| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import unittest |
| 5 | from unittest.mock import MagicMock |
| 6 | |
| 7 | import numpy as np |
| 8 | import torch |
| 9 | |
| 10 | from vllm.config import DeviceConfig, VllmConfig |
| 11 | from vllm.multimodal.inputs import ( |
| 12 | MultiModalFeatureSpec, |
| 13 | MultiModalKwargsItem, |
| 14 | PlaceholderRange, |
| 15 | ) |
| 16 | from vllm.sampling_params import SamplingParams |
| 17 | from vllm.v1.core.sched.scheduler import Scheduler |
| 18 | from vllm.v1.engine import FinishReason |
| 19 | from vllm.v1.kv_cache_interface import ( |
| 20 | FullAttentionSpec, |
| 21 | KVCacheConfig, |
| 22 | KVCacheGroupSpec, |
| 23 | ) |
| 24 | from vllm.v1.outputs import LogprobsLists, ModelRunnerOutput |
| 25 | from vllm.v1.request import Request, RequestStatus, StreamingUpdate |
| 26 | from vllm.v1.structured_output import StructuredOutputManager |
| 27 | |
| 28 | STOP_TOKEN = 128001 |
| 29 | |
| 30 | |
| 31 | class DummyRequest(Request): |
| 32 | def __init__( |
| 33 | self, |
| 34 | request_id, |
| 35 | resumable=True, |
| 36 | prompt_token_ids=None, |
| 37 | mm_features: list[MultiModalFeatureSpec] | None = None, |
| 38 | max_tokens: int | None = 16, |
| 39 | ): |
| 40 | super().__init__( |
| 41 | request_id=request_id, |
| 42 | prompt_token_ids=prompt_token_ids if prompt_token_ids is not None else [], |
| 43 | sampling_params=SamplingParams( |
| 44 | stop_token_ids=[STOP_TOKEN], max_tokens=max_tokens |
| 45 | ), |
| 46 | pooling_params=None, |
| 47 | mm_features=mm_features, |
| 48 | resumable=resumable, |
| 49 | ) |
| 50 | |
| 51 | |
| 52 | def create_scheduler() -> Scheduler: |
| 53 | vllm_config = VllmConfig(device_config=DeviceConfig("cpu")) |
| 54 | vllm_config.model_config = MagicMock() |
| 55 | vllm_config.model_config.skip_tokenizer_init = True |
| 56 | vllm_config.model_config.is_multimodal_model = False |
| 57 | vllm_config.model_config.supports_multimodal_inputs = False |
| 58 | vllm_config.model_config.is_encoder_decoder = False |
| 59 | vllm_config.model_config.is_diffusion = False |
| 60 | vllm_config.model_config.max_model_len = 1024 |
| 61 | vllm_config.model_config.enable_return_routed_experts = False |
| 62 | vllm_config.cache_config = MagicMock() |
| 63 | vllm_config.cache_config.num_gpu_blocks = 1000 |
| 64 | vllm_config.cache_config.enable_prefix_caching = False |
| 65 | kv_cache_config = KVCacheConfig( |
| 66 | num_blocks=1000, |
| 67 | kv_cache_tensors=[], |
| 68 | kv_cache_groups=[ |
| 69 | KVCacheGroupSpec( |
| 70 | ["layer"], |
| 71 | FullAttentionSpec( |
| 72 | block_size=16, num_kv_heads=1, head_size=1, dtype=torch.float32 |
| 73 | ), |
| 74 | ) |
| 75 | ], |
| 76 | ) |
| 77 | return Scheduler( |
| 78 | vllm_config=vllm_config, |
| 79 | kv_cache_config=kv_cache_config, |
| 80 | log_stats=True, |
| 81 | structured_output_manager=StructuredOutputManager(vllm_config), |
| 82 | block_size=16, |
| 83 | hash_block_size=16, |
| 84 | ) |
| 85 | |
| 86 | |
| 87 | class TestStreamingScheduler(unittest.TestCase): |
| 88 | def test_chunk_logprobs_do_not_depend_on_continuation_arrival(self): |
| 89 | """A queued continuation must not change the finishing chunk's logprobs.""" |
| 90 | for continuation_queued in (False, True): |
| 91 | with self.subTest(continuation_queued=continuation_queued): |
| 92 | scheduler = create_scheduler() |
| 93 | session = DummyRequest("session", prompt_token_ids=[1, 2, 3]) |
| 94 | session.sampling_params = SamplingParams( |
| 95 | stop_token_ids=[STOP_TOKEN], max_tokens=16, logprobs=0 |
| 96 | ) |
| 97 | scheduler.add_request(session) |
| 98 | scheduled = scheduler.schedule() |
| 99 | |
| 100 | continuation = DummyRequest("session", prompt_token_ids=[4, 5]) |
| 101 | assert continuation.sampling_params.logprobs is None |
| 102 | if continuation_queued: |
| 103 | scheduler.add_request(continuation) |
| 104 | |
| 105 | logprobs = LogprobsLists( |
| 106 | logprob_token_ids=np.array([[STOP_TOKEN]], dtype=np.int32), |
| 107 | logprobs=np.array([[-0.25]], dtype=np.float32), |
| 108 | sampled_token_ranks=np.array([1], dtype=np.int32), |
| 109 | ) |
| 110 | runner_output = ModelRunnerOutput( |
| 111 | req_ids=[session.request_id], |
| 112 | req_id_to_index={session.request_id: 0}, |
| 113 | sampled_token_ids=[[STOP_TOKEN]], |
| 114 | logprobs=logprobs, |
| 115 | prompt_logprobs_dict={}, |
| 116 | pooler_output=[], |
| 117 | ) |
| 118 | outputs = scheduler.update_from_output(scheduled, runner_output) |
| 119 | [output] = outputs[session.client_index].outputs |
| 120 | |
| 121 | if not continuation_queued: |
| 122 | assert session.status == RequestStatus.WAITING_FOR_STREAMING_REQ |
| 123 | scheduler.add_request(continuation) |
| 124 | assert session.status == RequestStatus.WAITING |
| 125 | assert session.sampling_params is continuation.sampling_params |
| 126 | assert output.new_token_ids == [STOP_TOKEN] |
| 127 | assert output.finish_reason == FinishReason.STOP |
| 128 | assert output.new_logprobs is not None, ( |
| 129 | "The finishing chunk requested logprobs; the continuation's " |
| 130 | "settings must not suppress them" |
| 131 | ) |
| 132 | np.testing.assert_array_equal( |
| 133 | output.new_logprobs.logprob_token_ids, logprobs.logprob_token_ids |
| 134 | ) |
| 135 | np.testing.assert_array_equal( |
| 136 | output.new_logprobs.logprobs, logprobs.logprobs |
| 137 | ) |
| 138 | np.testing.assert_array_equal( |
| 139 | output.new_logprobs.sampled_token_ranks, |
| 140 | logprobs.sampled_token_ranks, |
| 141 | ) |
| 142 | |
| 143 | def test_add_request(self): |
| 144 | scheduler = create_scheduler() |
| 145 | |
| 146 | request = DummyRequest( |
| 147 | request_id="test_request", |
| 148 | resumable=True, |
| 149 | ) |
| 150 | |
| 151 | scheduler.add_request(request) |
| 152 | |
| 153 | assert "test_request" in scheduler.requests |
| 154 | assert request.status == RequestStatus.WAITING |
| 155 | assert len(scheduler.waiting) == 1 |
| 156 | |
| 157 | next_request = DummyRequest( |
| 158 | request_id="test_request", |
| 159 | resumable=True, |
| 160 | ) |
| 161 | scheduler.add_request(next_request) |
| 162 | |
| 163 | assert next_request.status == RequestStatus.WAITING |
| 164 | assert len(scheduler.requests["test_request"].streaming_queue) == 1 |
| 165 | |
| 166 | def test_update_request_as_session_max_token(self): |
| 167 | scheduler = create_scheduler() |
| 168 | |
| 169 | session = DummyRequest( |
| 170 | request_id="session", |
| 171 | prompt_token_ids=[1, 2, 3], |
| 172 | ) |
| 173 | session.num_computed_tokens = len(session.prompt_token_ids) |
| 174 | session.max_tokens = 2 |
| 175 | |
| 176 | new_request = DummyRequest( |
| 177 | request_id="session", |
| 178 | prompt_token_ids=[4, 5, 6], |
| 179 | ) |
| 180 | new_request.sampling_params = SamplingParams(max_tokens=10) |
| 181 | new_request.max_tokens = 10 # Additional max_tokens from new request |
| 182 | |
| 183 | update = StreamingUpdate.from_request(new_request) |
| 184 | scheduler._update_request_as_session(session, update) |
| 185 | |
| 186 | assert session.sampling_params.max_tokens == 10 |
| 187 | assert session.max_tokens == 10 |
| 188 | |
| 189 | session.num_computed_tokens = len(session.prompt_token_ids) |
| 190 | |
| 191 | new_request2 = DummyRequest( |
| 192 | request_id="session", |
| 193 | prompt_token_ids=[7, 8, 9], |
| 194 | ) |
| 195 | new_request2.sampling_params = SamplingParams(max_tokens=4) |
| 196 | new_request2.max_tokens = 4 |
| 197 | update2 = StreamingUpdate.from_request(new_request2) |
| 198 | scheduler._update_request_as_session(session, update2) |
| 199 | |
| 200 | assert session.sampling_params.max_tokens == 4 |
| 201 | assert session.max_tokens == 4 |
| 202 | |
| 203 | def test_update_request_as_session(self): |
| 204 | scheduler = create_scheduler() |
| 205 | |
| 206 | session = DummyRequest( |
| 207 | request_id="session", |
| 208 | prompt_token_ids=[1, 2, 3], |
| 209 | ) |
| 210 | session.num_computed_tokens = len(session.prompt_token_ids) |
| 211 | |
| 212 | new_request = DummyRequest( |
| 213 | request_id="session", |
| 214 | prompt_token_ids=[4, 5, 6], |
| 215 | ) |
| 216 | new_request.sampling_params = SamplingParams(max_tokens=10) |
| 217 | |
| 218 | update = StreamingUpdate.from_request(new_request) |
| 219 | scheduler._update_request_as_session(session, update) |
| 220 | |
| 221 | assert session.prompt_token_ids == [1, 2, 3, 4, 5, 6] |
| 222 | assert session._all_token_ids == [1, 2, 3, 4, 5, 6] |
| 223 | assert session.sampling_params.max_tokens == 10 |
| 224 | assert session.status == RequestStatus.WAITING |
| 225 | |
| 226 | def test_update_request_as_session_with_multimodal(self): |
| 227 | scheduler = create_scheduler() |
| 228 | |
| 229 | mm_feature = MultiModalFeatureSpec( |
| 230 | data=MultiModalKwargsItem.dummy(), |
| 231 | modality="audio", |
| 232 | identifier="", |
| 233 | mm_position=PlaceholderRange(offset=1, length=1), |
| 234 | ) |
| 235 | session = DummyRequest( |
| 236 | request_id="session", |
| 237 | prompt_token_ids=[1, 2, 3], |
| 238 | mm_features=[mm_feature], |
| 239 | ) |
| 240 | session.num_computed_tokens = len(session.prompt_token_ids) |
| 241 | |
| 242 | mm_feature = MultiModalFeatureSpec( |
| 243 | data=MultiModalKwargsItem.dummy(), |
| 244 | modality="audio", |
| 245 | identifier="", |
| 246 | mm_position=PlaceholderRange(offset=2, length=1), |
| 247 | ) |
| 248 | new_request = DummyRequest( |
| 249 | request_id="session", |
| 250 | prompt_token_ids=[4, 5, 6, 7], |
| 251 | mm_features=[mm_feature], |
| 252 | ) |
| 253 | update = StreamingUpdate.from_request(new_request) |
| 254 | scheduler._update_request_as_session(session, update) |
| 255 | |
| 256 | assert len(session.mm_features) == 2 |
| 257 | assert session.mm_features[0].mm_position.offset == 1 |
| 258 | # 2 + len([1, 2, 3]) |
| 259 | assert session.mm_features[1].mm_position.offset == 5 |
| 260 | |
| 261 | def test_process_streaming_requests_with_finish_session(self): |
| 262 | """Test that a non-resumable request signals stream completion. |
| 263 | |
| 264 | With the new streaming API, completion is signaled by closing/finishing |
| 265 | the input generator. When a non-resumable request is added to a session |
| 266 | in WAITING_FOR_STREAMING_REQ state, the session is finished immediately |
| 267 | with FINISHED_ABORTED status. |
| 268 | """ |
| 269 | scheduler = create_scheduler() |
| 270 | |
| 271 | session = DummyRequest( |
| 272 | request_id="session", |
| 273 | prompt_token_ids=[1, 2, 3], |
| 274 | resumable=True, |
| 275 | ) |
| 276 | scheduler.add_request(session) |
| 277 | session.status = RequestStatus.WAITING_FOR_STREAMING_REQ |
| 278 | session.num_computed_tokens = len(session.prompt_token_ids) |
| 279 | |
| 280 | # A non-resumable request signals stream completion |
| 281 | close_request = DummyRequest( |
| 282 | request_id="session", |
| 283 | prompt_token_ids=[0], |
| 284 | resumable=False, |
| 285 | max_tokens=1, |
| 286 | ) |
| 287 | scheduler.add_request(close_request) |
| 288 | |
| 289 | # The session should be immediately finished (stream completed) |
| 290 | assert session.status == RequestStatus.FINISHED_ABORTED |
| 291 | # The session should be removed from the scheduler |
| 292 | assert session.request_id not in scheduler.requests |
| 293 | |
| 294 | def test_streaming_request_session_update(self): |
| 295 | """Test that a resumable request updates a waiting session directly. |
| 296 | |
| 297 | When a session is in WAITING_FOR_STREAMING_REQ state and a new resumable |
| 298 | request arrives, the update is applied directly via _update_request_as_session, |
| 299 | not queued. |
| 300 | """ |
| 301 | scheduler = create_scheduler() |
| 302 | |
| 303 | session = DummyRequest( |
| 304 | request_id="session", |
| 305 | prompt_token_ids=[1, 2, 3], |
| 306 | resumable=True, |
| 307 | ) |
| 308 | scheduler.add_request(session) |
| 309 | session.status = RequestStatus.WAITING_FOR_STREAMING_REQ |
| 310 | session.num_computed_tokens = len(session.prompt_token_ids) |
| 311 | |
| 312 | next_request = DummyRequest( |
| 313 | request_id="session", |
| 314 | prompt_token_ids=[4, 5], |
| 315 | resumable=True, |
| 316 | ) |
| 317 | |
| 318 | scheduler.add_request(next_request) |
| 319 | |
| 320 | # With the new behavior, when session is in WAITING_FOR_STREAMING_REQ, |
| 321 | # the update is applied directly (not queued), and session status |
| 322 | # becomes WAITING |
| 323 | assert session.status == RequestStatus.WAITING |
| 324 | assert session.prompt_token_ids == [1, 2, 3, 4, 5] |
| 325 | |
| 326 | _ = scheduler.schedule() |
| 327 | |
| 328 | assert session.status == RequestStatus.RUNNING |
| 329 | |
| 330 | def test_update_request_as_session_with_output_tokens(self): |
| 331 | scheduler = create_scheduler() |
| 332 | |
| 333 | session = DummyRequest( |
| 334 | request_id="session", |
| 335 | prompt_token_ids=[1, 2, 3], # 3 prompt tokens |
| 336 | ) |
| 337 | session.append_output_token_ids([10, 11]) |
| 338 | """ |
| 339 | The last output token (11) hasn't been "scheduled" yet, so `num_computed_tokens` |
| 340 | only includes: 3 prompt + 1 output (the 10) = 4 |
| 341 | """ |
| 342 | session.num_computed_tokens = 4 |
| 343 | |
| 344 | new_request = DummyRequest( |
| 345 | request_id="session", |
| 346 | prompt_token_ids=[4, 5], |
| 347 | ) |
| 348 | |
| 349 | update = StreamingUpdate.from_request(new_request) |
| 350 | scheduler._update_request_as_session(session, update) |
| 351 | |
| 352 | # _update_request_as_session keeps computed output tokens (they become |
| 353 | # part of the prompt) and only discards the final uncomputed sampled |
| 354 | # token. Computed output token 10 is kept, uncomputed token 11 is |
| 355 | # discarded. |
| 356 | assert session._all_token_ids == [1, 2, 3, 10, 4, 5] |
| 357 | assert session.prompt_token_ids == [1, 2, 3, 10, 4, 5] |
| 358 | # Output tokens list is cleared |
| 359 | assert session._output_token_ids == [] |
| 360 | # num_computed_tokens is unchanged (KV cache still valid for computed |
| 361 | # tokens) |
| 362 | assert session.num_computed_tokens == 4 |
| 363 | # Verify that the next schedule will only process the new prompt tokens |
| 364 | # num_new_tokens = num_tokens - num_computed_tokens = 6 - 4 = 2 |
| 365 | num_new_tokens = session.num_tokens - session.num_computed_tokens |
| 366 | assert num_new_tokens == 2 |
| 367 | |
| 368 | def test_streaming_e2e_lifecycle(self): |
| 369 | """Comprehensive integration test covering complete streaming request lifecycle |
| 370 | including scheduler state management and aliasing bug prevention. |
| 371 | |
| 372 | FULL LIFECYCLE: |
| 373 | ================ |
| 374 | CYCLE 1 (Initial Decode): |
| 375 | 1. Add streaming request (seq_id=0) with prompt tokens [1,2,3] |
| 376 | 2. Schedule() creates NewRequestData with prompt_token_ids |
| 377 | 3. Model runner caches this prompt_token_ids reference (simulated) |
| 378 | 4. Model executes and generates output token 10 |
| 379 | 5. update_from_output() appends token 10 to request._all_token_ids |
| 380 | 6. Request transitions to RUNNING state |
| 381 | |
| 382 | CYCLE 2 (Continue Decode): |
| 383 | 7. Schedule() again - request is now in scheduled_cached_reqs (not new) |
| 384 | 8. Model runner uses CACHED state to calculate num_tokens |
| 385 | 9. Model generates output token (STOP_TOKEN) |
| 386 | 10. update_from_output() appends STOP_TOKEN to request._all_token_ids |
| 387 | 11. Request transitions to WAITING_FOR_STREAMING_REQ |
| 388 | |
| 389 | CYCLE 3 (New Streaming Request): |
| 390 | 12. Add new streaming request (seq_id=1) with prompt tokens [4,5] |
| 391 | 13. Scheduler merges into session, creates NewRequestData again |
| 392 | 14. Model runner caches new prompt_token_ids reference |
| 393 | 15. Verify cached state from Cycle 1 wasn't corrupted by mutations |
| 394 | |
| 395 | CRITICAL BUG PREVENTION: |
| 396 | ======================== |
| 397 | Without .copy() in _create_new_request_data(): |
| 398 | - Cycle 1 Step 3: cached_state["prompt_token_ids"] aliases |
| 399 | request._all_token_ids |
| 400 | - Cycle 1 Step 5: When appending token 10, cached state mutates: |
| 401 | [1,2,3] -> [1,2,3,10] |
| 402 | - Cycle 2 Step 8: num_tokens = len([1,2,3,10]) + len([10]) |
| 403 | = 5 (WRONG! Should be 4) |
| 404 | - Cycle 2: Discard logic would see seq_lens=4 < num_tokens=5 |
| 405 | -> INCORRECTLY DISCARDS |
| 406 | |
| 407 | With .copy() in _create_new_request_data(): |
| 408 | - Cycle 1 Step 3: cached_state["prompt_token_ids"] is independent copy |
| 409 | - Cycle 1 Step 5: Only request._all_token_ids mutates, cached stays [1,2,3] |
| 410 | - Cycle 2 Step 8: num_tokens = len([1,2,3]) + len([10]) = 4 (CORRECT) |
| 411 | - Cycle 2: Discard logic works correctly |
| 412 | """ |
| 413 | scheduler = create_scheduler() |
| 414 | |
| 415 | # ═══════════════════════════════════════════════════════════════════ |
| 416 | # CYCLE 1: Initial Request Scheduling and First Decode |
| 417 | # ═══════════════════════════════════════════════════════════════════ |
| 418 | |
| 419 | session = DummyRequest( |
| 420 | request_id="session", |
| 421 | prompt_token_ids=[1, 2, 3], |
| 422 | ) |
| 423 | scheduler.add_request(session) |
| 424 | |
| 425 | # Step 2: Schedule creates NewRequestData |
| 426 | scheduler_output_cycle1 = scheduler.schedule() |
| 427 | |
| 428 | # Verify request is in scheduled_new_reqs (first time scheduling) |
| 429 | assert len(scheduler_output_cycle1.scheduled_new_reqs) == 1 |
| 430 | new_req_data_cycle1 = scheduler_output_cycle1.scheduled_new_reqs[0] |
| 431 | assert new_req_data_cycle1.prompt_token_ids == [1, 2, 3] |
| 432 | assert ( |
| 433 | scheduler_output_cycle1.num_scheduled_tokens[session.request_id] == 3 |
| 434 | ) # [1, 2, 3] |
| 435 | assert ( |
| 436 | session.request_id |
| 437 | not in scheduler_output_cycle1.scheduled_cached_reqs.req_ids |
| 438 | ) |
| 439 | |
| 440 | # Step 3: Simulate model runner caching the prompt_token_ids |
| 441 | # This simulates gpu_model_runner.py:706-720 CachedRequestState creation |
| 442 | # The model runner makes a copy of prompt_token_ids when creating |
| 443 | # CachedRequestState |
| 444 | cached_state_cycle1 = { |
| 445 | "req_id": session.request_id, |
| 446 | "prompt_token_ids": list( |
| 447 | new_req_data_cycle1.prompt_token_ids |
| 448 | ), # Explicit copy |
| 449 | "output_token_ids": [], |
| 450 | "num_computed_tokens": 0, |
| 451 | } |
| 452 | |
| 453 | # Store original for verification |
| 454 | original_cached_prompt_cycle1 = cached_state_cycle1["prompt_token_ids"].copy() |
| 455 | |
| 456 | # Step 4-5: Model execution generates token, scheduler updates request |
| 457 | output_token_1 = 10 |
| 458 | cached_state_cycle1["output_token_ids"].append(output_token_1) |
| 459 | |
| 460 | mro_cycle1 = ModelRunnerOutput( |
| 461 | req_ids=[session.request_id], |
| 462 | req_id_to_index={session.request_id: 0}, |
| 463 | sampled_token_ids=[[output_token_1]], |
| 464 | logprobs=None, |
| 465 | prompt_logprobs_dict={session.request_id: None}, |
| 466 | pooler_output=[], |
| 467 | ) |
| 468 | session.num_computed_tokens = len(session.prompt_token_ids) |
| 469 | eco_dict_cycle1 = scheduler.update_from_output( |
| 470 | scheduler_output_cycle1, mro_cycle1 |
| 471 | ) |
| 472 | |
| 473 | # Step 6: Verify request state after Cycle 1 |
| 474 | eco_cycle1 = eco_dict_cycle1[session.client_index].outputs[0] |
| 475 | assert eco_cycle1.finish_reason is None # Not stopped yet |
| 476 | assert session.status == RequestStatus.RUNNING |
| 477 | assert session in scheduler.running |
| 478 | assert session._all_token_ids == [1, 2, 3, 10] # Mutation happened here |
| 479 | |
| 480 | # CRITICAL ASSERTION: Cached prompt_token_ids must NOT have changed |
| 481 | assert ( |
| 482 | cached_state_cycle1["prompt_token_ids"] == original_cached_prompt_cycle1 |
| 483 | ), ( |
| 484 | f"ALIASING BUG DETECTED in Cycle 1! " |
| 485 | f"cached_state['prompt_token_ids'] was mutated from " |
| 486 | f"{original_cached_prompt_cycle1} to " |
| 487 | f"{cached_state_cycle1['prompt_token_ids']}. " |
| 488 | f"This means _create_new_request_data() didn't call .copy()!" |
| 489 | ) |
| 490 | assert cached_state_cycle1["prompt_token_ids"] is not session._all_token_ids, ( |
| 491 | "ALIASING BUG! cached_state['prompt_token_ids'] is the same object as " |
| 492 | "session._all_token_ids. They must be independent copies." |
| 493 | ) |
| 494 | |
| 495 | # ═══════════════════════════════════════════════════════════════════ |
| 496 | # CYCLE 2: Continue Decoding (Using Cached State) |
| 497 | # ═══════════════════════════════════════════════════════════════════ |
| 498 | |
| 499 | # Step 7: Schedule again - now request uses cached state |
| 500 | scheduler_output_cycle2 = scheduler.schedule() |
| 501 | |
| 502 | # Verify request is NOT in scheduled_new_reqs (already cached) |
| 503 | assert not scheduler_output_cycle2.scheduled_new_reqs |
| 504 | assert ( |
| 505 | session.request_id in scheduler_output_cycle2.scheduled_cached_reqs.req_ids |
| 506 | ) |
| 507 | assert ( |
| 508 | scheduler_output_cycle2.num_scheduled_tokens[session.request_id] == 1 |
| 509 | ) # Only the output token [10] |
| 510 | |
| 511 | # Step 8: Calculate num_tokens like gpu_model_runner.py:1284 does |
| 512 | # This is where the bug would manifest! |
| 513 | num_tokens_cycle2 = len(cached_state_cycle1["prompt_token_ids"]) + len( |
| 514 | cached_state_cycle1["output_token_ids"] |
| 515 | ) |
| 516 | |
| 517 | # CRITICAL ASSERTION: num_tokens must be correct (3 prompt + 1 output = 4) |
| 518 | # Without .copy(), cached_state["prompt_token_ids"] would be [1,2,3,10] |
| 519 | # and num_tokens would incorrectly be 5, causing the discard bug |
| 520 | expected_num_tokens_cycle2 = 4 |
| 521 | assert num_tokens_cycle2 == expected_num_tokens_cycle2, ( |
| 522 | f"DISCARD BUG WOULD TRIGGER! num_tokens calculation is wrong. " |
| 523 | f"Expected {expected_num_tokens_cycle2}, got {num_tokens_cycle2}. " |
| 524 | f"cached_state['prompt_token_ids'] = " |
| 525 | f"{cached_state_cycle1['prompt_token_ids']} (should be [1,2,3], not [1,2,3," |
| 526 | f"10]). Without .copy(), this would be 5 = len([1,2,3,10]) + len([10]). " |
| 527 | f"Discard logic would see: seq_lens={session.num_computed_tokens} " |
| 528 | f"< num_tokens={num_tokens_cycle2}, triggering incorrect discard!" |
| 529 | ) |
| 530 | |
| 531 | # Step 9-10: Model generates STOP_TOKEN, scheduler updates |
| 532 | output_token_2 = STOP_TOKEN |
| 533 | cached_state_cycle1["output_token_ids"].append(output_token_2) |
| 534 | |
| 535 | mro_cycle2 = ModelRunnerOutput( |
| 536 | req_ids=[session.request_id], |
| 537 | req_id_to_index={session.request_id: 0}, |
| 538 | sampled_token_ids=[[output_token_2]], |
| 539 | logprobs=None, |
| 540 | prompt_logprobs_dict={session.request_id: None}, |
| 541 | pooler_output=[], |
| 542 | ) |
| 543 | eco_dict_cycle2 = scheduler.update_from_output( |
| 544 | scheduler_output_cycle2, mro_cycle2 |
| 545 | ) |
| 546 | |
| 547 | # Step 11: Verify request transitioned to WAITING_FOR_STREAMING_REQ |
| 548 | eco_cycle2 = eco_dict_cycle2[session.client_index].outputs[0] |
| 549 | assert eco_cycle2.finish_reason == FinishReason.STOP |
| 550 | assert session.status == RequestStatus.WAITING_FOR_STREAMING_REQ |
| 551 | # Sessions paused for streaming input keep their KV blocks, so they |
| 552 | # live in the kv_holding_waiting queue. |
| 553 | assert session in scheduler.kv_holding_waiting |
| 554 | assert session._all_token_ids == [1, 2, 3, 10, STOP_TOKEN] |
| 555 | |
| 556 | # CRITICAL ASSERTION: Cached prompt_token_ids STILL must not have changed |
| 557 | assert cached_state_cycle1["prompt_token_ids"] == [1, 2, 3], ( |
| 558 | f"ALIASING BUG DETECTED in Cycle 2! " |
| 559 | f"cached_state['prompt_token_ids'] = " |
| 560 | f"{cached_state_cycle1['prompt_token_ids']} (should still be [1,2,3]). " |
| 561 | f"Mutations from update_from_output() leaked through!" |
| 562 | ) |
| 563 | |
| 564 | # ═══════════════════════════════════════════════════════════════════ |
| 565 | # CYCLE 3: New Streaming Request (Session Continuation) |
| 566 | # ═══════════════════════════════════════════════════════════════════ |
| 567 | |
| 568 | # Step 12: Add new streaming request with seq_id=1 |
| 569 | new_request = DummyRequest( |
| 570 | request_id="session", |
| 571 | prompt_token_ids=[4, 5], |
| 572 | ) |
| 573 | scheduler.add_request(new_request) |
| 574 | |
| 575 | # With the new streaming API, when session is in WAITING_FOR_STREAMING_REQ, |
| 576 | # the update is applied directly via _update_request_as_session (not queued). |
| 577 | # The session status becomes WAITING after the update is applied. |
| 578 | assert session.status == RequestStatus.WAITING |
| 579 | |
| 580 | # Step 13: Scheduler schedules the updated session |
| 581 | scheduler_output_cycle3 = scheduler.schedule() |
| 582 | |
| 583 | # Verify scheduler created NewRequestData with merged prompt_token_ids |
| 584 | assert len(scheduler_output_cycle3.scheduled_new_reqs) == 1 |
| 585 | assert ( |
| 586 | scheduler_output_cycle3.scheduled_new_reqs[0].prompt_token_ids |
| 587 | == session.prompt_token_ids |
| 588 | ) |
| 589 | assert ( |
| 590 | scheduler_output_cycle3.num_scheduled_tokens[session.request_id] == 2 |
| 591 | ) # Only new tokens [4, 5] |
| 592 | # Computed output tokens are kept (become part of prompt), only the |
| 593 | # final uncomputed sampled token (STOP_TOKEN) is discarded |
| 594 | assert session._all_token_ids == [1, 2, 3, 10, 4, 5] |
| 595 | assert session.prompt_token_ids == [1, 2, 3, 10, 4, 5] # Includes kept output |
| 596 | assert session._output_token_ids == [] # Output tokens are cleared |
| 597 | |
| 598 | # Step 14: Model runner caches NEW prompt_token_ids reference |
| 599 | # The model runner makes a copy of prompt_token_ids when creating |
| 600 | # CachedRequestState |
| 601 | new_req_data_cycle3 = scheduler_output_cycle3.scheduled_new_reqs[0] |
| 602 | cached_state_cycle3 = { |
| 603 | "req_id": session.request_id, |
| 604 | "prompt_token_ids": list( |
| 605 | new_req_data_cycle3.prompt_token_ids |
| 606 | ), # Explicit copy |
| 607 | "output_token_ids": [], |
| 608 | "num_computed_tokens": session.num_computed_tokens, |
| 609 | } |
| 610 | |
| 611 | # Step 15: FINAL CRITICAL VERIFICATION |
| 612 | # The old cached state from Cycle 1 must still be unchanged |
| 613 | assert cached_state_cycle1["prompt_token_ids"] == [1, 2, 3], ( |
| 614 | f"PERSISTENT ALIASING BUG! Even after new scheduling cycle, " |
| 615 | f"old cached_state was mutated to " |
| 616 | f"{cached_state_cycle1['prompt_token_ids']}. This proves the aliasing bug " |
| 617 | f"exists!" |
| 618 | ) |
| 619 | |
| 620 | # The new cached state must be independent |
| 621 | assert cached_state_cycle3["prompt_token_ids"] is not session._all_token_ids, ( |
| 622 | "ALIASING BUG in Cycle 3! Cached state is aliased to _all_token_ids." |
| 623 | ) |
| 624 | |
| 625 | # Both cached states must be independent of each other |
| 626 | assert ( |
| 627 | cached_state_cycle1["prompt_token_ids"] |
| 628 | is not cached_state_cycle3["prompt_token_ids"] |
| 629 | ), "Cached states from different cycles should be independent objects." |