| 756 | ".attn.k.": (".attn.qkv.", "k"), |
| 757 | ".attn.v.": (".attn.qkv.", "v"), |
| 758 | } |
| 759 | ) |
| 760 | |
| 761 | def __init__( |
| 762 | self, |
| 763 | vision_config, |
| 764 | norm_eps: float = 1e-6, |
| 765 | quant_config: QuantizationConfig | None = None, |
| 766 | prefix: str = "", |
| 767 | ) -> None: |
| 768 | super().__init__() |
| 769 | self.hidden_size = vision_config.hidden_size |
| 770 | self.num_heads = vision_config.num_heads |
| 771 | self.image_size = vision_config.image_size |
| 772 | self.patch_size = vision_config.patch_size |
| 773 | self.spatial_merge_size = vision_config.spatial_merge_size |
| 774 | self.spatial_merge_unit = self.spatial_merge_size**2 |
| 775 | self.temporal_patch_size = vision_config.temporal_patch_size |
| 776 | self.num_grid_per_side = self.image_size // self.patch_size |
| 777 | self.apply_vit_abs_pos_embed = vision_config.apply_vit_abs_pos_embed |
| 778 | self.deepstack_visual_indexes = vision_config.deepstack_visual_indexes |
| 779 | |
| 780 | self.patch_embed = Qwen3_VisionPatchEmbed( |
| 781 | patch_size=self.patch_size, |
| 782 | temporal_patch_size=self.temporal_patch_size, |
| 783 | in_channels=vision_config.in_channels, |
| 784 | hidden_size=self.hidden_size, |
| 785 | ) |
| 786 | |
| 787 | # vit pos embedding, TODO: spatial_patch_size vs patch_size |
| 788 | if self.apply_vit_abs_pos_embed: |
| 789 | self.pos_embed = nn.Embedding(self.num_grid_per_side**2, self.hidden_size) |
| 790 | else: |
| 791 | self.pos_embed = nn.Parameter( |
| 792 | torch.empty([1, self.num_grid_per_side**2, self.hidden_size]) |
| 793 | ) |
| 794 | |
| 795 | norm_layer = partial(nn.LayerNorm, eps=norm_eps) |
| 796 | head_dim = self.hidden_size // self.num_heads |
| 797 | self.rotary_pos_emb = get_rope( |
| 798 | head_size=head_dim, |
| 799 | max_position=8192, |
| 800 | is_neox_style=True, |
| 801 | rope_parameters={"partial_rotary_factor": 0.5}, |
| 802 | ) |
| 803 | |
| 804 | self.blocks = nn.ModuleList( |
| 805 | [ |
| 806 | Qwen3_VisionBlock( |
| 807 | dim=self.hidden_size, |
| 808 | num_heads=self.num_heads, |
| 809 | mlp_hidden_dim=vision_config.intermediate_size, |
| 810 | act_fn=_ACTIVATION_REGISTRY[vision_config.hidden_act], |
| 811 | norm_layer=norm_layer, |
| 812 | quant_config=quant_config, |
| 813 | prefix=f"{prefix}.blocks.{layer_idx}", |
| 814 | ) |
| 815 | for layer_idx in range(vision_config.depth) |
| 816 | ] |
| 817 | ) |
| 818 | self.merger = Qwen3_VisionPatchMerger( |
| 819 | d_model=vision_config.out_hidden_size, |
| 820 | context_dim=self.hidden_size, |
| 821 | norm_layer=norm_layer, |
| 822 | spatial_merge_size=self.spatial_merge_size, |
| 823 | quant_config=quant_config, |
| 824 | prefix=f"{prefix}.merger", |
| 825 | ) |
| 826 | if self.deepstack_visual_indexes is not None: |
| 827 | self.merger_list = nn.ModuleList( |
| 828 | [ |
| 829 | Qwen3_VisionPatchMerger( |
| 830 | d_model=vision_config.out_hidden_size, |
| 831 | context_dim=self.hidden_size, |
| 832 | spatial_merge_size=self.spatial_merge_size, |
| 833 | use_postshuffle_norm=True, |
| 834 | norm_layer=norm_layer, |
| 835 | quant_config=quant_config, |
| 836 | prefix=f"{prefix}.merger_list.{layer_idx}", |
| 837 | ) |
| 838 | for layer_idx in range(len(self.deepstack_visual_indexes)) |
| 839 | ] |
| 840 | ) |
| 841 | |
| 842 | self.attn_backend = get_vit_attn_backend( |
| 843 | head_size=head_dim, |
| 844 | dtype=torch.get_default_dtype(), |
| 845 | ) |
| 846 | |
| 847 | @property |
| 848 | def dtype(self) -> torch.dtype: |
| 849 | return self.patch_embed.proj.weight.dtype |
| 850 | |
| 851 | @property |
| 852 | def device(self) -> torch.device: |
| 853 | return self.patch_embed.proj.weight.device |
| 854 | |
| 855 | def rot_pos_emb(self, grid_thw): |
| 856 | max_grid_size = grid_thw[:, 1:].max() |
| 857 | |
| 858 | # Use pre-computed cos_sin_cache from RotaryEmbedding |
| 859 | cos, sin = self.rotary_pos_emb.get_cos_sin(max_grid_size) |
| 860 | |
| 861 | pos_ids = [] |
| 862 | for t, h, w in grid_thw: |
| 863 | hpos_ids = torch.arange(h, device=cos.device).unsqueeze(1).expand(-1, w) |
| 864 | hpos_ids = hpos_ids.reshape( |
| 865 | h // self.spatial_merge_size, |
| 866 | self.spatial_merge_size, |
| 867 | w // self.spatial_merge_size, |
| 868 | self.spatial_merge_size, |
| 869 | ) |
| 870 | hpos_ids = hpos_ids.permute(0, 2, 1, 3) |
| 871 | hpos_ids = hpos_ids.flatten() |
| 872 | |
| 873 | wpos_ids = torch.arange(w, device=cos.device).unsqueeze(0).expand(h, -1) |
| 874 | wpos_ids = wpos_ids.reshape( |
| 875 | h // self.spatial_merge_size, |
| 876 | self.spatial_merge_size, |
| 877 | w // self.spatial_merge_size, |
| 878 | self.spatial_merge_size, |
| 879 | ) |
| 880 | wpos_ids = wpos_ids.permute(0, 2, 1, 3) |
| 881 | wpos_ids = wpos_ids.flatten() |
| 882 | pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1)) |
| 883 | pos_ids = torch.cat(pos_ids, dim=0) |
| 884 | |
| 885 | cos_combined = cos[pos_ids].flatten(1) |
| 886 | sin_combined = sin[pos_ids].flatten(1) |
| 887 | |
| 888 | return cos_combined, sin_combined |
| 889 | |
| 890 | def fast_pos_embed_interpolate(self, grid_thw: list[list[int]]) -> torch.Tensor: |
| 891 | num_grid_per_side = self.num_grid_per_side |
| 892 | m_size = self.spatial_merge_size |
| 893 | hidden_dim = self.pos_embed.embedding_dim |
| 894 | |
| 895 | outputs = [] |
| 896 | for t, h, w in grid_thw: |
| 897 | h_idxs = torch.linspace( |
| 898 | 0, num_grid_per_side - 1, h, dtype=torch.float32, device=self.device |
| 899 | ) |
| 900 | w_idxs = torch.linspace( |
| 901 | 0, num_grid_per_side - 1, w, dtype=torch.float32, device=self.device |
| 902 | ) |
| 903 | |
| 904 | h_floor = h_idxs.to(torch.long) |
| 905 | w_floor = w_idxs.to(torch.long) |
| 906 | h_ceil = torch.clamp(h_floor + 1, max=num_grid_per_side - 1) |
| 907 | w_ceil = torch.clamp(w_floor + 1, max=num_grid_per_side - 1) |
| 908 | |
| 909 | dh = h_idxs - h_floor |
| 910 | dw = w_idxs - w_floor |
| 911 | |
| 912 | # Create meshgrid view for all h, w vars |
| 913 | dh_grid, dw_grid = torch.meshgrid(dh, dw, indexing="ij") |
| 914 | h_floor_grid, w_floor_grid = torch.meshgrid(h_floor, w_floor, indexing="ij") |
| 915 | h_ceil_grid, w_ceil_grid = torch.meshgrid(h_ceil, w_ceil, indexing="ij") |
| 916 | h_floor_grid_idx = h_floor_grid * num_grid_per_side |
| 917 | h_ceil_grid_idx = h_ceil_grid * num_grid_per_side |
| 918 | |
| 919 | # original computation of weights |
| 920 | # w00 = (1 - dh_grid) * (1 - dw_grid) |
| 921 | # w01 = (1 - dh_grid) * dw_grid |
| 922 | # w10 = dh_grid * (1 - dw_grid) |
| 923 | # w11 = dh_grid * dw_grid |
| 924 | # we reuse w11 here to avoid duplicate |
| 925 | # dh_grid * dw_grid computation |
| 926 | w11 = dh_grid * dw_grid |
| 927 | w10 = dh_grid - w11 |
| 928 | w01 = dw_grid - w11 |
| 929 | w00 = 1 - dh_grid - dw_grid + w11 |
| 930 | |
| 931 | idx00 = h_floor_grid_idx + w_floor_grid |
| 932 | idx01 = h_floor_grid_idx + w_ceil_grid |
| 933 | idx10 = h_ceil_grid_idx + w_floor_grid |
| 934 | idx11 = h_ceil_grid_idx + w_ceil_grid |
| 935 | |
| 936 | indices = torch.stack([idx00, idx01, idx10, idx11], dim=0).reshape(4, -1) |
| 937 | weights = torch.stack([w00, w01, w10, w11], dim=0).reshape(4, -1, 1) |
| 938 | weights = weights.to(dtype=self.dtype, device=self.device) |
| 939 | |
| 940 | embeds = self.pos_embed(indices) |
| 941 | weighted_embeds = embeds * weights |
| 942 | p0, p1, p2, p3 = weighted_embeds.unbind(dim=0) |
| 943 | combined = p0 + p1 + p2 + p3 |
| 944 | |
| 945 | combined = combined.view(h * w, hidden_dim) |
| 946 | repeated = combined.unsqueeze(0).expand(t, -1, -1).contiguous() |
| 947 | repeated = repeated.view( |
| 948 | t, h // m_size, m_size, w // m_size, m_size, hidden_dim |
| 949 | ) |
| 950 | repeated = repeated.permute(0, 1, 3, 2, 4, 5).reshape(-1, hidden_dim) |
| 951 | outputs.append(repeated) |
| 952 | |
| 953 | return torch.cat(outputs, dim=0) |
| 954 | |
| 955 | def compute_attn_mask_seqlen( |
| 956 | self, |
| 957 | cu_seqlens: torch.Tensor, |
| 958 | ) -> torch.Tensor: |
| 959 | max_seqlen = torch.zeros([], device=cu_seqlens.device) |
| 960 | if self.attn_backend in { |
| 961 | AttentionBackendEnum.FLASH_ATTN, |
| 962 | AttentionBackendEnum.ROCM_AITER_FA, |
| 963 | AttentionBackendEnum.TRITON_ATTN, |
| 964 | }: |
| 965 | max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max() |
| 966 | return max_seqlen |
| 967 | |
| 968 | def forward( |
| 969 | self, |
| 970 | x: torch.Tensor, |
| 971 | grid_thw: torch.Tensor, |
| 972 | ) -> torch.Tensor: |
| 973 | hidden_states = x.to(device=self.device, dtype=self.dtype) |
| 974 | hidden_states = self.patch_embed(hidden_states) |
| 975 | |
| 976 | if self.apply_vit_abs_pos_embed: |
| 977 | pos_embeds = self.fast_pos_embed_interpolate(grid_thw) |
| 978 | hidden_states = hidden_states + pos_embeds |
| 979 | rotary_pos_emb_cos, rotary_pos_emb_sin = self.rot_pos_emb(grid_thw) |
| 980 | |
| 981 | # RDNA3 (gfx11) specific bug workaround: torch.repeat_interleave triggers |
| 982 | # kernel crashes. We attempt the operation and catch the RuntimeError |
| 983 | # to switch to a vectorized cumsum + searchsorted approach. |
| 984 | try: |
| 985 | cu_seqlens = torch.repeat_interleave( |
| 986 | grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0] |
| 987 | ).cumsum( |
| 988 | dim=0, |
| 989 | dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32, |
| 990 | ) |
| 991 | cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0) |
| 992 | except RuntimeError: |
| 993 | logger.warning( |
| 994 | "torch.repeat_interleave not executable, " |
| 995 | "switching to vectorized searchsorted implementation." |
| 996 | ) |
| 997 | repeat_counts = grid_thw[:, 0] |
| 998 | values = grid_thw[:, 1] * grid_thw[:, 2] |
| 999 | repeat_cumsum = repeat_counts.cumsum(0) |
| 1000 | total_items = repeat_cumsum[-1].item() |
| 1001 | |
| 1002 | indices = torch.searchsorted( |
| 1003 | repeat_cumsum, |
| 1004 | torch.arange(total_items, device=grid_thw.device), |
| 1005 | right=True, |
| 1006 | ) |
| 1007 | cu_seqlens = values[indices].cumsum( |
| 1008 | dim=0, |
| 1009 | dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32, |
| 1010 | ) |
| 1011 | cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0) |
| 1012 | |
| 1013 | # Move cu_seqlens to GPU; grid_thw may be on CPU during profile_run |
| 1014 | # and FA3 vit attention requires cu_seqlens on CUDA. |
| 1015 | cu_seqlens = async_tensor_h2d(cu_seqlens, self.device) |
| 1016 | hidden_states = hidden_states.unsqueeze(1) |
| 1017 | rotary_pos_emb_cos = rotary_pos_emb_cos.to(hidden_states.device) |
| 1018 | rotary_pos_emb_sin = rotary_pos_emb_sin.to(hidden_states.device) |
| 1019 | max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens) |
| 1020 | |
| 1021 | # Recompute cu_seqlens in numpy from grid_thw to avoid GPU->CPU sync |
| 1022 | grid_thw_np = grid_thw.cpu().numpy() |
| 1023 | cu_seqlens_np = np.repeat( |
| 1024 | grid_thw_np[:, 1] * grid_thw_np[:, 2], grid_thw_np[:, 0] |
| 1025 | ).cumsum(axis=0, dtype=np.int32) |
| 1026 | cu_seqlens_np = np.concatenate([np.zeros(1, dtype=np.int32), cu_seqlens_np]) |
| 1027 | sequence_lengths = MMEncoderAttention.maybe_compute_seq_lens( |
| 1028 | self.attn_backend, |
| 1029 | cu_seqlens_np, |
| 1030 | self.device, |
| 1031 | ) |
| 1032 | |
| 1033 | hidden_states_list = [] |
| 1034 | deepstack_visual_indexes = self.deepstack_visual_indexes |
| 1035 | |
| 1036 | for layer_num, blk in enumerate(self.blocks): |
| 1037 | hidden_states = blk( |
| 1038 | hidden_states, |
| 1039 | cu_seqlens=cu_seqlens, |
| 1040 | rotary_pos_emb_cos=rotary_pos_emb_cos, |
| 1041 | rotary_pos_emb_sin=rotary_pos_emb_sin, |
| 1042 | max_seqlen=max_seqlen, |
| 1043 | sequence_lengths=sequence_lengths, |
| 1044 | ) |
| 1045 | if ( |
| 1046 | deepstack_visual_indexes is not None |
| 1047 | and layer_num in deepstack_visual_indexes |
| 1048 | ): |
| 1049 | hidden_states_list.append(hidden_states) |
| 1050 | |
| 1051 | hidden_states = self.merger(hidden_states) |
| 1052 | |
| 1053 | # processing deepstack |
| 1054 | if deepstack_visual_indexes is not None: |
| 1055 | processed_hidden_states_list = [hidden_states] |
| 1056 | for idx, x in enumerate(hidden_states_list): |
| 1057 | x = self.merger_list[idx](x) |
| 1058 | processed_hidden_states_list.append(x) |
| 1059 | # we cat the original visual features and deepstack features |
| 1060 | # along the feature dim |
| 1061 | hidden_states = torch.cat( |
| 1062 | processed_hidden_states_list, dim=1 |
| 1063 | ) # [seq_len, hidden_size * (1 + depth_of_deepstack)] |
| 1064 | |
| 1065 | return hidden_states |
| 1066 | |
| 1067 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1068 | loader = AutoWeightsLoader(self) |
| 1069 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1070 | |
| 1071 | |
| 1072 | @support_torch_compile( |
| 1073 | dynamic_arg_dims={ |
| 1074 | "input_ids": 0, |
| 1075 | "positions": -1, |
| 1076 | "intermediate_tensors": 0, |
| 1077 | "inputs_embeds": 0, |
| 1078 | "deepstack_input_embeds": 0, |
| 1079 | } |
| 1080 | ) |
| 1081 | class Qwen3MoeLLMModel(Qwen3MoeModel): |
| 1082 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1083 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 1084 | |
| 1085 | self.deepstack_multiscale_layer_start = 1 |
| 1086 | |
| 1087 | def forward( |
| 1088 | self, |
| 1089 | input_ids: torch.Tensor | None, |
| 1090 | positions: torch.Tensor, |
| 1091 | intermediate_tensors: IntermediateTensors | None = None, |
| 1092 | inputs_embeds: torch.Tensor | None = None, |
| 1093 | deepstack_input_embeds: IntermediateTensors | None = None, |
| 1094 | ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]: |
| 1095 | if get_pp_group().is_first_rank: |
| 1096 | if inputs_embeds is not None: |
| 1097 | hidden_states = inputs_embeds |
| 1098 | else: |
| 1099 | hidden_states = self.embed_input_ids(input_ids) |
| 1100 | residual = None |
| 1101 | else: |
| 1102 | assert intermediate_tensors is not None |
| 1103 | hidden_states = intermediate_tensors["hidden_states"] |
| 1104 | residual = intermediate_tensors["residual"] |
| 1105 | aux_hidden_states = self._maybe_add_hidden_state( |
| 1106 | [], self.start_layer, hidden_states, residual |
| 1107 | ) |
| 1108 | for layer_idx, layer in enumerate( |
| 1109 | self.layers[self.start_layer : self.end_layer] |
| 1110 | ): |
| 1111 | layer_idx = layer_idx + self.start_layer |
| 1112 | |
| 1113 | hidden_states, residual = layer( |
| 1114 | positions, |
| 1115 | hidden_states, |
| 1116 | residual, |
| 1117 | ) |
| 1118 | |
| 1119 | if deepstack_input_embeds is not None and layer_idx in range( |
| 1120 | 0, len(deepstack_input_embeds) |
| 1121 | ): |
| 1122 | hidden_states = ( |
| 1123 | hidden_states |
| 1124 | + deepstack_input_embeds[f"deepstack_input_embeds_{layer_idx}"] |
| 1125 | ) |
| 1126 | self._maybe_add_hidden_state( |
| 1127 | aux_hidden_states, layer_idx + 1, hidden_states, residual |
| 1128 | ) |
| 1129 | |
| 1130 | if not get_pp_group().is_last_rank: |
| 1131 | return IntermediateTensors( |
| 1132 | {"hidden_states": hidden_states, "residual": residual} |
| 1133 | ) |
| 1134 | hidden_states, _ = self.norm(hidden_states, residual) |
| 1135 | if aux_hidden_states: |
| 1136 | return hidden_states, aux_hidden_states |
| 1137 | return hidden_states |
| 1138 | |
| 1139 | |
| 1140 | class Qwen3MoeLLMForCausalLM(Qwen3MoeForCausalLM): |
| 1141 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1142 | super(Qwen3MoeForCausalLM, self).__init__() |
| 1143 | config = vllm_config.model_config.hf_config |
| 1144 | quant_config = vllm_config.quant_config |
| 1145 | self.config = config |
| 1146 | self.quant_config = quant_config |
| 1147 | self.model = Qwen3MoeLLMModel( |
| 1148 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1149 | ) |
| 1150 | self.lm_head = ParallelLMHead( |
| 1151 | config.vocab_size, config.hidden_size, quant_config=quant_config |
| 1152 | ) |
| 1153 | if self.config.tie_word_embeddings: |
| 1154 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 1155 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1156 | self.make_empty_intermediate_tensors = ( |
| 1157 | self.model.make_empty_intermediate_tensors |
| 1158 | ) |
| 1159 | |
| 1160 | |
| 1161 | class Qwen3OmniMoeThinkerMultiModalDataParser(Qwen2_5OmniThinkerMultiModalDataParser): |
| 1162 | # Keep the existing raw-media fallback for vision inputs. |
| 1163 | embedding_fields = { |
| 1164 | "audio": Qwen2_5OmniThinkerMultiModalDataParser.embedding_fields["audio"], |
| 1165 | } |
| 1166 | |
| 1167 | |
| 1168 | class Qwen3OmniMoeThinkerProcessingInfo( |
| 1169 | Qwen2AudioProcessingInfo, Qwen2_5_VLProcessingInfo |
| 1170 | ): |
| 1171 | def get_hf_config(self): |
| 1172 | return self.ctx.get_hf_config(Qwen3OmniMoeConfig).thinker_config |
| 1173 | |
| 1174 | def get_hf_processor(self, **kwargs: object) -> Qwen3OmniMoeProcessor: |
| 1175 | processor = self.ctx.get_hf_processor( |
| 1176 | Qwen3OmniMoeProcessor, |
| 1177 | **kwargs, |
| 1178 | ) |
| 1179 | if not hasattr(processor, "audio_token"): |
| 1180 | processor.audio_token = "<|audio_pad|>" |
| 1181 | if not hasattr(processor, "image_token"): |
| 1182 | processor.image_token = "<|image_pad|>" |
| 1183 | if not hasattr(processor, "video_token"): |
| 1184 | processor.video_token = "<|video_pad|>" |
| 1185 | return processor |
| 1186 | |
| 1187 | def get_feature_extractor(self, **kwargs: object): |
| 1188 | hf_processor = self.get_hf_processor(**kwargs) |
| 1189 | feature_extractor = hf_processor.feature_extractor # type: ignore |
| 1190 | assert isinstance(feature_extractor, WhisperFeatureExtractor) |
| 1191 | return feature_extractor |
| 1192 | |
| 1193 | def get_data_parser(self): |
| 1194 | return Qwen3OmniMoeThinkerMultiModalDataParser( |
| 1195 | spatial_merge_size=self.get_hf_config().vision_config.spatial_merge_size, |
| 1196 | target_sr=self.get_feature_extractor().sampling_rate, |
| 1197 | target_channels=self.get_target_channels(), |
| 1198 | expected_hidden_size=self._get_expected_hidden_size(), |
| 1199 | allow_missing_mm_embeddings=self.allow_missing_mm_embeddings, |
| 1200 | ) |
| 1201 | |
| 1202 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 1203 | return {"audio": None, "image": None, "video": None} |
| 1204 | |
| 1205 | def get_mm_max_tokens_per_item( |
| 1206 | self, |
| 1207 | seq_len: int, |
| 1208 | mm_counts: Mapping[str, int] | None = None, |
| 1209 | ) -> Mapping[str, int]: |
| 1210 | mm_counts = mm_counts or {} |
| 1211 | requested_modalities = {m for m, c in mm_counts.items() if c > 0} |
| 1212 | mm_max_tokens: dict[str, int] = {} |
| 1213 | |
| 1214 | if requested_modalities & {"image", "video"}: |
| 1215 | vl_tokens = Qwen2_5_VLProcessingInfo.get_mm_max_tokens_per_item( |
| 1216 | self, |
| 1217 | seq_len=seq_len, |
| 1218 | mm_counts=mm_counts, |
| 1219 | ) |
| 1220 | mm_max_tokens.update( |
| 1221 | { |
| 1222 | m: vl_tokens[m] |
| 1223 | for m in ["image", "video"] |
| 1224 | if m in requested_modalities |
| 1225 | } |
| 1226 | ) |
| 1227 | |
| 1228 | if "audio" in requested_modalities: |
| 1229 | audio_tokens = Qwen2AudioProcessingInfo.get_mm_max_tokens_per_item( |
| 1230 | self, |
| 1231 | seq_len=seq_len, |
| 1232 | mm_counts=mm_counts, |
| 1233 | ) |
| 1234 | mm_max_tokens["audio"] = audio_tokens["audio"] |
| 1235 | |
| 1236 | return mm_max_tokens |
| 1237 | |
| 1238 | |
| 1239 | Qwen3OmniMoeThinkerDummyInputsBuilder = Qwen2_5OmniThinkerDummyInputsBuilder[ |
| 1240 | Qwen3OmniMoeThinkerProcessingInfo |
| 1241 | ] |
| 1242 | |
| 1243 | _I = TypeVar("_I", bound=BaseProcessingInfo, default=Qwen3OmniMoeThinkerProcessingInfo) |
| 1244 | |
| 1245 | |
| 1246 | class Qwen3OmniMoeThinkerMultiModalProcessor( |
| 1247 | Qwen2_5OmniThinkerMultiModalProcessor[_I], |
| 1248 | ): |
| 1249 | def _get_hf_mm_inputs( |
| 1250 | self, |
| 1251 | mm_items: MultiModalDataItems, |
| 1252 | hf_kwargs: Mapping[str, object], |
| 1253 | ) -> HFMultiModalInputs: |
| 1254 | hf_inputs = super()._get_hf_mm_inputs(mm_items, hf_kwargs) |
| 1255 | |
| 1256 | audios = hf_inputs.hf_data.get("audio") |
| 1257 | if audios: |
| 1258 | assert isinstance(audios, (list, tuple)) |
| 1259 | feature_extractor = self.info.get_hf_processor().feature_extractor |
| 1260 | hop_length = feature_extractor.hop_length |
| 1261 | |
| 1262 | def pad_to_hop_length(x: np.ndarray) -> np.ndarray: |
| 1263 | length = x.shape[-1] |
| 1264 | if length % hop_length != 0: |
| 1265 | pad_length = hop_length - (length % hop_length) |
| 1266 | x = np.pad(x, (0, pad_length), mode="constant", constant_values=0) |
| 1267 | return x |
| 1268 | |
| 1269 | # To make sure the cache works with padding=True, we pre-padded |
| 1270 | # the audio to a multiple of hop_length. |
| 1271 | hf_inputs.hf_data["audio"] = [ |
| 1272 | pad_to_hop_length(audio) |
| 1273 | if isinstance(audio, np.ndarray) |
| 1274 | else (pad_to_hop_length(audio[0]), audio[1]) |
| 1275 | for audio in audios |
| 1276 | ] |
| 1277 | |
| 1278 | return hf_inputs |
| 1279 | |
| 1280 | def _postprocess_hf_mm_data( |
| 1281 | self, |
| 1282 | mm_data: Mapping[str, object], |
| 1283 | hf_processor_mm_kwargs: Mapping[str, object], |
| 1284 | processed_data: BatchFeature, |
| 1285 | ) -> BatchFeature: |
| 1286 | processed_data = super()._postprocess_hf_mm_data( |
| 1287 | mm_data, hf_processor_mm_kwargs, processed_data |
| 1288 | ) |
| 1289 | |
| 1290 | if ( |
| 1291 | "audio_feature_lengths" in processed_data |
| 1292 | and "feature_attention_mask" in processed_data |
| 1293 | ): |
| 1294 | audios = mm_data.get("audio", []) |
| 1295 | if audios: |
| 1296 | assert isinstance(audios, (list, tuple)) |
| 1297 | feature_extractor = self.info.get_hf_processor( |
| 1298 | **hf_processor_mm_kwargs |
| 1299 | ).feature_extractor |
| 1300 | hop_length = feature_extractor.hop_length |
| 1301 | |
| 1302 | audio_num_frames = [] |
| 1303 | for audio in audios: |
| 1304 | audio_length = ( |
| 1305 | len(audio[0]) if isinstance(audio, tuple) else len(audio) |
| 1306 | ) |
| 1307 | num_frame = ( |
| 1308 | (audio_length // hop_length) |
| 1309 | if audio_length % hop_length == 0 |
| 1310 | else (audio_length // hop_length - 1) |
| 1311 | ) |
| 1312 | if hf_processor_mm_kwargs.get("truncation", False): |
| 1313 | num_frame = min( |
| 1314 | num_frame, feature_extractor.n_samples // hop_length |
| 1315 | ) |
| 1316 | audio_num_frames.append(num_frame) |
| 1317 | |
| 1318 | processed_data["feature_attention_mask"] = [ |
| 1319 | torch.ones(num_frame) for num_frame in audio_num_frames |
| 1320 | ] |
| 1321 | processed_data["audio_feature_lengths"] = torch.tensor(audio_num_frames) |
| 1322 | |
| 1323 | return processed_data |
| 1324 | |
| 1325 | def _maybe_apply_prompt_updates( |
| 1326 | self, |
| 1327 | mm_items: MultiModalDataItems, |
| 1328 | mm_res: MultiModalProcessingResult, |
| 1329 | ) -> tuple[list[int], Mapping[str, list[PlaceholderFeaturesInfo]]]: |
| 1330 | """Qwen3-Omni reimplements this function to handle `use_audio_in_video`.""" |
| 1331 | mm_item_counts = mm_items.get_all_counts() |
| 1332 | prompt_ids = mm_res.prompt_ids |
| 1333 | mm_kwargs = mm_res.kwargs |
| 1334 | mm_prompt_updates = mm_res.prompt_updates |
| 1335 | |
| 1336 | self._validate_mm_kwargs(mm_kwargs, mm_item_counts) |
| 1337 | |
| 1338 | use_audio_in_video = False |
| 1339 | if "video" in mm_kwargs: |
| 1340 | for item in mm_kwargs["video"]: |
| 1341 | if item and item["use_audio_in_video"].data: |
| 1342 | use_audio_in_video = True |
| 1343 | else: |
| 1344 | use_audio_in_video = False |
| 1345 | # for mutilmodality cache |
| 1346 | if any(item is None for item in mm_kwargs["video"]): |
| 1347 | video_token_id = self.info.get_hf_config().video_token_id |
| 1348 | audio_token_id = self.info.get_hf_config().audio_token_id |
| 1349 | video_audio_item_num = sum( |
| 1350 | id in (video_token_id, audio_token_id) for id in prompt_ids |
| 1351 | ) |
| 1352 | audio_updates_num = len(mm_prompt_updates.get("audio", [])) |
| 1353 | video_updates_num = len(mm_prompt_updates.get("video", [])) |
| 1354 | if video_audio_item_num != video_updates_num + audio_updates_num: |
| 1355 | use_audio_in_video = True |
| 1356 | |
| 1357 | if use_audio_in_video and "audio" in mm_prompt_updates: |
| 1358 | filtered_updates = { |
| 1359 | k: v for k, v in mm_prompt_updates.items() if k != "audio" |
| 1360 | } |
| 1361 | prompt_ids, mm_placeholders = self._apply_prompt_updates( |
| 1362 | prompt_ids, |
| 1363 | filtered_updates, |
| 1364 | ) |
| 1365 | # Derive audio placeholders from video placeholders |
| 1366 | mm_placeholders = self._derive_audio_from_video_placeholders( |
| 1367 | mm_placeholders, mm_prompt_updates |
| 1368 | ) |
| 1369 | else: |
| 1370 | prompt_ids, mm_placeholders = self._apply_prompt_updates( |
| 1371 | prompt_ids, |
| 1372 | mm_prompt_updates, |
| 1373 | ) |
| 1374 | |
| 1375 | self._validate_mm_placeholders(mm_placeholders, mm_item_counts) |
| 1376 | |
| 1377 | return prompt_ids, mm_placeholders |
| 1378 | |
| 1379 | def get_updates_use_audio_in_video( |
| 1380 | self, |
| 1381 | thinker_config: PreTrainedConfig, |
| 1382 | audio_len: int, |
| 1383 | video_grid_thw: list[int] | torch.Tensor, |
| 1384 | video_second_per_grid_t: float, |
| 1385 | ) -> list[int]: |
| 1386 | shift = 0 |
| 1387 | audio_token_id = thinker_config.audio_token_id |
| 1388 | video_token_id = thinker_config.video_token_id |
| 1389 | audio_start_token_id = thinker_config.audio_start_token_id |
| 1390 | audio_end_token_id = thinker_config.audio_end_token_id |
| 1391 | spatial_merge_size = thinker_config.vision_config.spatial_merge_size |
| 1392 | position_id_per_seconds = thinker_config.position_id_per_seconds |
| 1393 | audio_token_indices = np.arange(next(iter([audio_len]))) |
| 1394 | curr_video_grid_thw = next(iter([video_grid_thw])) |
| 1395 | height = curr_video_grid_thw[1] // spatial_merge_size |
| 1396 | width = curr_video_grid_thw[2] // spatial_merge_size |
| 1397 | video_token_indices = np.arange(curr_video_grid_thw[0]).reshape(-1, 1, 1) |
| 1398 | video_token_indices = np.broadcast_to( |
| 1399 | video_token_indices, (video_token_indices.shape[0], height, width) |
| 1400 | ).reshape(-1) |
| 1401 | video_token_indices = ( |
| 1402 | (video_token_indices + shift) |
| 1403 | * next(iter([video_second_per_grid_t])) |
| 1404 | * position_id_per_seconds |
| 1405 | ) |
| 1406 | video_data_index, audio_data_index = 0, 0 |
| 1407 | updates = [audio_start_token_id] |
| 1408 | while video_data_index < len(video_token_indices) and audio_data_index < len( |
| 1409 | audio_token_indices |
| 1410 | ): |
| 1411 | if ( |
| 1412 | video_token_indices[video_data_index] |
| 1413 | <= audio_token_indices[audio_data_index] |
| 1414 | ): |
| 1415 | updates += [video_token_id] |
| 1416 | video_data_index += 1 |
| 1417 | else: |
| 1418 | updates += [audio_token_id] |
| 1419 | audio_data_index += 1 |
| 1420 | if video_data_index < len(video_token_indices): |
| 1421 | updates += [video_token_id] * (len(video_token_indices) - video_data_index) |
| 1422 | if audio_data_index < len(audio_token_indices): |
| 1423 | updates += [audio_token_id] * (len(audio_token_indices) - audio_data_index) |
| 1424 | updates += [audio_end_token_id] |
| 1425 | return updates |
| 1426 | |
| 1427 | def _get_prompt_updates( |
| 1428 | self, |
| 1429 | mm_items: MultiModalDataItems, |
| 1430 | hf_processor_mm_kwargs: Mapping[str, Any], |
| 1431 | out_mm_kwargs: MultiModalKwargsItems, |
| 1432 | ) -> Sequence[PromptUpdate]: |
| 1433 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 1434 | tokenizer = self.info.get_tokenizer() |
| 1435 | assert isinstance(self.info, Qwen2_5_VLProcessingInfo) |
| 1436 | image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs) |
| 1437 | vocab = tokenizer.get_vocab() |
| 1438 | |
| 1439 | audio_token = processor.audio_token |
| 1440 | image_token = processor.image_token |
| 1441 | video_token = processor.video_token |
| 1442 | audio_token_id = vocab[audio_token] |
| 1443 | image_token_id = vocab[image_token] |
| 1444 | video_token_id = vocab[video_token] |
| 1445 | |
| 1446 | out_mm_data = out_mm_kwargs.get_data() |
| 1447 | audio_feature_lengths = out_mm_data.get("audio_feature_lengths") |
| 1448 | feature_attention_mask = out_mm_data.get("feature_attention_mask") |
| 1449 | if audio_feature_lengths is None and feature_attention_mask is None: |
| 1450 | audio_output_lengths = [] |
| 1451 | elif audio_feature_lengths is not None: |
| 1452 | audio_output_lens = _get_feat_extract_output_lengths(audio_feature_lengths) |
| 1453 | audio_output_lengths = audio_output_lens.tolist() |
| 1454 | elif feature_attention_mask is not None: |
| 1455 | assert isinstance(feature_attention_mask, torch.Tensor) |
| 1456 | audio_output_lens = _get_feat_extract_output_lengths( |
| 1457 | feature_attention_mask.sum(-1) |
| 1458 | ) |
| 1459 | audio_output_lengths = audio_output_lens.tolist() |
| 1460 | |
| 1461 | # number of audios read from video. |
| 1462 | audio_in_video_item_idx = 0 |
| 1463 | audio_item_idx = 0 |
| 1464 | |
| 1465 | def get_replacement_qwen2_audio(item_idx: int): |
| 1466 | nonlocal audio_item_idx |
| 1467 | item_idx += audio_in_video_item_idx |
| 1468 | |
| 1469 | audio_item_idx += 1 |
| 1470 | |
| 1471 | num_features = audio_output_lengths[item_idx] |
| 1472 | if num_features == 0: |
| 1473 | audios = mm_items.get_items("audio", AudioProcessorItems) |
| 1474 | audio = audios.get(item_idx) |
| 1475 | assert audio is not None |
| 1476 | raise ValueError( |
| 1477 | f"The audio {audio} (len={len(audio)}) is too short " |
| 1478 | "to be represented inside the model" |
| 1479 | ) |
| 1480 | |
| 1481 | return [audio_token_id] * num_features |
| 1482 | |
| 1483 | def get_replacement_qwen2_vision(item_idx: int, modality: str): |
| 1484 | grid_thw = out_mm_data[f"{modality}_grid_thw"][item_idx] |
| 1485 | assert isinstance(grid_thw, torch.Tensor) |
| 1486 | merge_length = image_processor.merge_size**2 |
| 1487 | |
| 1488 | token_id = image_token_id if modality == "image" else video_token_id |
| 1489 | return [token_id] * (int(grid_thw.prod()) // merge_length) |
| 1490 | |
| 1491 | use_audio_in_video = hf_processor_mm_kwargs.get("use_audio_in_video", False) |
| 1492 | thinker_config = self.info.get_hf_config() |
| 1493 | |
| 1494 | def get_replacement_qwen2_use_audio_in_video(item_idx: int): |
| 1495 | nonlocal audio_in_video_item_idx |
| 1496 | audio_num_features = audio_output_lengths[audio_in_video_item_idx] |
| 1497 | video_grid_thw = out_mm_data["video_grid_thw"][item_idx] |
| 1498 | |
| 1499 | audio_in_video_item_idx += 1 |
| 1500 | |
| 1501 | second_per_grid_ts = hf_processor_mm_kwargs.get("second_per_grid_ts", None) |
| 1502 | if second_per_grid_ts: |
| 1503 | video_second_per_grid_t = second_per_grid_ts[item_idx] |
| 1504 | else: |
| 1505 | video_second_per_grid_t = 2.0 |
| 1506 | |
| 1507 | placeholder = self.get_updates_use_audio_in_video( |
| 1508 | thinker_config=thinker_config, |
| 1509 | audio_len=audio_num_features, |
| 1510 | video_grid_thw=video_grid_thw, |
| 1511 | video_second_per_grid_t=video_second_per_grid_t, |
| 1512 | ) |
| 1513 | return PromptUpdateDetails.select_token_id( |
| 1514 | placeholder, embed_token_id=video_token_id |
| 1515 | ) |
| 1516 | |
| 1517 | video_replacement_fn = ( |
| 1518 | get_replacement_qwen2_use_audio_in_video |
| 1519 | if use_audio_in_video |
| 1520 | else partial(get_replacement_qwen2_vision, modality="video") |
| 1521 | ) |
| 1522 | |
| 1523 | return [ |
| 1524 | PromptReplacement( |
| 1525 | modality="audio", |
| 1526 | target=[audio_token_id], |
| 1527 | replacement=get_replacement_qwen2_audio, |
| 1528 | ), |
| 1529 | PromptReplacement( |
| 1530 | modality="image", |
| 1531 | target=[image_token_id], |
| 1532 | replacement=partial(get_replacement_qwen2_vision, modality="image"), |
| 1533 | ), |
| 1534 | PromptReplacement( |
| 1535 | modality="video", |
| 1536 | target=[video_token_id], |
| 1537 | replacement=video_replacement_fn, |
| 1538 | ), |
| 1539 | ] |
| 1540 | |
| 1541 | def _derive_audio_from_video_placeholders( |
| 1542 | self, |
| 1543 | placeholders: Mapping[str, list[PlaceholderFeaturesInfo]], |
| 1544 | mm_prompt_updates: MultiModalPromptUpdates, |
| 1545 | ) -> Mapping[str, list[PlaceholderFeaturesInfo]]: |
| 1546 | """Helper to derive audio placeholders from video placeholders when |
| 1547 | use_audio_in_video=True. |
| 1548 | """ |
| 1549 | if "video" not in placeholders: |
| 1550 | return placeholders |
| 1551 | |
| 1552 | # Validate audio and video counts match |
| 1553 | num_videos = len(placeholders["video"]) |
| 1554 | num_audios = len(mm_prompt_updates.get("audio", [])) |
| 1555 | if num_audios != num_videos: |
| 1556 | raise ValueError( |
| 1557 | f"use_audio_in_video requires equal number of audio and video items, " |
| 1558 | f"got {num_audios=}, {num_videos=}" |
| 1559 | ) |
| 1560 | |
| 1561 | tokenizer = self.info.get_tokenizer() |
| 1562 | processor = self.info.get_hf_processor() |
| 1563 | audio_token_id = tokenizer.get_vocab()[processor.audio_token] |
| 1564 | |
| 1565 | result_placeholders = dict(placeholders) |
| 1566 | audio_placeholders = [] |
| 1567 | |
| 1568 | # Each video is paired with one audio |
| 1569 | for video_idx, video_placeholder in enumerate(placeholders["video"]): |
| 1570 | # Create is_embed mask selecting only audio tokens |
| 1571 | audio_is_embed = torch.tensor(video_placeholder.tokens) == audio_token_id |
| 1572 | |
| 1573 | audio_placeholder = PlaceholderFeaturesInfo( |
| 1574 | modality="audio", |
| 1575 | item_idx=video_idx, |
| 1576 | start_idx=video_placeholder.start_idx, |
| 1577 | tokens=video_placeholder.tokens, |
| 1578 | is_embed=audio_is_embed, |
| 1579 | ) |
| 1580 | audio_placeholders.append(audio_placeholder) |
| 1581 | |
| 1582 | result_placeholders["audio"] = audio_placeholders |
| 1583 | return result_placeholders |
| 1584 | |
| 1585 | |
| 1586 | class Qwen3OmniMoeConditionalGenerationMixin(Qwen2_5OmniConditionalGenerationMixin): |
| 1587 | def _process_audio_input( |
| 1588 | self, |
| 1589 | audio_input: Qwen2_5OmniAudioFeatureInputs, |
| 1590 | ) -> tuple[torch.Tensor, ...]: |
| 1591 | input_features = audio_input["input_features"] |
| 1592 | # audio_feature_lengths is keep_on_cpu; the audio tower derives |
| 1593 | # device placement from feature_lens, so move it explicitly. |
| 1594 | audio_feature_lengths = async_tensor_h2d( |
| 1595 | audio_input["audio_feature_lengths"], input_features.device |
| 1596 | ) |
| 1597 | |
| 1598 | audio_output_lengths = _get_feat_extract_output_lengths(audio_feature_lengths) |
| 1599 | |
| 1600 | audio_features = self.audio_tower( |
| 1601 | input_features.to(self.audio_tower.dtype), |
| 1602 | feature_lens=audio_feature_lengths, |
| 1603 | aftercnn_lens=audio_output_lengths, |
| 1604 | ) |
| 1605 | return audio_features.split(audio_output_lengths.tolist()) |
| 1606 | |
| 1607 | |
| 1608 | @MULTIMODAL_REGISTRY.register_processor( |
| 1609 | Qwen3OmniMoeThinkerMultiModalProcessor, |
| 1610 | info=Qwen3OmniMoeThinkerProcessingInfo, |
| 1611 | dummy_inputs=Qwen3OmniMoeThinkerDummyInputsBuilder, |
| 1612 | ) |
| 1613 | class Qwen3OmniMoeThinkerForConditionalGeneration( |
| 1614 | nn.Module, |
| 1615 | SupportsMultiModal, |
| 1616 | SupportsLoRA, |
| 1617 | SupportsPP, |
| 1618 | SupportsMRoPE, |
| 1619 | SupportsEagle3, |
| 1620 | Qwen3OmniMoeConditionalGenerationMixin, |
| 1621 | SupportsTranscription, |
| 1622 | ): |
| 1623 | is_3d_moe_weight: ClassVar[bool] = True |
| 1624 | supports_tower_connector_lora: ClassVar[bool] = True |
| 1625 | hf_to_vllm_mapper = WeightsMapper( |
| 1626 | orig_to_new_prefix={ |
| 1627 | "thinker.lm_head.": "language_model.lm_head.", |
| 1628 | "thinker.model.": "language_model.model.", |
| 1629 | "thinker.": "", |
| 1630 | "talker.": None, |
| 1631 | "code2wav.": None, |
| 1632 | # Adapter trained/saved directly from the thinker model |
| 1633 | "model.": "language_model.model.", |
| 1634 | "lm_head.": "language_model.lm_head.", |
| 1635 | } |
| 1636 | ) |
| 1637 | |
| 1638 | packed_modules_mapping = { |
| 1639 | "qkv_proj": [ |
| 1640 | "q_proj", |
| 1641 | "k_proj", |
| 1642 | "v_proj", |
| 1643 | ], |
| 1644 | "gate_up_proj": [ |
| 1645 | "gate_proj", |
| 1646 | "up_proj", |
| 1647 | ], |
| 1648 | "qkv": ["qkv"], # for visual encoder |
| 1649 | } |
| 1650 | |
| 1651 | supported_languages = ISO639_1_SUPPORTED_LANGS |
| 1652 | |
| 1653 | @classmethod |
| 1654 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 1655 | if modality.startswith("image"): |
| 1656 | return "<|vision_start|><|image_pad|><|vision_end|>" |
| 1657 | if modality.startswith("video"): |
| 1658 | return "<|vision_start|><|video_pad|><|vision_end|>" |
| 1659 | if modality.startswith("audio"): |
| 1660 | return "<|audio_start|><|audio_pad|><|audio_end|>" |
| 1661 | |
| 1662 | raise ValueError("Only image, video or audio modality is supported") |
| 1663 | |
| 1664 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1665 | super().__init__() |
| 1666 | self.vllm_config = vllm_config # needed for torch compile forward context |
| 1667 | thinker_config: Qwen3OmniMoeThinkerConfig = ( |
| 1668 | vllm_config.model_config.hf_config.thinker_config |
| 1669 | ) |
| 1670 | # MoE LoRA uses the architecture to determine its weight layout. |
| 1671 | thinker_config.architectures = vllm_config.model_config.hf_config.architectures |
| 1672 | quant_config = vllm_config.quant_config |
| 1673 | multimodal_config = vllm_config.model_config.multimodal_config |
| 1674 | self.config = thinker_config |
| 1675 | self.multimodal_config = multimodal_config |
| 1676 | self.quant_config = quant_config |
| 1677 | |
| 1678 | with self._mark_tower_model(vllm_config, "audio"): |
| 1679 | self.audio_tower = Qwen3OmniMoeAudioEncoder( |
| 1680 | thinker_config.audio_config, |
| 1681 | prefix=maybe_prefix(prefix, "audio_tower"), |
| 1682 | ) |
| 1683 | |
| 1684 | with self._mark_tower_model(vllm_config, {"image", "video"}): |
| 1685 | self.visual = Qwen3Omni_VisionTransformer( |
| 1686 | vision_config=thinker_config.vision_config, |
| 1687 | norm_eps=getattr(thinker_config.text_config, "rms_norm_eps", 1e-6), |
| 1688 | quant_config=quant_config, |
| 1689 | prefix=maybe_prefix(prefix, "visual"), |
| 1690 | ) |
| 1691 | |
| 1692 | self.use_deepstack = hasattr( |
| 1693 | thinker_config.vision_config, "deepstack_visual_indexes" |
| 1694 | ) and not isinstance(self.visual, StageMissingLayer) |
| 1695 | self.deepstack_num_level = ( |
| 1696 | len(thinker_config.vision_config.deepstack_visual_indexes) |
| 1697 | if self.use_deepstack |
| 1698 | else 0 |
| 1699 | ) |
| 1700 | self.visual_dim = thinker_config.vision_config.out_hidden_size |
| 1701 | self.multiscale_dim = self.visual_dim * self.deepstack_num_level |
| 1702 | |
| 1703 | if self.use_deepstack: |
| 1704 | self.deepstack_input_embeds = [ |
| 1705 | torch.zeros( |
| 1706 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 1707 | thinker_config.text_config.hidden_size, |
| 1708 | ) |
| 1709 | for _ in range(self.deepstack_num_level) |
| 1710 | ] |
| 1711 | # Tracks the valid token span currently stored in the buffer. |
| 1712 | # Zero means there is no active deepstack payload to consume. |
| 1713 | self.deepstack_input_embeds_num_tokens = 0 |
| 1714 | |
| 1715 | with self._mark_language_model(vllm_config): |
| 1716 | self.language_model = Qwen3MoeLLMForCausalLM( |
| 1717 | vllm_config=vllm_config.with_hf_config( |
| 1718 | thinker_config.text_config, |
| 1719 | architectures=["Qwen3MoeForCausalLM"], |
| 1720 | ), |
| 1721 | prefix=maybe_prefix(prefix, "language_model"), |
| 1722 | ) |
| 1723 | |
| 1724 | self.make_empty_intermediate_tensors = ( |
| 1725 | self.language_model.make_empty_intermediate_tensors |
| 1726 | ) |
| 1727 | |
| 1728 | def _get_deepstack_input_embeds( |
| 1729 | self, |
| 1730 | num_tokens: int, |
| 1731 | ) -> IntermediateTensors | None: |
| 1732 | if not getattr(self, "deepstack_input_embeds", None): |
| 1733 | return None # If vision tower is skipped |
| 1734 | if num_tokens > self.deepstack_input_embeds[0].size(0): |
| 1735 | self._resize_deepstack_input_embeds(num_tokens) |
| 1736 | |
| 1737 | # get deepstack_input_embeds from buffer, and clear the buffer |
| 1738 | return IntermediateTensors( |
| 1739 | { |
| 1740 | f"deepstack_input_embeds_{idx}": self.deepstack_input_embeds[idx][ |
| 1741 | :num_tokens |
| 1742 | ] |
| 1743 | for idx in range(self.deepstack_num_level) |
| 1744 | } |
| 1745 | ) |
| 1746 | |
| 1747 | def _resize_deepstack_input_embeds(self, num_tokens: int) -> None: |
| 1748 | self.deepstack_input_embeds = [ |
| 1749 | torch.zeros( |
| 1750 | num_tokens, |
| 1751 | self.config.text_config.hidden_size, |
| 1752 | device=self.deepstack_input_embeds[0].device, |
| 1753 | dtype=self.deepstack_input_embeds[0].dtype, |
| 1754 | ) |
| 1755 | for _ in range(self.deepstack_num_level) |
| 1756 | ] |
| 1757 | |
| 1758 | def _set_deepstack_input_embeds(self, deepstack_input_embeds: torch.Tensor) -> None: |
| 1759 | if not getattr(self, "deepstack_input_embeds", None): |
| 1760 | return |
| 1761 | |
| 1762 | # set deepstack_input_embeds to buffer |
| 1763 | num_tokens = deepstack_input_embeds.size(1) |
| 1764 | if num_tokens > self.deepstack_input_embeds[0].size(0): |
| 1765 | self._resize_deepstack_input_embeds(num_tokens) |
| 1766 | for idx in range(self.deepstack_num_level): |
| 1767 | self.deepstack_input_embeds[idx][:num_tokens].copy_( |
| 1768 | deepstack_input_embeds[idx] |
| 1769 | ) |
| 1770 | self.deepstack_input_embeds_num_tokens = num_tokens |
| 1771 | |
| 1772 | def _clear_deepstack_input_embeds(self, num_tokens: int) -> None: |
| 1773 | if not getattr(self, "deepstack_input_embeds", None): |
| 1774 | return |
| 1775 | if getattr(self, "deepstack_input_embeds_num_tokens", 0) == 0: |
| 1776 | return |
| 1777 | |
| 1778 | # clear deepstack_input_embeds in buffer |
| 1779 | if num_tokens > 0: |
| 1780 | for idx in range(self.deepstack_num_level): |
| 1781 | self.deepstack_input_embeds[idx][:num_tokens].zero_() |
| 1782 | self.deepstack_input_embeds_num_tokens = 0 |
| 1783 | |
| 1784 | def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict: |
| 1785 | mm_input_by_modality: dict[ |
| 1786 | str, |
| 1787 | Qwen2_5_VLImageInputs |
| 1788 | | Qwen2_5_VLVideoInputs |
| 1789 | | Qwen2_5OmniAudioFeatureInputs |
| 1790 | | None, |
| 1791 | ] = {} |
| 1792 | |
| 1793 | # Preserve the order of modalities if there are multiple of them |
| 1794 | # from the order of kwargs. |
| 1795 | for input_key in kwargs: |
| 1796 | if ( |
| 1797 | input_key in ("pixel_values", "image_embeds") |
| 1798 | and "image" not in mm_input_by_modality |
| 1799 | ): |
| 1800 | mm_input_by_modality["image"] = self._parse_and_validate_image_input( |
| 1801 | **kwargs |
| 1802 | ) |
| 1803 | if ( |
| 1804 | input_key in ("pixel_values_videos", "video_embeds") |
| 1805 | and "video" not in mm_input_by_modality |
| 1806 | ): |
| 1807 | mm_input_by_modality["video"] = self._parse_and_validate_video_input( |
| 1808 | **kwargs |
| 1809 | ) |
| 1810 | if ( |
| 1811 | input_key in ("input_audio_features") |
| 1812 | and "audio" not in mm_input_by_modality |
| 1813 | ): |
| 1814 | mm_input_by_modality["audio"] = self._parse_and_validate_audio_input( |
| 1815 | **kwargs |
| 1816 | ) |
| 1817 | return mm_input_by_modality |
| 1818 | |
| 1819 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None: |
| 1820 | mm_input_by_modality = self._parse_and_validate_multimodal_inputs(**kwargs) |
| 1821 | if not mm_input_by_modality: |
| 1822 | return [] |
| 1823 | |
| 1824 | # The result multimodal_embeddings is tuple of tensors, with each |
| 1825 | # tensor corresponding to a multimodal data item (image or video). |
| 1826 | multimodal_embeddings: tuple[torch.Tensor, ...] = () |
| 1827 | |
| 1828 | # NOTE: It is important to iterate over the keys in this dictionary |
| 1829 | # to preserve the order of the modalities. |
| 1830 | for modality in mm_input_by_modality: |
| 1831 | multimodal_input = mm_input_by_modality[modality] |
| 1832 | if modality == "image": |
| 1833 | image_embeddings = self._process_image_input(multimodal_input) |
| 1834 | multimodal_embeddings += tuple(image_embeddings) |
| 1835 | if modality == "video": |
| 1836 | video_embeddings = self._process_video_input(multimodal_input) |
| 1837 | multimodal_embeddings += tuple(video_embeddings) |
| 1838 | if modality == "audio": |
| 1839 | audio_embeddings = self._process_audio_input(multimodal_input) |
| 1840 | multimodal_embeddings += tuple(audio_embeddings) |
| 1841 | return multimodal_embeddings |
| 1842 | |
| 1843 | def embed_input_ids( |
| 1844 | self, |
| 1845 | input_ids: torch.Tensor, |
| 1846 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 1847 | *, |
| 1848 | is_multimodal: torch.Tensor | None = None, |
| 1849 | ) -> torch.Tensor: |
| 1850 | inputs_embeds = self._embed_text_input_ids( |
| 1851 | input_ids, |
| 1852 | self.language_model.embed_input_ids, |
| 1853 | is_multimodal=is_multimodal, |
| 1854 | ) |
| 1855 | |
| 1856 | if multimodal_embeddings is None or len(multimodal_embeddings) == 0: |
| 1857 | return inputs_embeds |
| 1858 | |
| 1859 | # Detect interleaved audio-in-video early, since it affects |
| 1860 | # both the deepstack path and the final embedding merge. |
| 1861 | video_token_id = self.config.video_token_id |
| 1862 | audio_token_id = self.config.audio_token_id |
| 1863 | image_token_id = self.config.image_token_id |
| 1864 | input_ids_cpu = input_ids.cpu() |
| 1865 | is_video = is_multimodal & (input_ids_cpu == video_token_id) |
| 1866 | is_audio = is_multimodal & (input_ids_cpu == audio_token_id) |
| 1867 | is_image = is_multimodal & (input_ids_cpu == image_token_id) |
| 1868 | num_video = is_video.sum().item() |
| 1869 | num_audio = is_audio.sum().item() |
| 1870 | |
| 1871 | is_interleaved = check_interleaved_audio_video( |
| 1872 | is_video, is_audio, num_video, num_audio |
| 1873 | ) |
| 1874 | |
| 1875 | deepstack_input_embeds = None |
| 1876 | # split the feat dim to obtain multi-scale visual feature |
| 1877 | has_vision_embeddings = [ |
| 1878 | embeddings.shape[-1] != self.config.text_config.hidden_size |
| 1879 | for embeddings in multimodal_embeddings |
| 1880 | ] |
| 1881 | if self.visual.deepstack_visual_indexes is not None and any( |
| 1882 | has_vision_embeddings |
| 1883 | ): |
| 1884 | multimodal_embeddings = list(multimodal_embeddings) |
| 1885 | multiscale_len = len(self.visual.deepstack_visual_indexes) |
| 1886 | multimodal_embeddings_multiscale = [] |
| 1887 | |
| 1888 | if is_interleaved: |
| 1889 | # Use input_ids-based mask for all vision positions when |
| 1890 | # audio and video tokens are interleaved. |
| 1891 | is_vision = is_video | is_image |
| 1892 | else: |
| 1893 | is_vision = torch.zeros_like(is_multimodal) |
| 1894 | mm_positions = torch.nonzero(is_multimodal, as_tuple=True)[0] |
| 1895 | mm_position_idx = 0 |
| 1896 | |
| 1897 | for index, embeddings in enumerate(multimodal_embeddings): |
| 1898 | num_tokens = embeddings.shape[0] |
| 1899 | |
| 1900 | # Vision embeddings |
| 1901 | if embeddings.shape[-1] != self.config.text_config.hidden_size: |
| 1902 | visual_dim = embeddings.shape[-1] // (multiscale_len + 1) |
| 1903 | multi_dim = visual_dim * multiscale_len |
| 1904 | embeddings_main, embeddings_multiscale = torch.split( |
| 1905 | embeddings, [visual_dim, multi_dim], dim=-1 |
| 1906 | ) |
| 1907 | multimodal_embeddings[index] = embeddings_main |
| 1908 | multimodal_embeddings_multiscale.append(embeddings_multiscale) |
| 1909 | if not is_interleaved: |
| 1910 | current_positions = mm_positions[ |
| 1911 | mm_position_idx : mm_position_idx + num_tokens |
| 1912 | ] |
| 1913 | is_vision[current_positions] = True |
| 1914 | |
| 1915 | # Audio embeddings |
| 1916 | else: |
| 1917 | if not is_interleaved: |
| 1918 | current_positions = mm_positions[ |
| 1919 | mm_position_idx : mm_position_idx + num_tokens |
| 1920 | ] |
| 1921 | is_vision[current_positions] = False |
| 1922 | |
| 1923 | if not is_interleaved: |
| 1924 | mm_position_idx += num_tokens |
| 1925 | |
| 1926 | deepstack_input_embeds = inputs_embeds.new_zeros( |
| 1927 | inputs_embeds.size(0), multiscale_len * inputs_embeds.size(1) |
| 1928 | ) |
| 1929 | deepstack_input_embeds = _merge_multimodal_embeddings( |
| 1930 | inputs_embeds=deepstack_input_embeds, |
| 1931 | multimodal_embeddings=multimodal_embeddings_multiscale, |
| 1932 | is_multimodal=is_vision, |
| 1933 | ) |
| 1934 | deepstack_input_embeds = ( |
| 1935 | deepstack_input_embeds.view( |
| 1936 | inputs_embeds.shape[0], multiscale_len, visual_dim |
| 1937 | ) |
| 1938 | .permute(1, 0, 2) |
| 1939 | .contiguous() |
| 1940 | ) |
| 1941 | self._set_deepstack_input_embeds(deepstack_input_embeds) |
| 1942 | |
| 1943 | if is_interleaved: |
| 1944 | return merge_interleaved_embeddings( |
| 1945 | inputs_embeds, |
| 1946 | multimodal_embeddings, |
| 1947 | is_video, |
| 1948 | is_audio, |
| 1949 | is_multimodal, |
| 1950 | ) |
| 1951 | |
| 1952 | # Default: standard merge (no interleaving), same as parent class. |
| 1953 | # multimodal_embeddings may have been updated above (deepstack |
| 1954 | # main-scale). Use super() to stay consistent with the parent |
| 1955 | # implementation and avoid issues seen in Qwen2.5-Omni (#34506). |
| 1956 | return super().embed_input_ids( |
| 1957 | input_ids, |
| 1958 | multimodal_embeddings=multimodal_embeddings, |
| 1959 | is_multimodal=is_multimodal, |
| 1960 | ) |
| 1961 | |
| 1962 | def forward( |
| 1963 | self, |
| 1964 | input_ids: torch.Tensor | None, |
| 1965 | positions: torch.Tensor, |
| 1966 | intermediate_tensors: IntermediateTensors | None = None, |
| 1967 | inputs_embeds: torch.Tensor | None = None, |
| 1968 | **kwargs: object, |
| 1969 | ) -> torch.Tensor | IntermediateTensors: |
| 1970 | if intermediate_tensors is not None: |
| 1971 | inputs_embeds = None |
| 1972 | |
| 1973 | if inputs_embeds is not None and get_pp_group().is_first_rank: |
| 1974 | deepstack_input_embeds = self._get_deepstack_input_embeds( |
| 1975 | inputs_embeds.size(0) |
| 1976 | ) |
| 1977 | else: |
| 1978 | deepstack_input_embeds = None |
| 1979 | |
| 1980 | hidden_states = self.language_model.model( |
| 1981 | input_ids, |
| 1982 | positions, |
| 1983 | intermediate_tensors, |
| 1984 | inputs_embeds=inputs_embeds, |
| 1985 | # args for deepstack |
| 1986 | deepstack_input_embeds=deepstack_input_embeds, |
| 1987 | ) |
| 1988 | |
| 1989 | if inputs_embeds is not None and get_pp_group().is_first_rank: |
| 1990 | self._clear_deepstack_input_embeds(inputs_embeds.size(0)) |
| 1991 | |
| 1992 | return hidden_states |
| 1993 | |
| 1994 | def compute_logits( |
| 1995 | self, |
| 1996 | hidden_states: torch.Tensor, |
| 1997 | ) -> torch.Tensor | None: |
| 1998 | return self.language_model.compute_logits(hidden_states) |
| 1999 | |
| 2000 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 2001 | loader = AutoWeightsLoader(self) |
| 2002 | loaded_weights = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 2003 | |
| 2004 | return loaded_weights |
| 2005 | |
| 2006 | def _compute_audio_token_count(self, audio_feature_length: int) -> int: |
| 2007 | """Compute audio tokens from feature length using Qwen3-Omni formula.""" |
| 2008 | return _get_feat_extract_output_lengths( |
| 2009 | torch.tensor([audio_feature_length]) |
| 2010 | ).item() |
| 2011 | |
| 2012 | def _get_audio_for_video_mapping( |
| 2013 | self, mm_features: list[MultiModalFeatureSpec] |
| 2014 | ) -> tuple[dict[int, int], set[int]]: |
| 2015 | """Map video offset -> paired audio_feature_length for use_audio_in_video. |
| 2016 | |
| 2017 | When use_audio_in_video=True, audio is interleaved within video. |
| 2018 | The pairing is based on feature order in mm_features. |
| 2019 | |
| 2020 | Returns: |
| 2021 | Tuple of (video_offset -> audio_feature_length mapping, |
| 2022 | set of paired audio offsets to skip) |
| 2023 | |
| 2024 | """ |
| 2025 | videos_with_audio = [ |
| 2026 | f |
| 2027 | for f in mm_features |
| 2028 | if f.modality == "video" |
| 2029 | and f.data is not None |
| 2030 | and f.data.get("use_audio_in_video") |
| 2031 | and _get_feature_tensor(f, "use_audio_in_video").item() |
| 2032 | ] |
| 2033 | audios = [f for f in mm_features if f.modality == "audio"] |
| 2034 | |
| 2035 | mapping: dict[int, int] = {} |
| 2036 | paired_audio_offsets: set[int] = set() |
| 2037 | for i, video_f in enumerate(videos_with_audio): |
| 2038 | if i < len(audios): |
| 2039 | audio_len = int( |
| 2040 | _get_feature_tensor(audios[i], "audio_feature_lengths").item() |
| 2041 | ) |
| 2042 | mapping[video_f.mm_position.offset] = audio_len |
| 2043 | paired_audio_offsets.add(audios[i].mm_position.offset) |
| 2044 | return mapping, paired_audio_offsets |
| 2045 | |
| 2046 | def iter_mm_features( |
| 2047 | self, mm_features: list[MultiModalFeatureSpec] |
| 2048 | ) -> Iterator[tuple[int, str, dict[str, Any]]]: |
| 2049 | """Iterate over multimodal features sorted by position offset. |
| 2050 | |
| 2051 | Yields: (offset, modality, feature_data) where feature_data contains: |
| 2052 | - image: {"grid_t", "grid_h", "grid_w", "t_factor"} |
| 2053 | - video: {"grid_t", "grid_h", "grid_w", "t_factor", |
| 2054 | "use_audio_in_video", "audio_feature_length"} |
| 2055 | - audio: {"audio_feature_length"} |
| 2056 | """ |
| 2057 | config = self.config |
| 2058 | spatial_merge_size = config.vision_config.spatial_merge_size |
| 2059 | position_id_per_seconds = config.position_id_per_seconds |
| 2060 | |
| 2061 | sorted_features = sorted(mm_features, key=lambda f: f.mm_position.offset) |
| 2062 | audio_for_video, paired_audio_offsets = self._get_audio_for_video_mapping( |
| 2063 | sorted_features |
| 2064 | ) |
| 2065 | |
| 2066 | for mm_feature in sorted_features: |
| 2067 | offset = mm_feature.mm_position.offset |
| 2068 | modality = mm_feature.modality |
| 2069 | assert mm_feature.data is not None |
| 2070 | |
| 2071 | if modality == "image": |
| 2072 | t, h, w = _get_feature_tensor(mm_feature, "image_grid_thw").tolist() |
| 2073 | yield ( |
| 2074 | offset, |
| 2075 | "image", |
| 2076 | { |
| 2077 | "grid_t": t, |
| 2078 | "grid_h": h // spatial_merge_size, |
| 2079 | "grid_w": w // spatial_merge_size, |
| 2080 | "t_factor": position_id_per_seconds, |
| 2081 | }, |
| 2082 | ) |
| 2083 | elif modality == "video": |
| 2084 | t, h, w = _get_feature_tensor(mm_feature, "video_grid_thw").tolist() |
| 2085 | second_per_grid_ts = 2.0 |
| 2086 | if mm_feature.data.get("second_per_grid_ts"): |
| 2087 | second_per_grid_ts = _get_feature_tensor( |
| 2088 | mm_feature, "second_per_grid_ts" |
| 2089 | ).item() |
| 2090 | use_audio_in_video = bool( |
| 2091 | mm_feature.data.get("use_audio_in_video") |
| 2092 | and _get_feature_tensor(mm_feature, "use_audio_in_video").item() |
| 2093 | ) |
| 2094 | |
| 2095 | yield ( |
| 2096 | offset, |
| 2097 | "video", |
| 2098 | { |
| 2099 | "grid_t": t, |
| 2100 | "grid_h": h // spatial_merge_size, |
| 2101 | "grid_w": w // spatial_merge_size, |
| 2102 | "t_factor": second_per_grid_ts * position_id_per_seconds, |
| 2103 | "use_audio_in_video": use_audio_in_video, |
| 2104 | "audio_feature_length": audio_for_video.get(offset), |
| 2105 | }, |
| 2106 | ) |
| 2107 | elif modality == "audio": |
| 2108 | if offset not in paired_audio_offsets: |
| 2109 | audio_len = _get_feature_tensor( |
| 2110 | mm_feature, "audio_feature_lengths" |
| 2111 | ).item() |
| 2112 | yield offset, "audio", {"audio_feature_length": audio_len} |
| 2113 | |
| 2114 | def _compute_interleaved_positions( |
| 2115 | self, start_idx: int, data: dict[str, Any] |
| 2116 | ) -> tuple[np.ndarray, int]: |
| 2117 | """Compute positions for interleaved video+audio using Qwen3 token-by-token |
| 2118 | interleaving logic. |
| 2119 | |
| 2120 | Returns: (position_ids [3, N], total_token_count) |
| 2121 | """ |
| 2122 | grid_t = data["grid_t"] |
| 2123 | grid_h = data["grid_h"] |
| 2124 | grid_w = data["grid_w"] |
| 2125 | t_factor = data["t_factor"] |
| 2126 | audio_feature_length = data["audio_feature_length"] |
| 2127 | |
| 2128 | audio_len = self._compute_audio_token_count(audio_feature_length) |
| 2129 | |
| 2130 | h_index = np.tile( |
| 2131 | np.arange(grid_h).reshape(1, -1, 1), (grid_t, 1, grid_w) |
| 2132 | ).flatten() |
| 2133 | w_index = np.tile( |
| 2134 | np.arange(grid_w).reshape(1, 1, -1), (grid_t, grid_h, 1) |
| 2135 | ).flatten() |
| 2136 | t_index_raw = np.arange(grid_t) |
| 2137 | t_index_scaled = (t_index_raw * t_factor).astype(np.int64) |
| 2138 | t_index = np.repeat(t_index_scaled, grid_h * grid_w) |
| 2139 | |
| 2140 | video_pos = np.stack([t_index, h_index, w_index]) + start_idx |
| 2141 | audio_pos = np.broadcast_to(np.arange(audio_len), (3, audio_len)) + start_idx |
| 2142 | |
| 2143 | video_t_values = video_pos[0] |
| 2144 | audio_t_values = audio_pos[0] |
| 2145 | |
| 2146 | pos_ids_list: list[np.ndarray] = [] |
| 2147 | video_idx, audio_idx = 0, 0 |
| 2148 | num_video = grid_t * grid_h * grid_w |
| 2149 | |
| 2150 | while video_idx < num_video and audio_idx < audio_len: |
| 2151 | if video_t_values[video_idx] <= audio_t_values[audio_idx]: |
| 2152 | pos_ids_list.append(video_pos[:, video_idx : video_idx + 1]) |
| 2153 | video_idx += 1 |
| 2154 | else: |
| 2155 | pos_ids_list.append(audio_pos[:, audio_idx : audio_idx + 1]) |
| 2156 | audio_idx += 1 |
| 2157 | |
| 2158 | if video_idx < num_video: |
| 2159 | pos_ids_list.append(video_pos[:, video_idx:]) |
| 2160 | if audio_idx < audio_len: |
| 2161 | pos_ids_list.append(audio_pos[:, audio_idx:]) |
| 2162 | |
| 2163 | total_tokens = num_video + audio_len |
| 2164 | return np.concatenate(pos_ids_list, axis=1), total_tokens |
| 2165 | |
| 2166 | @classmethod |
| 2167 | def get_speech_to_text_config( |
| 2168 | cls, model_config: ModelConfig, task_type: str |
| 2169 | ) -> SpeechToTextConfig: |
| 2170 | processor = cached_processor_from_config( |
| 2171 | model_config, processor_cls=Qwen3OmniMoeProcessor |
| 2172 | ) |
| 2173 | return SpeechToTextConfig( |
| 2174 | max_audio_clip_s=processor.feature_extractor.chunk_length, |
| 2175 | sample_rate=processor.feature_extractor.sampling_rate, |
| 2176 | min_energy_split_window_size=None, |
| 2177 | ) |
| 2178 | |
| 2179 | @classmethod |
| 2180 | def get_generation_prompt(cls, stt_params: SpeechToTextParams) -> PromptType: |
| 2181 | """Construct a transcription/translation prompt for Qwen3-Omni.""" |
| 2182 | audio = stt_params.audio |
| 2183 | stt_config = stt_params.stt_config |
| 2184 | model_config = stt_params.model_config |
| 2185 | language = stt_params.language |
| 2186 | task_type = stt_params.task_type |
| 2187 | to_language = stt_params.to_language |
| 2188 | request_prompt = stt_params.request_prompt |
| 2189 | # Transcribe this audio [into <language>] | for transcription |
| 2190 | # Translate this audio [from <language> into <to_language>] | for translation |
| 2191 | instruction = "Transcribe" if task_type == "transcribe" else "Translate" |
| 2192 | instruction += " this audio" |
| 2193 | |
| 2194 | # Default to_language to English for translation |
| 2195 | if task_type == "translate" and to_language is None: |
| 2196 | to_language = "en" |
| 2197 | |
| 2198 | # Get full language names from supported_languages mapping |
| 2199 | full_lang_name = ( |
| 2200 | cls.supported_languages.get(language, "") if language is not None else "" |
| 2201 | ) |
| 2202 | full_lang_name_to = ( |
| 2203 | cls.supported_languages.get(to_language, "") |
| 2204 | if to_language is not None |
| 2205 | else "" |
| 2206 | ) |
| 2207 | |
| 2208 | if task_type == "transcribe" and full_lang_name: |
| 2209 | instruction += f" into {full_lang_name}" |
| 2210 | elif task_type == "translate": |
| 2211 | if full_lang_name: |
| 2212 | instruction += f" from {full_lang_name}" |
| 2213 | if full_lang_name_to: |
| 2214 | instruction += f" into {full_lang_name_to}" |
| 2215 | |
| 2216 | instruction += "." |
| 2217 | |
| 2218 | if request_prompt: |
| 2219 | instruction += f" {request_prompt}" |
| 2220 | |
| 2221 | processor = cached_processor_from_config( |
| 2222 | model_config, processor_cls=Qwen3OmniMoeProcessor |
| 2223 | ) |
| 2224 | # Audio placeholder format: <|audio_start|><|audio_pad|><|audio_end|> |
| 2225 | audio_placeholder = "<|audio_start|><|audio_pad|><|audio_end|>" |
| 2226 | user_content = f"{audio_placeholder}{instruction}" |
| 2227 | |
| 2228 | messages = [{"role": "user", "content": user_content}] |
| 2229 | prompt = processor.apply_chat_template( |
| 2230 | messages, |
| 2231 | tokenize=False, |
| 2232 | add_generation_prompt=True, |
| 2233 | ) |
| 2234 | |
| 2235 | audio_data = (audio, stt_config.sample_rate) |
| 2236 | prompts_dict: TextPrompt = { |
| 2237 | "multi_modal_data": {"audio": audio_data}, |
| 2238 | "prompt": prompt, |
| 2239 | } |
| 2240 | return prompts_dict |
| 2241 | |
| 2242 | def get_mrope_input_positions( |
| 2243 | self, |
| 2244 | input_tokens: list[int], |
| 2245 | mm_features: list[MultiModalFeatureSpec], |
| 2246 | ) -> tuple[torch.Tensor, int]: |
| 2247 | """Compute M-RoPE input positions using mm_features directly.""" |
| 2248 | seq_len = len(input_tokens) |
| 2249 | |
| 2250 | llm_pos_ids_list: list[np.ndarray] = [] |
| 2251 | st = 0 |
| 2252 | |
| 2253 | for offset, modality, data in self.iter_mm_features(mm_features): |
| 2254 | text_len = offset - st |
| 2255 | st_idx = int(llm_pos_ids_list[-1].max()) + 1 if llm_pos_ids_list else 0 |
| 2256 | |
| 2257 | if text_len > 0: |
| 2258 | llm_pos_ids_list.append( |
| 2259 | np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx |
| 2260 | ) |
| 2261 | st_idx += text_len |
| 2262 | |
| 2263 | bos_pos = np.broadcast_to(np.array([st_idx]), (3, 1)) |
| 2264 | llm_pos_ids_list.append(bos_pos) |
| 2265 | st_idx += 1 |
| 2266 | |
| 2267 | if modality == "audio": |
| 2268 | audio_tokens = self._compute_audio_token_count( |
| 2269 | data["audio_feature_length"] |
| 2270 | ) |
| 2271 | audio_pos = ( |
| 2272 | np.broadcast_to(np.arange(audio_tokens), (3, audio_tokens)) + st_idx |
| 2273 | ) |
| 2274 | llm_pos_ids_list.append(audio_pos) |
| 2275 | st_idx = int(audio_pos.max()) + 1 |
| 2276 | |
| 2277 | eos_pos = np.broadcast_to(np.array([st_idx]), (3, 1)) |
| 2278 | llm_pos_ids_list.append(eos_pos) |
| 2279 | st = offset + 1 + audio_tokens + 1 |
| 2280 | |
| 2281 | elif modality == "image": |
| 2282 | grid_t = data["grid_t"] |
| 2283 | grid_h = data["grid_h"] |
| 2284 | grid_w = data["grid_w"] |
| 2285 | t_factor = data["t_factor"] |
| 2286 | |
| 2287 | grid_indices = np.indices((grid_t, grid_h, grid_w)) |
| 2288 | if t_factor != 1.0: |
| 2289 | grid_indices[0] = (grid_indices[0] * t_factor).astype(np.int64) |
| 2290 | llm_pos_ids_list.append(grid_indices.reshape(3, -1) + st_idx) |
| 2291 | |
| 2292 | image_len = grid_t * grid_h * grid_w |
| 2293 | st_idx = int(llm_pos_ids_list[-1].max()) + 1 |
| 2294 | |
| 2295 | eos_pos = np.broadcast_to(np.array([st_idx]), (3, 1)) |
| 2296 | llm_pos_ids_list.append(eos_pos) |
| 2297 | st = offset + 1 + image_len + 1 |
| 2298 | |
| 2299 | elif modality == "video": |
| 2300 | grid_t = data["grid_t"] |
| 2301 | grid_h = data["grid_h"] |
| 2302 | grid_w = data["grid_w"] |
| 2303 | t_factor = data["t_factor"] |
| 2304 | |
| 2305 | if not data["use_audio_in_video"]: |
| 2306 | grid_indices = np.indices((grid_t, grid_h, grid_w)) |
| 2307 | if t_factor != 1.0: |
| 2308 | grid_indices[0] = (grid_indices[0] * t_factor).astype(np.int64) |
| 2309 | llm_pos_ids_list.append(grid_indices.reshape(3, -1) + st_idx) |
| 2310 | |
| 2311 | video_len = grid_t * grid_h * grid_w |
| 2312 | st_idx = int(llm_pos_ids_list[-1].max()) + 1 |
| 2313 | |
| 2314 | eos_pos = np.broadcast_to(np.array([st_idx]), (3, 1)) |
| 2315 | llm_pos_ids_list.append(eos_pos) |
| 2316 | st = offset + 1 + video_len + 1 |
| 2317 | else: |
| 2318 | audio_bos_pos = np.broadcast_to(np.array([st_idx - 1]), (3, 1)) |
| 2319 | llm_pos_ids_list.append(audio_bos_pos) |
| 2320 | |
| 2321 | pos_ids, _ = self._compute_interleaved_positions(st_idx, data) |
| 2322 | llm_pos_ids_list.append(pos_ids) |
| 2323 | st_idx = int(pos_ids.max()) + 1 |
| 2324 | |
| 2325 | eos_pos = np.broadcast_to(np.array([st_idx]), (3, 1)) |
| 2326 | llm_pos_ids_list.append(eos_pos) |
| 2327 | llm_pos_ids_list.append(eos_pos) |
| 2328 | |
| 2329 | video_len = grid_t * grid_h * grid_w |
| 2330 | audio_len = self._compute_audio_token_count( |
| 2331 | data["audio_feature_length"] |
| 2332 | ) |
| 2333 | st = offset + 2 + video_len + audio_len + 2 |
| 2334 | |
| 2335 | if st < seq_len: |
| 2336 | st_idx = int(llm_pos_ids_list[-1].max()) + 1 if llm_pos_ids_list else 0 |
| 2337 | text_len = seq_len - st |
| 2338 | llm_pos_ids_list.append( |
| 2339 | np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx |
| 2340 | ) |
| 2341 | |
| 2342 | llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1) |
| 2343 | if llm_positions.shape[1] != seq_len: |
| 2344 | raise RuntimeError("Position ids length mismatch with input ids length") |
| 2345 | |
| 2346 | mrope_position_delta = int(llm_positions.max()) + 1 - seq_len |
| 2347 | return torch.from_numpy(llm_positions), mrope_position_delta |
| 2348 | |
| 2349 | def get_mm_mapping(self) -> MultiModelKeys: |
| 2350 | """Get the module prefix in multimodal models""" |
| 2351 | return MultiModelKeys.from_string_field( |
| 2352 | language_model="language_model", |
| 2353 | connector="visual.merger", |
| 2354 | tower_model=["visual.", "audio_tower."], |
| 2355 | ) |
| 2356 | |
| 2357 | def get_mm_lora_token_counts( |
| 2358 | self, |
| 2359 | *, |
| 2360 | modality: str, |
| 2361 | mm_kwargs: MultiModalKwargsItem | None, |
| 2362 | num_mm_embeds: int, |
| 2363 | ) -> tuple[int, int | None]: |
| 2364 | if modality in ("image", "video"): |
| 2365 | hf_config = self.config |
| 2366 | vision_config = hf_config.vision_config |
| 2367 | merge_size = vision_config.spatial_merge_size |
| 2368 | tower_tokens = num_mm_embeds * merge_size**2 |
| 2369 | connector_tokens = num_mm_embeds |
| 2370 | |
| 2371 | if modality == "audio": |
| 2372 | tower_tokens = num_mm_embeds |
| 2373 | connector_tokens = num_mm_embeds |
| 2374 | |
| 2375 | return tower_tokens, connector_tokens |