| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for EncoderRunner.gather_mm_embeddings (model runner V2). |
| 4 | |
| 5 | Covers the speculative-drafter encoder-cache handling: the drafter reads one |
| 6 | position ahead of the target model (``draft_lookahead``). The +1 look-ahead |
| 7 | feature past the processed boundary is used when its encoder output is present |
| 8 | and tolerated (token-embedding fallback) when it is not, while a miss within |
| 9 | the processed range still fails loudly. |
| 10 | """ |
| 11 | |
| 12 | from types import SimpleNamespace |
| 13 | from unittest.mock import MagicMock, patch |
| 14 | |
| 15 | import numpy as np |
| 16 | import pytest |
| 17 | import torch |
| 18 | |
| 19 | from vllm.config.compilation import CUDAGraphMode |
| 20 | from vllm.model_executor.models.interfaces import SupportsEncoderCudaGraph |
| 21 | from vllm.multimodal.inputs import ( |
| 22 | MultiModalFeatureSpec, |
| 23 | MultiModalFieldElem, |
| 24 | MultiModalKwargsItem, |
| 25 | MultiModalSharedField, |
| 26 | PlaceholderRange, |
| 27 | ) |
| 28 | from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache |
| 29 | from vllm.v1.worker.gpu.mm.encoder_runner import EncoderRunner |
| 30 | from vllm.v1.worker.gpu.model_states.interface import ModelState |
| 31 | |
| 32 | pytestmark = pytest.mark.cpu_test |
| 33 | |
| 34 | HIDDEN = 4 |
| 35 | |
| 36 | |
| 37 | @pytest.mark.parametrize( |
| 38 | "enabled,enforce_eager,supported,expected", |
| 39 | [ |
| 40 | (True, False, True, True), |
| 41 | (False, False, True, False), |
| 42 | (True, True, True, False), |
| 43 | (True, False, False, False), |
| 44 | ], |
| 45 | ) |
| 46 | def test_encoder_graph_configuration(enabled, enforce_eager, supported, expected): |
| 47 | """Encoder graph eligibility is independent of decoder graph mode.""" |
| 48 | config = MagicMock() |
| 49 | config.model_config.enforce_eager = enforce_eager |
| 50 | config.model_config.dtype = torch.float32 |
| 51 | config.model_config.get_inputs_embeds_size.return_value = HIDDEN |
| 52 | config.scheduler_config.max_num_batched_tokens = 8 |
| 53 | config.compilation_config.cudagraph_mode = CUDAGraphMode.NONE |
| 54 | config.compilation_config.cudagraph_mm_encoder = enabled |
| 55 | model = MagicMock(spec=SupportsEncoderCudaGraph) if supported else torch.nn.Module() |
| 56 | state = _model_state(EncoderCache()) |
| 57 | with patch("vllm.v1.worker.gpu.model_states.interface.EncoderCudaGraphManager"): |
| 58 | ModelState.__init__(state, config, model, state.encoder_cache, state.device) |
| 59 | assert state.encoder_runner.has_cudagraph() is expected |
| 60 | |
| 61 | |
| 62 | def test_encoder_graph_unsupported_modality_uses_eager_output(): |
| 63 | """An image graph must not intercept another modality's encoder output.""" |
| 64 | from vllm.v1.worker.encoder_cudagraph import EncoderCudaGraphManager |
| 65 | |
| 66 | runner = _make_runner([], []) |
| 67 | output = torch.ones(2, HIDDEN) |
| 68 | runner.model = MagicMock() |
| 69 | runner.model.embed_multimodal.return_value = [output] |
| 70 | manager = MagicMock(spec=EncoderCudaGraphManager) |
| 71 | manager.config = SimpleNamespace(modalities=["image"]) |
| 72 | manager.is_captured.return_value = True |
| 73 | manager.supports_modality.side_effect = lambda modality: ( |
| 74 | EncoderCudaGraphManager.supports_modality(manager, modality) |
| 75 | ) |
| 76 | runner.cudagraph_manager = manager |
| 77 | with patch( |
| 78 | "vllm.v1.worker.gpu.mm.encoder_runner.group_and_batch_mm_kwargs", |
| 79 | return_value=[("audio", 1, {"audio_values": torch.zeros(2)})], |
| 80 | ): |
| 81 | result = runner.execute_mm_encoder([]) |
| 82 | assert len(result) == 1 |
| 83 | assert result[0] is output |
| 84 | manager.execute.assert_not_called() |
| 85 | |
| 86 | |
| 87 | def _model_state(cache: EncoderCache) -> MagicMock: |
| 88 | """A mock ModelState backed by a real EncoderCache.""" |
| 89 | state = MagicMock() |
| 90 | state.encoder_cache = cache |
| 91 | state.device = torch.device("cpu") |
| 92 | return state |
| 93 | |
| 94 | |
| 95 | def _embeds_item(embeds: torch.Tensor) -> MultiModalKwargsItem: |
| 96 | """A `prompt_embeds` kwargs item, as the HF renderer builds it.""" |
| 97 | return MultiModalKwargsItem( |
| 98 | { |
| 99 | "embedding": MultiModalFieldElem( |
| 100 | data=embeds, field=MultiModalSharedField(batch_size=1) |
| 101 | ) |
| 102 | } |
| 103 | ) |
| 104 | |
| 105 | |
| 106 | def _feature(identifier: str, offset: int, length: int) -> MultiModalFeatureSpec: |
| 107 | return MultiModalFeatureSpec( |
| 108 | data=None, |
| 109 | modality="image", |
| 110 | identifier=identifier, |
| 111 | mm_position=PlaceholderRange(offset=offset, length=length), |
| 112 | ) |
| 113 | |
| 114 | |
| 115 | def _make_runner( |
| 116 | features: list[MultiModalFeatureSpec], |
| 117 | cached: list[MultiModalFeatureSpec], |
| 118 | ) -> EncoderRunner: |
| 119 | cache = EncoderCache() |
| 120 | cache.mm_features["req0"] = features |
| 121 | for f in cached: |
| 122 | length = f.mm_position.length |
| 123 | cache.encoder_outputs[f.identifier] = torch.arange( |
| 124 | length * HIDDEN, dtype=torch.float32 |
| 125 | ).reshape(length, HIDDEN) |
| 126 | return EncoderRunner( |
| 127 | model=None, # unused by gather_mm_embeddings |
| 128 | max_num_tokens=64, |
| 129 | hidden_size=HIDDEN, |
| 130 | encoder_cache=cache, |
| 131 | dtype=torch.float32, |
| 132 | device=torch.device("cpu"), |
| 133 | ) |
| 134 | |
| 135 | |
| 136 | def _gather(runner: EncoderRunner, *, num_scheduled: int, draft_lookahead: int): |
| 137 | # Single prefilling request, num_computed_tokens=0, prefill_len large. |
| 138 | return runner.gather_mm_embeddings( |
| 139 | req_ids=["req0"], |
| 140 | total_num_scheduled_tokens=num_scheduled, |
| 141 | num_scheduled_tokens=np.array([num_scheduled]), |
| 142 | query_start_loc=np.array([0]), |
| 143 | prefill_lens=np.array([1000]), |
| 144 | num_computed_tokens=np.array([0]), |
| 145 | draft_lookahead=draft_lookahead, |
| 146 | ) |
| 147 | |
| 148 | |
| 149 | def test_draft_lookahead_uses_boundary_feature_when_cached(): |
| 150 | """The drafter's +1 look-ahead can reach the feature at offset == |
| 151 | processed_end (the next chunk). When its encoder output is already cached |
| 152 | (the scheduler encoded it ahead), it is used for the look-ahead position |
| 153 | rather than ignored.""" |
| 154 | f0 = _feature("h0", offset=0, length=8) |
| 155 | f1 = _feature("h1", offset=8, length=8) # starts exactly at processed_end |
| 156 | runner = _make_runner([f0, f1], cached=[f0, f1]) |
| 157 | |
| 158 | mm_embeds, is_mm_embed = _gather(runner, num_scheduled=8, draft_lookahead=1) |
| 159 | |
| 160 | # f0 covers positions 0..6 (+1 skew); f1's first embed covers position 7. |
| 161 | assert len(mm_embeds) == 2 |
| 162 | assert [e.modality for e in mm_embeds] == ["image", "image"] |
| 163 | assert bool(is_mm_embed[7]) |
| 164 | assert int(is_mm_embed.sum()) == 8 |
| 165 | |
| 166 | |
| 167 | def test_draft_lookahead_tolerates_missing_boundary_feature(): |
| 168 | """When the +1 look-ahead feature past the processed boundary is not yet |
| 169 | encoded, fall back to the token embedding (the draft token is verified by |
| 170 | the target) instead of raising.""" |
| 171 | f0 = _feature("h0", offset=0, length=8) |
| 172 | f1 = _feature("h1", offset=8, length=8) # boundary feature, not cached |
| 173 | runner = _make_runner([f0, f1], cached=[f0]) |
| 174 | |
| 175 | mm_embeds, is_mm_embed = _gather(runner, num_scheduled=8, draft_lookahead=1) |
| 176 | |
| 177 | # Only f0 is gathered; f1's boundary position falls back silently. |
| 178 | assert len(mm_embeds) == 1 |
| 179 | assert [e.modality for e in mm_embeds] == ["image"] |
| 180 | assert not bool(is_mm_embed[7]) |
| 181 | assert int(is_mm_embed.sum()) == 7 |
| 182 | |
| 183 | |
| 184 | def test_draft_lookahead_raises_on_interior_miss(): |
| 185 | """A miss for a feature within the processed range (not the look-ahead |
| 186 | boundary) is a real invariant violation and must fail loudly, even on the |
| 187 | drafter path.""" |
| 188 | f0 = _feature("h0", offset=0, length=8) # interior, within processed range |
| 189 | runner = _make_runner([f0], cached=[]) |
| 190 | |
| 191 | with pytest.raises(RuntimeError, match="Encoder cache miss"): |
| 192 | _gather(runner, num_scheduled=8, draft_lookahead=1) |
| 193 | |
| 194 | |
| 195 | def test_target_path_raises_on_encoder_cache_miss(): |
| 196 | """On the target path (no look-ahead) a miss is a real invariant |
| 197 | violation and must fail loudly.""" |
| 198 | f0 = _feature("h0", offset=0, length=8) |
| 199 | runner = _make_runner([f0], cached=[]) |
| 200 | |
| 201 | with pytest.raises(RuntimeError, match="Encoder cache miss"): |
| 202 | _gather(runner, num_scheduled=8, draft_lookahead=0) |
| 203 | |
| 204 | |
| 205 | @pytest.mark.parametrize("draft_lookahead", [0, 1]) |
| 206 | def test_multi_request_batch_gathers_per_request(draft_lookahead): |
| 207 | """Two prefilling requests in one batch: per-request query bounds must be |
| 208 | indexed by request, not applied as whole arrays.""" |
| 209 | a0 = _feature("a0", offset=0, length=8) |
| 210 | b0 = _feature("b0", offset=0, length=8) |
| 211 | cache = EncoderCache() |
| 212 | cache.mm_features["req0"] = [a0] |
| 213 | cache.mm_features["req1"] = [b0] |
| 214 | for f in (a0, b0): |
| 215 | cache.encoder_outputs[f.identifier] = torch.arange( |
| 216 | f.mm_position.length * HIDDEN, dtype=torch.float32 |
| 217 | ).reshape(f.mm_position.length, HIDDEN) |
| 218 | runner = EncoderRunner( |
| 219 | model=None, |
| 220 | max_num_tokens=64, |
| 221 | hidden_size=HIDDEN, |
| 222 | encoder_cache=cache, |
| 223 | dtype=torch.float32, |
| 224 | device=torch.device("cpu"), |
| 225 | ) |
| 226 | |
| 227 | mm_embeds, is_mm_embed = runner.gather_mm_embeddings( |
| 228 | req_ids=["req0", "req1"], |
| 229 | total_num_scheduled_tokens=16, |
| 230 | num_scheduled_tokens=np.array([8, 8]), |
| 231 | query_start_loc=np.array([0, 8]), |
| 232 | prefill_lens=np.array([1000, 1000]), |
| 233 | num_computed_tokens=np.array([0, 0]), |
| 234 | draft_lookahead=draft_lookahead, |
| 235 | ) |
| 236 | |
| 237 | # Both requests contribute a feature; with the +1 skew each marks 7 of its |
| 238 | # 8 positions (the skew drops one), otherwise all 8. |
| 239 | assert len(mm_embeds) == 2 |
| 240 | assert [e.modality for e in mm_embeds] == ["image", "image"] |
| 241 | assert int(is_mm_embed.sum()) == (14 if draft_lookahead else 16) |
| 242 | |
| 243 | |
| 244 | def test_gather_preserves_mixed_modalities(): |
| 245 | """Modalities must be attached on tensors in gather order.""" |
| 246 | video = MultiModalFeatureSpec( |
| 247 | data=None, |
| 248 | modality="video", |
| 249 | identifier="v0", |
| 250 | mm_position=PlaceholderRange(offset=0, length=4), |
| 251 | ) |
| 252 | audio = MultiModalFeatureSpec( |
| 253 | data=None, |
| 254 | modality="audio", |
| 255 | identifier="a0", |
| 256 | mm_position=PlaceholderRange(offset=4, length=4), |
| 257 | ) |
| 258 | runner = _make_runner([video, audio], cached=[video, audio]) |
| 259 | |
| 260 | mm_embeds, is_mm_embed = _gather(runner, num_scheduled=8, draft_lookahead=0) |
| 261 | |
| 262 | assert len(mm_embeds) == 2 |
| 263 | assert [e.modality for e in mm_embeds] == ["video", "audio"] |
| 264 | assert int(is_mm_embed.sum()) == 8 |
| 265 | |
| 266 | |
| 267 | def test_execute_mm_encoder_caches_outputs_without_gathering(): |
| 268 | """An encoder instance encodes and publishes, and must stop there. |
| 269 | |
| 270 | `ModelState.execute_mm_encoder` is the half of `get_mm_embeddings` that an |
| 271 | EPD encoder instance needs: it runs no language model, so gathering would |
| 272 | build an `inputs_embeds` nobody reads -- and the gather raises |
| 273 | `Encoder cache miss` for any scheduled item absent from the local cache, |
| 274 | which on a producer takes the whole engine down (the scheduler hands it |
| 275 | items the connector already holds, and a producer has no load path). |
| 276 | """ |
| 277 | cache = EncoderCache() |
| 278 | state = _model_state(cache) |
| 279 | embedding = torch.ones(2, HIDDEN) |
| 280 | # (mm_hashes, [(modality, kwargs item), ...]), as prepare_mm_inputs returns. |
| 281 | state.encoder_runner.prepare_mm_inputs.return_value = ( |
| 282 | ["hash0"], |
| 283 | [("image", MagicMock())], |
| 284 | ) |
| 285 | state.encoder_runner.execute_mm_encoder.return_value = [embedding] |
| 286 | |
| 287 | ModelState.execute_mm_encoder(state, {"req0": [0]}) |
| 288 | |
| 289 | assert cache.encoder_outputs == {"hash0": embedding} |
| 290 | state.encoder_runner.gather_mm_embeddings.assert_not_called() |
| 291 | |
| 292 | |
| 293 | def test_execute_mm_encoder_is_a_noop_without_scheduled_items(): |
| 294 | """A step that schedules no encoder input must not touch the encoder.""" |
| 295 | cache = EncoderCache() |
| 296 | state = _model_state(cache) |
| 297 | state.encoder_runner.prepare_mm_inputs.return_value = ([], []) |
| 298 | |
| 299 | ModelState.execute_mm_encoder(state, {}) |
| 300 | |
| 301 | assert not cache.encoder_outputs |
| 302 | state.encoder_runner.execute_mm_encoder.assert_not_called() |
| 303 | |
| 304 | |
| 305 | def _pe_feature(identifier: str, embeds: torch.Tensor, offset: int = 0): |
| 306 | return MultiModalFeatureSpec( |
| 307 | data=_embeds_item(embeds), |
| 308 | modality="prompt_embeds", |
| 309 | identifier=identifier, |
| 310 | mm_position=PlaceholderRange(offset=offset, length=embeds.shape[0]), |
| 311 | ) |
| 312 | |
| 313 | |
| 314 | def test_prepare_mm_inputs_passes_prompt_embeds_through(): |
| 315 | """`prompt_embeds` is already in embedding space, so no encoder may run. |
| 316 | |
| 317 | The renderer delivers prompt_embeds mixed with real media as an ordinary MM |
| 318 | modality. prepare_mm_inputs must cache the tensor directly and keep it out |
| 319 | of the encoder batch -- the vision encoder cannot consume it, and a missing |
| 320 | cache entry makes the subsequent gather raise "Encoder cache miss". |
| 321 | """ |
| 322 | prompt_embeds = torch.arange(2 * HIDDEN, dtype=torch.float32).view(2, HIDDEN) |
| 323 | image_feature = MultiModalFeatureSpec( |
| 324 | data=MagicMock(), |
| 325 | modality="image", |
| 326 | identifier="hash_img", |
| 327 | mm_position=PlaceholderRange(offset=2, length=2), |
| 328 | ) |
| 329 | runner = _make_runner( |
| 330 | [_pe_feature("hash_pe", prompt_embeds), image_feature], cached=[] |
| 331 | ) |
| 332 | |
| 333 | mm_hashes, mm_kwargs = runner.prepare_mm_inputs({"req0": [0, 1]}) |
| 334 | |
| 335 | # Only the image remains for the encoder; the embeds are already cached. |
| 336 | assert mm_hashes == ["hash_img"] |
| 337 | assert [modality for modality, _ in mm_kwargs] == ["image"] |
| 338 | assert torch.equal(runner.encoder_cache.encoder_outputs["hash_pe"], prompt_embeds) |
| 339 | |
| 340 | |
| 341 | def test_prepare_mm_inputs_skips_cached_prompt_embeds(): |
| 342 | """A prompt_embeds item already in the cache must not be re-uploaded.""" |
| 343 | prompt_embeds = torch.ones(3, HIDDEN) |
| 344 | feature = _pe_feature("hash_pe", prompt_embeds) |
| 345 | runner = _make_runner([feature], cached=[feature]) |
| 346 | sentinel = runner.encoder_cache.encoder_outputs["hash_pe"] |
| 347 | |
| 348 | mm_hashes, mm_kwargs = runner.prepare_mm_inputs({"req0": [0]}) |
| 349 | |
| 350 | assert mm_hashes == [] and mm_kwargs == [] |
| 351 | assert runner.encoder_cache.encoder_outputs["hash_pe"] is sentinel |
| 352 | |
| 353 | |
| 354 | def test_execute_mm_encoder_skips_encoder_for_prompt_embeds_only(): |
| 355 | """A batch of nothing but prompt_embeds must not invoke the encoder.""" |
| 356 | prompt_embeds = torch.ones(3, HIDDEN) |
| 357 | runner = _make_runner([_pe_feature("hash_pe", prompt_embeds)], cached=[]) |
| 358 | state = _model_state(runner.encoder_cache) |
| 359 | state.encoder_runner.prepare_mm_inputs.side_effect = runner.prepare_mm_inputs |
| 360 | |
| 361 | ModelState.execute_mm_encoder(state, {"req0": [0]}) |
| 362 | |
| 363 | state.encoder_runner.execute_mm_encoder.assert_not_called() |
| 364 | assert torch.equal(runner.encoder_cache.encoder_outputs["hash_pe"], prompt_embeds) |
| 365 | |
| 366 | |
| 367 | def test_encoder_timing_stats_registry(): |
| 368 | runner = _make_runner([], []) |
| 369 | runner.enable_timing = True |
| 370 | |
| 371 | with runner.timed_encoder_operation({"r1"}): |
| 372 | pass |
| 373 | with runner.timed_encoder_operation({"r1"}): |
| 374 | pass |
| 375 | |
| 376 | stats = runner.get_encoder_timing_stats() |
| 377 | assert set(stats) == {"r1"} |
| 378 | assert stats["r1"]["num_encoder_calls"] == 2 |
| 379 | assert stats["r1"]["encoder_forward_secs"] >= 0 |
| 380 | assert runner.get_encoder_timing_stats() == {} |