| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from contextlib import nullcontext |
| 5 | from types import SimpleNamespace |
| 6 | from unittest.mock import Mock, patch |
| 7 | |
| 8 | import pytest |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | from transformers import PreTrainedConfig |
| 12 | |
| 13 | from vllm.config import ParallelConfig |
| 14 | from vllm.multimodal.processing import InputProcessingContext |
| 15 | |
| 16 | |
| 17 | # Helper function to print input IDs with coalesced audio/video tokens. |
| 18 | def print_input_ids(input_ids): |
| 19 | """Print input IDs, compressing consecutive special tokens. |
| 20 | - 151675: <|audio_pad|> |
| 21 | - 151656: <|video_pad|> |
| 22 | """ |
| 23 | if not input_ids: |
| 24 | print("[]") |
| 25 | return |
| 26 | |
| 27 | result = [] |
| 28 | i = 0 |
| 29 | |
| 30 | while i < len(input_ids): |
| 31 | current_id = input_ids[i] |
| 32 | |
| 33 | # Check if it's a special token that should be compressed |
| 34 | if current_id in [151675, 151656]: |
| 35 | # Count consecutive occurrences |
| 36 | count = 1 |
| 37 | while i + count < len(input_ids) and input_ids[i + count] == current_id: |
| 38 | count += 1 |
| 39 | |
| 40 | # Add compressed representation |
| 41 | token_name = "<|audio_pad|>" if current_id == 151675 else "<|video_pad|>" |
| 42 | result.append(f"{token_name} * {count}") |
| 43 | i += count |
| 44 | else: |
| 45 | # Regular token, just add it |
| 46 | result.append(str(current_id)) |
| 47 | i += 1 |
| 48 | |
| 49 | print(", ".join(result)) |
| 50 | |
| 51 | |
| 52 | @pytest.fixture |
| 53 | def mock_qwen3_omni_config(): |
| 54 | """Create a mock Qwen3OmniMoeThinker config.""" |
| 55 | config = Mock(spec=PreTrainedConfig) |
| 56 | # Token IDs from https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/blob/main/tokenizer_config.json |
| 57 | config.audio_token_id = 151675 # <|audio_pad|> |
| 58 | config.video_token_id = 151656 # <|video_pad|> |
| 59 | config.image_token_id = 151655 # <|image_pad|> |
| 60 | config.audio_start_token_id = 151669 # <|audio_start|> |
| 61 | config.audio_end_token_id = 151670 # <|audio_end|> |
| 62 | config.vision_start_token_id = 151652 # <|vision_start|> |
| 63 | config.position_id_per_seconds = 12.5 |
| 64 | |
| 65 | # Vision config |
| 66 | vision_config = Mock() |
| 67 | vision_config.spatial_merge_size = 2 |
| 68 | config.vision_config = vision_config |
| 69 | |
| 70 | return config |
| 71 | |
| 72 | |
| 73 | @pytest.fixture |
| 74 | def mock_processor(): |
| 75 | """Create a mock HF processor.""" |
| 76 | from transformers.models.whisper import WhisperFeatureExtractor |
| 77 | |
| 78 | processor = Mock() |
| 79 | processor.audio_token = "<|audio_pad|>" |
| 80 | processor.image_token = "<|image_pad|>" |
| 81 | processor.video_token = "<|video_pad|>" |
| 82 | |
| 83 | # Create a real WhisperFeatureExtractor instance for the feature_extractor attribute |
| 84 | feature_extractor = WhisperFeatureExtractor() |
| 85 | processor.feature_extractor = feature_extractor |
| 86 | |
| 87 | return processor |
| 88 | |
| 89 | |
| 90 | @pytest.fixture |
| 91 | def mock_tokenizer(): |
| 92 | """Create a mock tokenizer.""" |
| 93 | tokenizer = Mock() |
| 94 | # Token IDs from https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/blob/main/tokenizer_config.json |
| 95 | tokenizer.get_vocab = Mock( |
| 96 | return_value={ |
| 97 | "<|audio_pad|>": 151675, |
| 98 | "<|video_pad|>": 151656, |
| 99 | "<|image_pad|>": 151655, |
| 100 | "<|audio_start|>": 151669, |
| 101 | "<|audio_end|>": 151670, |
| 102 | "<|vision_start|>": 151652, |
| 103 | "<|vision_end|>": 151653, |
| 104 | } |
| 105 | ) |
| 106 | tokenizer.encode = Mock( |
| 107 | side_effect=lambda x: { |
| 108 | "<|vision_start|>": [151652], |
| 109 | "<|vision_end|>": [151653], |
| 110 | "<|audio_start|>": [151669], |
| 111 | "<|audio_end|>": [151670], |
| 112 | "<|audio_pad|>": [151675], |
| 113 | "<|image_pad|>": [151655], |
| 114 | "<|video_pad|>": [151656], |
| 115 | }.get(x, [0]) |
| 116 | ) |
| 117 | tokenizer.vision_bos_token = "<|vision_start|>" |
| 118 | tokenizer.vision_eos_token = "<|vision_end|>" |
| 119 | tokenizer.audio_bos_token = "<|audio_start|>" |
| 120 | tokenizer.audio_eos_token = "<|audio_end|>" |
| 121 | return tokenizer |
| 122 | |
| 123 | |
| 124 | @pytest.fixture |
| 125 | def mock_image_processor(): |
| 126 | """Create a mock image processor.""" |
| 127 | image_processor = Mock() |
| 128 | image_processor.merge_size = 2 |
| 129 | return image_processor |
| 130 | |
| 131 | |
| 132 | def test_qwen3_omni_get_updates_use_audio_in_video( |
| 133 | mock_qwen3_omni_config, |
| 134 | mock_processor, |
| 135 | mock_tokenizer, |
| 136 | mock_image_processor, |
| 137 | ): |
| 138 | """Test the get_updates_use_audio_in_video method directly.""" |
| 139 | from vllm.model_executor.models.qwen3_omni_moe_thinker import ( |
| 140 | Qwen3OmniMoeThinkerMultiModalProcessor, |
| 141 | Qwen3OmniMoeThinkerProcessingInfo, |
| 142 | ) |
| 143 | |
| 144 | # Create a mock context |
| 145 | mock_ctx = Mock(spec=InputProcessingContext) |
| 146 | mock_ctx.tokenizer = mock_tokenizer |
| 147 | # `model_config` is an instance attribute, so it is not covered by the |
| 148 | # spec; the data parser reads it for `allow_missing_mm_embeddings`. |
| 149 | mock_ctx.model_config = SimpleNamespace( |
| 150 | multimodal_config=SimpleNamespace(allow_missing_mm_embeddings=False) |
| 151 | ) |
| 152 | |
| 153 | # Create processing info |
| 154 | info = Qwen3OmniMoeThinkerProcessingInfo(mock_ctx) |
| 155 | info._get_expected_hidden_size = lambda: 100 |
| 156 | info.get_hf_config = Mock(return_value=mock_qwen3_omni_config) |
| 157 | info.get_hf_processor = Mock(return_value=mock_processor) |
| 158 | info.get_tokenizer = Mock(return_value=mock_tokenizer) |
| 159 | info.get_image_processor = Mock(return_value=mock_image_processor) |
| 160 | |
| 161 | # Create a mock dummy_inputs builder |
| 162 | mock_dummy_inputs = Mock() |
| 163 | |
| 164 | # Create the processor |
| 165 | processor = Qwen3OmniMoeThinkerMultiModalProcessor(info, mock_dummy_inputs) |
| 166 | |
| 167 | # Test parameters from reference video |
| 168 | # https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/draw.mp4 |
| 169 | audio_len = 85 |
| 170 | video_grid_thw = [6, 36, 64] |
| 171 | video_second_per_grid_t = 2.0 |
| 172 | |
| 173 | # Call the method |
| 174 | updates = processor.get_updates_use_audio_in_video( |
| 175 | thinker_config=mock_qwen3_omni_config, |
| 176 | audio_len=audio_len, |
| 177 | video_grid_thw=video_grid_thw, |
| 178 | video_second_per_grid_t=video_second_per_grid_t, |
| 179 | ) |
| 180 | |
| 181 | # Updated input ids should align with HF implementation. |
| 182 | # 151669, |
| 183 | # <|video_pad|> * 576, <|audio_pad|> * 25, |
| 184 | # <|video_pad|> * 576, <|audio_pad|> * 25, |
| 185 | # <|video_pad|> * 576, <|audio_pad|> * 25, |
| 186 | # <|video_pad|> * 576, <|audio_pad|> * 10, |
| 187 | # <|video_pad|> * 1152, |
| 188 | # 151670 |
| 189 | print_input_ids(updates) |
| 190 | |
| 191 | # Verify structure |
| 192 | assert isinstance(updates, list) |
| 193 | assert len(updates) > 0 |
| 194 | |
| 195 | # Verify start and end tokens |
| 196 | audio_start_token_id = mock_qwen3_omni_config.audio_start_token_id |
| 197 | audio_end_token_id = mock_qwen3_omni_config.audio_end_token_id |
| 198 | |
| 199 | assert updates[0] == audio_start_token_id |
| 200 | assert updates[-1] == audio_end_token_id |
| 201 | |
| 202 | # Verify both audio and video tokens are present |
| 203 | audio_token_id = mock_qwen3_omni_config.audio_token_id |
| 204 | video_token_id = mock_qwen3_omni_config.video_token_id |
| 205 | |
| 206 | audio_count = updates.count(audio_token_id) |
| 207 | video_count = updates.count(video_token_id) |
| 208 | |
| 209 | assert audio_count == audio_len, ( |
| 210 | f"Expected {audio_len} audio tokens, got {audio_count}" |
| 211 | ) |
| 212 | |
| 213 | # Calculate expected video token count |
| 214 | spatial_merge_size = mock_qwen3_omni_config.vision_config.spatial_merge_size |
| 215 | height = video_grid_thw[1] // spatial_merge_size |
| 216 | width = video_grid_thw[2] // spatial_merge_size |
| 217 | expected_video_count = video_grid_thw[0] * height * width |
| 218 | |
| 219 | assert video_count == expected_video_count, ( |
| 220 | f"Expected {expected_video_count} video tokens, got {video_count}" |
| 221 | ) |
| 222 | |
| 223 | # Total tokens should be: 1 (start) + audio_len + video_count + 1 (end) |
| 224 | expected_total = 1 + audio_len + expected_video_count + 1 |
| 225 | assert len(updates) == expected_total, ( |
| 226 | f"Expected {expected_total} total tokens, got {len(updates)}" |
| 227 | ) |
| 228 | |
| 229 | |
| 230 | @pytest.mark.skip_global_cleanup |
| 231 | def test_qwen3_omni_exposes_eagle3_to_its_text_backbone(): |
| 232 | from vllm.model_executor.models.interfaces import EagleModelMixin, supports_eagle3 |
| 233 | from vllm.model_executor.models.qwen3_omni_moe_thinker import ( |
| 234 | Qwen3OmniMoeThinkerForConditionalGeneration, |
| 235 | ) |
| 236 | |
| 237 | class DummyBackbone(nn.Module, EagleModelMixin): |
| 238 | def __init__(self): |
| 239 | super().__init__() |
| 240 | self.layers = nn.ModuleList([nn.Identity(), nn.Identity()]) |
| 241 | |
| 242 | class DummyLanguageModel(nn.Module): |
| 243 | def __init__(self): |
| 244 | super().__init__() |
| 245 | self.model = DummyBackbone() |
| 246 | |
| 247 | def embed_input_ids(self, input_ids): |
| 248 | return input_ids |
| 249 | |
| 250 | model = Qwen3OmniMoeThinkerForConditionalGeneration.__new__( |
| 251 | Qwen3OmniMoeThinkerForConditionalGeneration |
| 252 | ) |
| 253 | nn.Module.__init__(model) |
| 254 | model.language_model = DummyLanguageModel() |
| 255 | |
| 256 | assert supports_eagle3(model) |
| 257 | model.set_aux_hidden_state_layers((1, 2)) |
| 258 | assert model.language_model.model.aux_hidden_state_layers == (1, 2) |
| 259 | |
| 260 | |
| 261 | @pytest.mark.skip_global_cleanup |
| 262 | def test_qwen3_omni_text_model_collects_post_deepstack_aux_hidden_states(): |
| 263 | from vllm.model_executor.models.qwen3_omni_moe_thinker import Qwen3MoeLLMModel |
| 264 | |
| 265 | class DummyLayer(nn.Module): |
| 266 | def forward(self, positions, hidden_states, residual): |
| 267 | return hidden_states + 1, torch.full_like(hidden_states, 10) |
| 268 | |
| 269 | class DummyNorm(nn.Module): |
| 270 | def forward(self, hidden_states, residual): |
| 271 | return hidden_states + residual, None |
| 272 | |
| 273 | model = Qwen3MoeLLMModel.__new__(Qwen3MoeLLMModel) |
| 274 | nn.Module.__init__(model) |
| 275 | model.start_layer = 0 |
| 276 | model.end_layer = 1 |
| 277 | model.layers = nn.ModuleList([DummyLayer()]) |
| 278 | model.norm = DummyNorm() |
| 279 | model.aux_hidden_state_layers = (1,) |
| 280 | |
| 281 | pp_group = Mock(is_first_rank=True, is_last_rank=True) |
| 282 | inputs_embeds = torch.tensor([[1.0]]) |
| 283 | deepstack_inputs = {"deepstack_input_embeds_0": torch.tensor([[3.0]])} |
| 284 | with patch( |
| 285 | "vllm.model_executor.models.qwen3_omni_moe_thinker.get_pp_group", |
| 286 | return_value=pp_group, |
| 287 | ): |
| 288 | output, aux_hidden_states = model.forward( |
| 289 | input_ids=None, |
| 290 | positions=torch.tensor([0]), |
| 291 | inputs_embeds=inputs_embeds, |
| 292 | deepstack_input_embeds=deepstack_inputs, |
| 293 | ) |
| 294 | |
| 295 | torch.testing.assert_close(output, torch.tensor([[15.0]])) |
| 296 | assert len(aux_hidden_states) == 1 |
| 297 | torch.testing.assert_close(aux_hidden_states[0], torch.tensor([[15.0]])) |
| 298 | |
| 299 | |
| 300 | def _dspark_vocab_stub(input_vocab_size: int, draft_vocab_size: int): |
| 301 | """A Qwen3 DSpark model with only what `load_weights` reads.""" |
| 302 | from vllm.model_executor.models.qwen3_dspark import Qwen3DSparkForCausalLM |
| 303 | |
| 304 | model = Qwen3DSparkForCausalLM.__new__(Qwen3DSparkForCausalLM) |
| 305 | nn.Module.__init__(model) |
| 306 | object.__setattr__( |
| 307 | model, |
| 308 | "config", |
| 309 | SimpleNamespace( |
| 310 | vocab_size=input_vocab_size, |
| 311 | draft_vocab_size=draft_vocab_size, |
| 312 | ), |
| 313 | ) |
| 314 | object.__setattr__(model, "target_vocab_size", 100) |
| 315 | object.__setattr__(model, "model", Mock(confidence_head=None)) |
| 316 | return model |
| 317 | |
| 318 | |
| 319 | @pytest.mark.skip_global_cleanup |
| 320 | @pytest.mark.parametrize( |
| 321 | ("input_vocab_size", "draft_vocab_size", "weights", "error"), |
| 322 | [ |
| 323 | (101, 100, [], "must include embed_tokens weights"), |
| 324 | (99, 99, [], "must include lm_head weights"), |
| 325 | (100, 40, [], "must include lm_head weights"), |
| 326 | ( |
| 327 | 100, |
| 328 | 40, |
| 329 | [("lm_head.weight", torch.empty(40, 8))], |
| 330 | "must include a d2t mapping", |
| 331 | ), |
| 332 | ], |
| 333 | ) |
| 334 | def test_qwen3_dspark_rejects_incomplete_vocab_weights( |
| 335 | input_vocab_size, draft_vocab_size, weights, error |
| 336 | ): |
| 337 | model = _dspark_vocab_stub(input_vocab_size, draft_vocab_size) |
| 338 | |
| 339 | with pytest.raises(ValueError, match=error): |
| 340 | model.load_weights(weights) |
| 341 | |
| 342 | |
| 343 | @pytest.mark.skip_global_cleanup |
| 344 | @pytest.mark.parametrize("draft_vocab_size", [100, 101]) |
| 345 | def test_qwen3_dspark_accepts_draft_vocab_at_least_target(draft_vocab_size): |
| 346 | """A draft vocab >= the target's maps identically, so needs no lm_head/d2t. |
| 347 | |
| 348 | Guards the padded case (draft > target): the draft's unembedding is padded |
| 349 | to its physical size, which is not a reduced vocabulary. |
| 350 | """ |
| 351 | model = _dspark_vocab_stub(100, draft_vocab_size) |
| 352 | |
| 353 | model.load_weights([]) |
| 354 | |
| 355 | |
| 356 | @pytest.mark.skip_global_cleanup |
| 357 | def test_dspark_shares_target_embedding_with_smaller_draft_vocabulary(): |
| 358 | from vllm.v1.worker.gpu.spec_decode.dspark import utils as dspark_utils |
| 359 | |
| 360 | target_embedding = nn.Embedding(100, 8) |
| 361 | draft_embedding = nn.Embedding(99, 8) |
| 362 | target_model = SimpleNamespace(model=SimpleNamespace(embed_tokens=target_embedding)) |
| 363 | draft_model = SimpleNamespace( |
| 364 | model=SimpleNamespace(embed_tokens=draft_embedding), |
| 365 | has_own_embed_tokens=False, |
| 366 | ) |
| 367 | draft_model_config = SimpleNamespace( |
| 368 | hf_config=SimpleNamespace(model_type="qwen3"), |
| 369 | get_vocab_size=Mock(return_value=99), |
| 370 | ) |
| 371 | vllm_config = SimpleNamespace( |
| 372 | speculative_config=SimpleNamespace( |
| 373 | draft_model_config=draft_model_config, |
| 374 | draft_parallel_config=SimpleNamespace(tensor_parallel_size=1), |
| 375 | attention_backend=None, |
| 376 | kv_cache_dtype=None, |
| 377 | draft_load_config=None, |
| 378 | ), |
| 379 | parallel_config=ParallelConfig(), |
| 380 | attention_config=SimpleNamespace(backend=None), |
| 381 | cache_config=SimpleNamespace(), |
| 382 | load_config=SimpleNamespace(load_format="auto"), |
| 383 | model_config=SimpleNamespace(get_vocab_size=Mock(return_value=100)), |
| 384 | ) |
| 385 | |
| 386 | def fake_replace(config, **changes): |
| 387 | values = vars(config).copy() |
| 388 | values.update(changes) |
| 389 | return SimpleNamespace(**values) |
| 390 | |
| 391 | with ( |
| 392 | patch.object(dspark_utils, "replace", side_effect=fake_replace), |
| 393 | patch( |
| 394 | "vllm.v1.worker.gpu.spec_decode.eagle.utils.get_pp_group", |
| 395 | return_value=SimpleNamespace(world_size=1), |
| 396 | ), |
| 397 | patch( |
| 398 | "vllm.v1.worker.gpu.spec_decode.utils.get_pp_group", |
| 399 | return_value=SimpleNamespace(world_size=1), |
| 400 | ), |
| 401 | patch( |
| 402 | "vllm.compilation.backends.set_model_tag", |
| 403 | return_value=nullcontext(), |
| 404 | ), |
| 405 | patch( |
| 406 | "vllm.model_executor.model_loader.get_model", |
| 407 | return_value=draft_model, |
| 408 | ), |
| 409 | patch( |
| 410 | "vllm.model_executor.models.qwen3_dflash.dflash_has_any_non_causal", |
| 411 | return_value=False, |
| 412 | ), |
| 413 | patch( |
| 414 | "vllm.model_executor.models.utils.get_draft_quant_config", |
| 415 | return_value=None, |
| 416 | ), |
| 417 | ): |
| 418 | loaded_model = dspark_utils.load_dspark_model(target_model, vllm_config) |
| 419 | |
| 420 | assert loaded_model.model.embed_tokens is target_embedding |
| 421 | |
| 422 | |
| 423 | if __name__ == "__main__": |
| 424 | pytest.main([__file__, "-v"]) |