| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import gc |
| 5 | from contextlib import nullcontext |
| 6 | from types import SimpleNamespace |
| 7 | from unittest.mock import Mock |
| 8 | |
| 9 | import numpy as np |
| 10 | import pytest |
| 11 | import torch |
| 12 | |
| 13 | import vllm.v1.worker.gpu_model_runner as gpu_model_runner_module |
| 14 | from vllm.config import ( |
| 15 | AttentionConfig, |
| 16 | CacheConfig, |
| 17 | ModelConfig, |
| 18 | ParallelConfig, |
| 19 | SchedulerConfig, |
| 20 | VllmConfig, |
| 21 | set_current_vllm_config, |
| 22 | ) |
| 23 | from vllm.config.reasoning import ReasoningConfig |
| 24 | from vllm.distributed.parallel_state import ( |
| 25 | init_distributed_environment, |
| 26 | initialize_model_parallel, |
| 27 | ) |
| 28 | from vllm.lora.layers import LoRAMappingType |
| 29 | from vllm.lora.request import LoRARequest |
| 30 | from vllm.model_executor.layers.attention import Attention |
| 31 | from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2 |
| 32 | from vllm.multimodal.inputs import MultiModalFeatureSpec, PlaceholderRange |
| 33 | from vllm.platforms import current_platform |
| 34 | from vllm.platforms.interface import Platform |
| 35 | from vllm.sampling_params import SamplingParams |
| 36 | from vllm.utils.mem_constants import GiB_bytes |
| 37 | from vllm.utils.system_utils import update_environment_variables |
| 38 | from vllm.utils.torch_utils import set_random_seed |
| 39 | from vllm.v1.attention.backend import MultipleOf |
| 40 | from vllm.v1.attention.backends.mla.indexer import DeepseekV32IndexerBackend |
| 41 | from vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse import ( |
| 42 | ROCMAiterMLASparseBackend, |
| 43 | ) |
| 44 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 45 | from vllm.v1.core.kv_cache_utils import estimate_max_model_len, get_kv_cache_configs |
| 46 | from vllm.v1.core.sched.output import CachedRequestData, NewRequestData, SchedulerOutput |
| 47 | from vllm.v1.kv_cache_interface import ( |
| 48 | FullAttentionSpec, |
| 49 | KVCacheConfig, |
| 50 | KVCacheGroupSpec, |
| 51 | KVCacheTensor, |
| 52 | ) |
| 53 | from vllm.v1.outputs import EMPTY_MODEL_RUNNER_OUTPUT |
| 54 | from vllm.v1.sample.metadata import SamplingMetadata |
| 55 | from vllm.v1.spec_decode.metadata import SpecDecodeMetadata |
| 56 | from vllm.v1.worker.block_table import ( |
| 57 | MultiGroupBlockTable, |
| 58 | SlotMappingMode, |
| 59 | get_block_table_width, |
| 60 | ) |
| 61 | from vllm.v1.worker.gpu.lora_utils import LoraState |
| 62 | from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache |
| 63 | from vllm.v1.worker.gpu.mm.lora import set_active_mm_loras |
| 64 | from vllm.v1.worker.gpu_input_batch import InputBatch |
| 65 | from vllm.v1.worker.gpu_model_runner import GPUModelRunner |
| 66 | from vllm.v1.worker.utils import select_common_block_size |
| 67 | |
| 68 | BLOCK_SIZE = 16 |
| 69 | NUM_BLOCKS = 10 |
| 70 | DEVICE_TYPE = current_platform.device_type |
| 71 | |
| 72 | |
| 73 | @pytest.fixture(autouse=True) |
| 74 | def _restore_default_dtype(): |
| 75 | """Several tests here set the process-wide default dtype to float16 and |
| 76 | previously leaked it, corrupting later float-sensitive tests in the same |
| 77 | pytest process (torch.randn silently produced fp16).""" |
| 78 | old = torch.get_default_dtype() |
| 79 | yield |
| 80 | torch.set_default_dtype(old) |
| 81 | |
| 82 | |
| 83 | def initialize_kv_cache(runner: GPUModelRunner): |
| 84 | """Only perform necessary steps in GPUModelRunner.initialize_kv_cache().""" |
| 85 | attn_spec = FullAttentionSpec( |
| 86 | block_size=BLOCK_SIZE, |
| 87 | num_kv_heads=runner.model_config.get_num_kv_heads(runner.parallel_config), |
| 88 | head_size=runner.model_config.get_head_size(), |
| 89 | dtype=runner.kv_cache_dtype, |
| 90 | ) |
| 91 | tensor_size = attn_spec.page_size_bytes * NUM_BLOCKS |
| 92 | kv_cache_config = KVCacheConfig( |
| 93 | num_blocks=NUM_BLOCKS, |
| 94 | kv_cache_tensors=[ |
| 95 | KVCacheTensor( |
| 96 | size=tensor_size, |
| 97 | layers=["layer.0"], |
| 98 | layer_stride=tensor_size, |
| 99 | block_stride=attn_spec.page_size_bytes, |
| 100 | ), |
| 101 | ], |
| 102 | kv_cache_groups=[ |
| 103 | KVCacheGroupSpec(layer_names=["layer.0"], kv_cache_spec=attn_spec) |
| 104 | ], |
| 105 | ) |
| 106 | runner.kv_cache_config = kv_cache_config |
| 107 | runner.input_batch = InputBatch( |
| 108 | max_num_reqs=runner.max_num_reqs, |
| 109 | max_model_len=runner.max_model_len, |
| 110 | max_num_batched_tokens=runner.max_num_tokens, |
| 111 | device=runner.device, |
| 112 | vocab_size=runner.model_config.get_vocab_size(), |
| 113 | block_sizes=[kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size], |
| 114 | kernel_block_sizes=[ |
| 115 | kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size |
| 116 | ], |
| 117 | max_num_blocks_per_req=[NUM_BLOCKS], |
| 118 | ) |
| 119 | runner.initialize_attn_backend(kv_cache_config) |
| 120 | |
| 121 | |
| 122 | def get_vllm_config(): |
| 123 | model_config = ModelConfig( |
| 124 | model="facebook/opt-125m", |
| 125 | dtype="float16", |
| 126 | seed=42, |
| 127 | ) |
| 128 | scheduler_config = SchedulerConfig( |
| 129 | max_num_seqs=10, |
| 130 | max_num_batched_tokens=512, |
| 131 | max_model_len=512, |
| 132 | is_encoder_decoder=model_config.is_encoder_decoder, |
| 133 | ) |
| 134 | cache_config = CacheConfig( |
| 135 | block_size=BLOCK_SIZE, |
| 136 | gpu_memory_utilization=0.9, |
| 137 | cache_dtype="auto", |
| 138 | ) |
| 139 | cache_config.kv_cache_layout = "LBNHC" |
| 140 | parallel_config = ParallelConfig() |
| 141 | vllm_config = VllmConfig( |
| 142 | model_config=model_config, |
| 143 | cache_config=cache_config, |
| 144 | scheduler_config=scheduler_config, |
| 145 | parallel_config=parallel_config, |
| 146 | ) |
| 147 | return vllm_config |
| 148 | |
| 149 | |
| 150 | @pytest.mark.parametrize("gc_initially_enabled", [True, False]) |
| 151 | def test_freeze_gc_disables_and_restores_automatic_gc( |
| 152 | monkeypatch: pytest.MonkeyPatch, |
| 153 | gc_initially_enabled: bool, |
| 154 | ): |
| 155 | monkeypatch.setenv("VLLM_ENABLE_CUDAGRAPH_GC", "0") |
| 156 | original_gc_state = gc.isenabled() |
| 157 | try: |
| 158 | if gc_initially_enabled: |
| 159 | gc.enable() |
| 160 | else: |
| 161 | gc.disable() |
| 162 | |
| 163 | with ( |
| 164 | pytest.raises(RuntimeError, match="capture failed"), |
| 165 | GPUModelRunner._freeze_gc(), |
| 166 | ): |
| 167 | assert not gc.isenabled() |
| 168 | raise RuntimeError("capture failed") |
| 169 | |
| 170 | assert gc.isenabled() is gc_initially_enabled |
| 171 | finally: |
| 172 | if original_gc_state: |
| 173 | gc.enable() |
| 174 | else: |
| 175 | gc.disable() |
| 176 | |
| 177 | |
| 178 | def test_prepare_padding_mask_marks_sequence_parallel_padding(): |
| 179 | runner = GPUModelRunner.__new__(GPUModelRunner) |
| 180 | runner.is_padding = torch.empty(8, dtype=torch.bool) |
| 181 | |
| 182 | mask = runner._prepare_padding_mask(1, 8) |
| 183 | |
| 184 | assert mask.tolist() == [False, True, True, True, True, True, True, True] |
| 185 | assert mask.data_ptr() == runner.is_padding.data_ptr() |
| 186 | |
| 187 | mask = runner._prepare_padding_mask(0, 8) |
| 188 | |
| 189 | assert mask.all() |
| 190 | |
| 191 | |
| 192 | @pytest.fixture |
| 193 | def model_runner(): |
| 194 | vllm_config = get_vllm_config() |
| 195 | with set_current_vllm_config(vllm_config): |
| 196 | model_config = vllm_config.model_config |
| 197 | num_heads = model_config.get_num_kv_heads(vllm_config.parallel_config) |
| 198 | head_size = model_config.get_head_size() |
| 199 | vllm_config.compilation_config.static_forward_context["layer.0"] = Attention( |
| 200 | num_heads, head_size, 0.1 |
| 201 | ) |
| 202 | runner = GPUModelRunner(vllm_config, DEVICE_TYPE) |
| 203 | initialize_kv_cache(runner) |
| 204 | yield runner |
| 205 | |
| 206 | |
| 207 | model_runner_2 = model_runner |
| 208 | |
| 209 | |
| 210 | def _schedule_new_request(*req_ids: str) -> SchedulerOutput: |
| 211 | new_reqs = [] |
| 212 | num_scheduled_tokens = {} |
| 213 | total_num_scheduled_tokens = 0 |
| 214 | for req_id in req_ids: |
| 215 | new_reqs.append( |
| 216 | NewRequestData( |
| 217 | req_id=req_id, |
| 218 | prompt_token_ids=[1, 2, 3], |
| 219 | mm_features=[], |
| 220 | sampling_params=SamplingParams(), |
| 221 | pooling_params=None, |
| 222 | block_ids=([0],), |
| 223 | num_computed_tokens=0, |
| 224 | lora_request=None, |
| 225 | ) |
| 226 | ) |
| 227 | num_scheduled_tokens[req_id] = 3 |
| 228 | total_num_scheduled_tokens += num_scheduled_tokens[req_id] |
| 229 | |
| 230 | return SchedulerOutput( |
| 231 | scheduled_new_reqs=new_reqs, |
| 232 | scheduled_cached_reqs=CachedRequestData.make_empty(), |
| 233 | num_scheduled_tokens=num_scheduled_tokens, |
| 234 | total_num_scheduled_tokens=total_num_scheduled_tokens, |
| 235 | scheduled_spec_decode_tokens={}, |
| 236 | scheduled_encoder_inputs={}, |
| 237 | num_common_prefix_blocks=[], |
| 238 | finished_req_ids=set(), |
| 239 | free_encoder_mm_hashes=[], |
| 240 | ) |
| 241 | |
| 242 | |
| 243 | def _schedule_cached_requests( |
| 244 | req_ids: list[str], |
| 245 | num_scheduled_tokens: dict[str, int], |
| 246 | new_token_ids: list[list[int]], |
| 247 | num_computed_tokens: list[int], |
| 248 | num_output_tokens: list[int], |
| 249 | ) -> SchedulerOutput: |
| 250 | return SchedulerOutput( |
| 251 | scheduled_new_reqs=[], |
| 252 | scheduled_cached_reqs=CachedRequestData( |
| 253 | req_ids=req_ids, |
| 254 | resumed_req_ids=set(), |
| 255 | new_token_ids=new_token_ids, |
| 256 | all_token_ids={}, |
| 257 | new_block_ids=[None] * len(req_ids), |
| 258 | num_computed_tokens=num_computed_tokens, |
| 259 | num_output_tokens=num_output_tokens, |
| 260 | ), |
| 261 | num_scheduled_tokens=num_scheduled_tokens, |
| 262 | total_num_scheduled_tokens=sum(num_scheduled_tokens.values()), |
| 263 | scheduled_spec_decode_tokens={}, |
| 264 | scheduled_encoder_inputs={}, |
| 265 | num_common_prefix_blocks=[], |
| 266 | finished_req_ids=set(), |
| 267 | free_encoder_mm_hashes=[], |
| 268 | ) |
| 269 | |
| 270 | |
| 271 | def _is_req_scheduled(model_runner, req_id: str) -> bool: |
| 272 | return req_id in model_runner.input_batch.req_id_to_index |
| 273 | |
| 274 | |
| 275 | def _is_req_added(model_runner, req_id: str) -> bool: |
| 276 | return req_id in model_runner.requests |
| 277 | |
| 278 | |
| 279 | def _is_sampling_metadata_changed( |
| 280 | model_runner, sampling_metadata_before: SamplingMetadata |
| 281 | ): |
| 282 | return model_runner.input_batch.sampling_metadata is not (sampling_metadata_before) |
| 283 | |
| 284 | |
| 285 | def _is_req_state_block_table_match(model_runner, req_id: str) -> bool: |
| 286 | req_index = model_runner.input_batch.req_id_to_index[req_id] |
| 287 | block_table = model_runner.input_batch.block_table[0] |
| 288 | req_state = model_runner.requests[req_id] |
| 289 | if block_table.num_blocks_per_row[req_index] != len(req_state.block_ids[0]): |
| 290 | return False |
| 291 | num_blocks = block_table.num_blocks_per_row[req_index] |
| 292 | return ( |
| 293 | block_table.block_table.np[req_index, :num_blocks] == req_state.block_ids[0] |
| 294 | ).all() |
| 295 | |
| 296 | |
| 297 | def _make_mock_backend_for_kernel_block_size( |
| 298 | supported_sizes: list[int | MultipleOf], |
| 299 | ): |
| 300 | class _MockBackend: |
| 301 | @staticmethod |
| 302 | def get_name(): |
| 303 | return "MOCK" |
| 304 | |
| 305 | @staticmethod |
| 306 | def get_supported_kernel_block_sizes(): |
| 307 | return supported_sizes |
| 308 | |
| 309 | return _MockBackend() |
| 310 | |
| 311 | |
| 312 | def _make_kv_cache_spec() -> FullAttentionSpec: |
| 313 | return FullAttentionSpec(block_size=1, num_kv_heads=1, head_size=1, dtype="float16") |
| 314 | |
| 315 | |
| 316 | def test_select_common_block_size_prefers_manager_block_size(): |
| 317 | backend_a = _make_mock_backend_for_kernel_block_size([MultipleOf(32)]) |
| 318 | backend_b = _make_mock_backend_for_kernel_block_size([64, MultipleOf(16)]) |
| 319 | |
| 320 | selected_size = select_common_block_size(128, [backend_a, backend_b]) |
| 321 | assert selected_size == 128 |
| 322 | |
| 323 | |
| 324 | def test_select_common_block_size_uses_largest_shared_int(): |
| 325 | backend_a = _make_mock_backend_for_kernel_block_size([128, 64]) |
| 326 | backend_b = _make_mock_backend_for_kernel_block_size([64, 32]) |
| 327 | |
| 328 | selected_size = select_common_block_size(256, [backend_a, backend_b]) |
| 329 | assert selected_size == 64 |
| 330 | |
| 331 | |
| 332 | def test_select_common_block_size_without_active_backends_uses_manager_size(): |
| 333 | assert select_common_block_size(256, []) == 256 |
| 334 | |
| 335 | |
| 336 | def test_select_common_block_size_accepts_rocm_sparse_block_size_16(monkeypatch): |
| 337 | monkeypatch.setattr(current_platform, "is_rocm", lambda: True) |
| 338 | |
| 339 | selected_size = select_common_block_size( |
| 340 | 16, |
| 341 | [DeepseekV32IndexerBackend, ROCMAiterMLASparseBackend], |
| 342 | ) |
| 343 | assert selected_size == 16 |
| 344 | |
| 345 | |
| 346 | def test_reasoning_config_without_custom_logitsprocs_does_not_need_output_token_ids( |
| 347 | dist_init, |
| 348 | ): |
| 349 | vllm_config = get_vllm_config() |
| 350 | assert vllm_config.model_config.logits_processors is None |
| 351 | reasoning_config = ReasoningConfig( |
| 352 | reasoning_start_str="<think>", reasoning_end_str="</think>" |
| 353 | ) |
| 354 | reasoning_config._reasoning_start_token_ids = [1] |
| 355 | reasoning_config._reasoning_end_token_ids = [2] |
| 356 | vllm_config.reasoning_config = reasoning_config |
| 357 | |
| 358 | with set_current_vllm_config(vllm_config): |
| 359 | model_config = vllm_config.model_config |
| 360 | num_heads = model_config.get_num_kv_heads(vllm_config.parallel_config) |
| 361 | head_size = model_config.get_head_size() |
| 362 | vllm_config.compilation_config.static_forward_context["layer.0"] = Attention( |
| 363 | num_heads, head_size, 0.1 |
| 364 | ) |
| 365 | runner = GPUModelRunner(vllm_config, torch.device("cpu")) |
| 366 | |
| 367 | assert runner.input_batch.thinking_budget_state_holder is not None |
| 368 | assert runner.input_batch.logitsprocs_need_output_token_ids is False |
| 369 | |
| 370 | |
| 371 | @pytest.mark.skip_global_cleanup |
| 372 | @pytest.mark.parametrize( |
| 373 | ("world_size", "is_last_rank", "expected_calls"), |
| 374 | [(1, True, 0), (2, True, 0), (2, False, 1)], |
| 375 | ) |
| 376 | def test_sample_tokens_receives_pp_sampled_ids_only_on_non_last_rank( |
| 377 | monkeypatch: pytest.MonkeyPatch, |
| 378 | world_size: int, |
| 379 | is_last_rank: bool, |
| 380 | expected_calls: int, |
| 381 | ): |
| 382 | runner = GPUModelRunner.__new__(GPUModelRunner) |
| 383 | runner.execute_model_state = None |
| 384 | runner.kv_connector_output = None |
| 385 | runner.use_async_scheduling = True |
| 386 | receive_calls = 0 |
| 387 | |
| 388 | def receive_prev_sampled_token_ids(): |
| 389 | nonlocal receive_calls |
| 390 | receive_calls += 1 |
| 391 | |
| 392 | runner._pp_receive_prev_sampled_token_ids_to_input_batch = ( |
| 393 | receive_prev_sampled_token_ids |
| 394 | ) |
| 395 | monkeypatch.setattr( |
| 396 | gpu_model_runner_module, |
| 397 | "get_pp_group", |
| 398 | lambda: SimpleNamespace(world_size=world_size, is_last_rank=is_last_rank), |
| 399 | ) |
| 400 | |
| 401 | output = GPUModelRunner.sample_tokens(runner, None) |
| 402 | assert output in (EMPTY_MODEL_RUNNER_OUTPUT, None) |
| 403 | assert receive_calls == expected_calls |
| 404 | |
| 405 | |
| 406 | @pytest.mark.skip_global_cleanup |
| 407 | def test_sample_tokens_skips_pp_group_lookup_without_async_scheduling( |
| 408 | monkeypatch: pytest.MonkeyPatch, |
| 409 | ): |
| 410 | runner = GPUModelRunner.__new__(GPUModelRunner) |
| 411 | runner.execute_model_state = None |
| 412 | runner.kv_connector_output = None |
| 413 | runner.use_async_scheduling = False |
| 414 | |
| 415 | monkeypatch.setattr( |
| 416 | gpu_model_runner_module, |
| 417 | "get_pp_group", |
| 418 | pytest.fail, |
| 419 | ) |
| 420 | |
| 421 | output = GPUModelRunner.sample_tokens(runner, None) |
| 422 | assert output in (EMPTY_MODEL_RUNNER_OUTPUT, None) |
| 423 | |
| 424 | |
| 425 | def test_select_common_block_size_no_valid_option(): |
| 426 | backend_a = _make_mock_backend_for_kernel_block_size([64]) |
| 427 | backend_b = _make_mock_backend_for_kernel_block_size([MultipleOf(16)]) |
| 428 | |
| 429 | with pytest.raises(ValueError): |
| 430 | select_common_block_size(48, [backend_a, backend_b]) |
| 431 | |
| 432 | |
| 433 | def _mock_backend(supported: list, *, exact: bool = False): |
| 434 | """Backend accepting multiples of ``supported``, or only those exact sizes |
| 435 | when ``exact`` (as CPU_MLA does).""" |
| 436 | from vllm.v1.attention.backend import AttentionBackend |
| 437 | |
| 438 | class _MockBackendCls(AttentionBackend): |
| 439 | @staticmethod |
| 440 | def get_name() -> str: |
| 441 | return "MOCK_EXACT" if exact else "MOCK" |
| 442 | |
| 443 | @staticmethod |
| 444 | def get_supported_kernel_block_sizes(): |
| 445 | return list(supported) |
| 446 | |
| 447 | if exact: |
| 448 | |
| 449 | @classmethod |
| 450 | def supports_block_size(cls, block_size: int | None) -> bool: |
| 451 | return block_size is None or block_size in supported |
| 452 | |
| 453 | return _MockBackendCls |
| 454 | |
| 455 | |
| 456 | @pytest.mark.parametrize( |
| 457 | "backends,expected", |
| 458 | [ |
| 459 | # A lone backend keeps the default; MultipleOf(8) makes a |
| 460 | # minimum-based regression return 8 instead of 16. |
| 461 | ([[MultipleOf(8)]], 16), |
| 462 | # Sparse-MLA main backend (32 or 64) beside an indexer supporting only |
| 463 | # 64: the main backend alone picks 32, which the indexer rejects at |
| 464 | # select_common_block_size time. |
| 465 | ([[32, 64], [64]], 64), |
| 466 | ([[32, 64]], 32), |
| 467 | # Neither divides the other, so only the LCM satisfies both. |
| 468 | ([[32], [48]], 96), |
| 469 | ], |
| 470 | ) |
| 471 | def test_preferred_block_size_satisfies_every_backend(backends, expected): |
| 472 | classes = [_mock_backend(s) for s in backends] |
| 473 | assert Platform._preferred_block_size_for_backends(classes, 16, None) == expected |
| 474 | |
| 475 | |
| 476 | def test_preferred_block_size_searches_past_an_exact_size_backend(): |
| 477 | # Extending greedily picks lcm(16, 32) = 32, which the exact backend |
| 478 | # rejects; 96 is accepted by both. |
| 479 | classes = [_mock_backend([16, 96], exact=True), _mock_backend([MultipleOf(32)])] |
| 480 | assert Platform._preferred_block_size_for_backends(classes, 16, None) == 96 |
| 481 | |
| 482 | |
| 483 | def test_preferred_block_size_rejects_backends_with_no_common_size(): |
| 484 | classes = [_mock_backend([16], exact=True), _mock_backend([MultipleOf(64)])] |
| 485 | with pytest.raises(ValueError, match="share no supported KV cache block size"): |
| 486 | Platform._preferred_block_size_for_backends(classes, 16, None) |
| 487 | |
| 488 | |
| 489 | def test_set_active_mm_loras_builds_tower_and_connector_mappings(): |
| 490 | model = Mock() |
| 491 | model.get_mm_lora_token_counts.side_effect = ( |
| 492 | lambda *, modality, mm_kwargs, num_mm_embeds: ( |
| 493 | num_mm_embeds + 1, |
| 494 | num_mm_embeds + 11, |
| 495 | ) |
| 496 | ) |
| 497 | model.get_mm_mapping.return_value = SimpleNamespace(connector=True) |
| 498 | |
| 499 | lora_manager = Mock() |
| 500 | lora_manager.supports_tower_connector_lora.return_value = True |
| 501 | |
| 502 | encoder_cache = EncoderCache() |
| 503 | encoder_cache.mm_features["req-with-lora"] = [ |
| 504 | MultiModalFeatureSpec( |
| 505 | data=None, |
| 506 | modality="image", |
| 507 | identifier="img-0", |
| 508 | mm_position=PlaceholderRange(offset=0, length=2), |
| 509 | ), |
| 510 | MultiModalFeatureSpec( |
| 511 | data=None, |
| 512 | modality="image", |
| 513 | identifier="img-1", |
| 514 | mm_position=PlaceholderRange(offset=2, length=3), |
| 515 | ), |
| 516 | ] |
| 517 | encoder_cache.mm_features["req-no-lora"] = [ |
| 518 | MultiModalFeatureSpec( |
| 519 | data=None, |
| 520 | modality="image", |
| 521 | identifier="img-2", |
| 522 | mm_position=PlaceholderRange(offset=0, length=1), |
| 523 | ) |
| 524 | ] |
| 525 | |
| 526 | lora_state = LoraState(max_num_reqs=4) |
| 527 | lora_request = LoRARequest("vision-lora", 7, "/tmp/vision-lora") |
| 528 | lora_state.add_request("req-with-lora", 0, lora_request) |
| 529 | lora_state.add_request("req-no-lora", 1, None) |
| 530 | |
| 531 | set_active_mm_loras( |
| 532 | model=model, |
| 533 | lora_manager=lora_manager, |
| 534 | encoder_cache=encoder_cache, |
| 535 | req_id_to_index={ |
| 536 | "req-with-lora": 0, |
| 537 | "req-no-lora": 1, |
| 538 | }, |
| 539 | lora_state=lora_state, |
| 540 | scheduled_encoder_inputs={ |
| 541 | "req-with-lora": [1, 0], |
| 542 | "req-no-lora": [0], |
| 543 | "missing-req": [0], |
| 544 | }, |
| 545 | ) |
| 546 | |
| 547 | assert lora_manager.set_active_adapters.call_count == 2 |
| 548 | |
| 549 | tower_requests, tower_mapping = lora_manager.set_active_adapters.call_args_list[ |
| 550 | 0 |
| 551 | ].args |
| 552 | assert tower_requests == {lora_request} |
| 553 | assert tower_mapping.type is LoRAMappingType.TOWER |
| 554 | assert tower_mapping.prompt_mapping == (7, 7, 0) |
| 555 | assert tower_mapping.index_mapping == (7, 7, 7, 7, 7, 7, 7, 0, 0) |
| 556 | |
| 557 | connector_requests, connector_mapping = ( |
| 558 | lora_manager.set_active_adapters.call_args_list[1].args |
| 559 | ) |
| 560 | assert connector_requests == {lora_request} |
| 561 | assert connector_mapping.type is LoRAMappingType.CONNECTOR |
| 562 | assert connector_mapping.prompt_mapping == (7, 7, 0) |
| 563 | assert connector_mapping.index_mapping == ((7,) * 14 + (7,) * 13 + (0,) * 12) |
| 564 | |
| 565 | |
| 566 | def test_update_states_new_request(model_runner, dist_init): |
| 567 | req_id = "req_0" |
| 568 | |
| 569 | # new req |
| 570 | scheduler_output = _schedule_new_request(req_id) |
| 571 | |
| 572 | metadata_before = model_runner.input_batch.sampling_metadata |
| 573 | model_runner._update_states(scheduler_output) |
| 574 | assert _is_sampling_metadata_changed(model_runner, metadata_before) |
| 575 | assert _is_req_added(model_runner, req_id) |
| 576 | assert _is_req_scheduled(model_runner, req_id) |
| 577 | assert _is_req_state_block_table_match(model_runner, req_id) |
| 578 | |
| 579 | |
| 580 | def test_update_states_request_finished(model_runner, dist_init): |
| 581 | req_id = "req_0" |
| 582 | |
| 583 | # new req |
| 584 | scheduler_output = _schedule_new_request(req_id) |
| 585 | |
| 586 | model_runner._update_states(scheduler_output) |
| 587 | assert _is_req_added(model_runner, req_id) |
| 588 | assert _is_req_scheduled(model_runner, req_id) |
| 589 | |
| 590 | # finish req |
| 591 | scheduler_output = SchedulerOutput( |
| 592 | scheduled_new_reqs=[], |
| 593 | scheduled_cached_reqs=CachedRequestData.make_empty(), |
| 594 | num_scheduled_tokens={}, |
| 595 | total_num_scheduled_tokens=0, |
| 596 | scheduled_spec_decode_tokens={}, |
| 597 | scheduled_encoder_inputs={}, |
| 598 | num_common_prefix_blocks=[], |
| 599 | finished_req_ids={req_id}, |
| 600 | free_encoder_mm_hashes=[], |
| 601 | ) |
| 602 | |
| 603 | metadata_before = model_runner.input_batch.sampling_metadata |
| 604 | model_runner._update_states(scheduler_output) |
| 605 | assert _is_sampling_metadata_changed(model_runner, metadata_before) |
| 606 | assert not _is_req_added(model_runner, req_id) |
| 607 | assert not _is_req_scheduled(model_runner, req_id) |
| 608 | |
| 609 | |
| 610 | def test_update_states_request_resumed(model_runner, dist_init): |
| 611 | req_id = "req_0" |
| 612 | |
| 613 | # new req |
| 614 | scheduler_output = _schedule_new_request(req_id) |
| 615 | |
| 616 | model_runner._update_states(scheduler_output) |
| 617 | assert _is_req_added(model_runner, req_id) |
| 618 | assert _is_req_scheduled(model_runner, req_id) |
| 619 | |
| 620 | # unschedule req |
| 621 | scheduler_output = SchedulerOutput( |
| 622 | scheduled_new_reqs=[], |
| 623 | scheduled_cached_reqs=CachedRequestData.make_empty(), |
| 624 | num_scheduled_tokens={}, |
| 625 | total_num_scheduled_tokens=0, |
| 626 | scheduled_spec_decode_tokens={}, |
| 627 | scheduled_encoder_inputs={}, |
| 628 | num_common_prefix_blocks=[], |
| 629 | finished_req_ids=set(), |
| 630 | free_encoder_mm_hashes=[], |
| 631 | ) |
| 632 | |
| 633 | model_runner._update_states(scheduler_output) |
| 634 | assert _is_req_added(model_runner, req_id) |
| 635 | assert not _is_req_scheduled(model_runner, req_id) |
| 636 | |
| 637 | # resume req |
| 638 | cached_req_data = CachedRequestData( |
| 639 | req_ids=[req_id], |
| 640 | resumed_req_ids=set(), |
| 641 | new_token_ids=[[]], |
| 642 | all_token_ids={}, |
| 643 | new_block_ids=[([0],)], |
| 644 | num_computed_tokens=[0], |
| 645 | num_output_tokens=[0], |
| 646 | ) |
| 647 | |
| 648 | scheduler_output = SchedulerOutput( |
| 649 | scheduled_new_reqs=[], |
| 650 | scheduled_cached_reqs=cached_req_data, |
| 651 | num_scheduled_tokens={req_id: 1}, |
| 652 | total_num_scheduled_tokens=1, |
| 653 | scheduled_spec_decode_tokens={}, |
| 654 | scheduled_encoder_inputs={}, |
| 655 | num_common_prefix_blocks=[], |
| 656 | finished_req_ids=set(), |
| 657 | free_encoder_mm_hashes=[], |
| 658 | ) |
| 659 | |
| 660 | metadata_before = model_runner.input_batch.sampling_metadata |
| 661 | model_runner._update_states(scheduler_output) |
| 662 | assert _is_sampling_metadata_changed(model_runner, metadata_before) |
| 663 | assert _is_req_added(model_runner, req_id) |
| 664 | assert _is_req_scheduled(model_runner, req_id) |
| 665 | assert _is_req_state_block_table_match(model_runner, req_id) |
| 666 | |
| 667 | |
| 668 | def test_get_nans_in_logits(model_runner, dist_init, monkeypatch): |
| 669 | monkeypatch.setattr(gpu_model_runner_module.envs, "VLLM_RAISE_ON_LOGIT_NANS", False) |
| 670 | req_ids = ("req_0", "req_1") |
| 671 | |
| 672 | scheduler_output = _schedule_new_request(*req_ids) |
| 673 | model_runner._update_states(scheduler_output) |
| 674 | |
| 675 | logits = torch.tensor( |
| 676 | [ |
| 677 | [1.0, 2.0, 3.0], |
| 678 | [3.0, 2.0, 1.0], |
| 679 | ], |
| 680 | device=DEVICE_TYPE, |
| 681 | ) |
| 682 | result = model_runner._get_nans_in_logits(logits) |
| 683 | assert result == {"req_0": 0, "req_1": 0} |
| 684 | |
| 685 | logits = torch.tensor( |
| 686 | [ |
| 687 | [1.0, float("nan"), 3.0], |
| 688 | [4.0, float("nan"), float("nan")], |
| 689 | ], |
| 690 | device=DEVICE_TYPE, |
| 691 | ) |
| 692 | result = model_runner._get_nans_in_logits(logits) |
| 693 | assert result == {"req_0": 1, "req_1": 2} |
| 694 | |
| 695 | logits = torch.tensor( |
| 696 | [ |
| 697 | [1.0, 2.0, 3.0], |
| 698 | [4.0, float("nan"), float("nan")], |
| 699 | ], |
| 700 | device=DEVICE_TYPE, |
| 701 | ) |
| 702 | result = model_runner._get_nans_in_logits(logits) |
| 703 | assert result == {"req_0": 0, "req_1": 2} |
| 704 | |
| 705 | result = model_runner._get_nans_in_logits(logits=None) |
| 706 | assert result == {"req_0": 0, "req_1": 0} |
| 707 | |
| 708 | logits = torch.tensor( |
| 709 | [ |
| 710 | [1.0, float("nan"), 3.0], |
| 711 | ], |
| 712 | device=DEVICE_TYPE, |
| 713 | ) |
| 714 | result = model_runner._get_nans_in_logits(logits) |
| 715 | assert result == {"req_0": 1, "req_1": 0} |
| 716 | |
| 717 | logits = torch.tensor( |
| 718 | [ |
| 719 | [float("nan"), float("nan"), 2.0], |
| 720 | [1.0, 2.0, 3.0], |
| 721 | [float("nan"), 2.0, 3.0], |
| 722 | ], |
| 723 | device=DEVICE_TYPE, |
| 724 | ) |
| 725 | result = model_runner._get_nans_in_logits(logits) |
| 726 | assert result == {"req_0": 2, "req_1": 0} |
| 727 | |
| 728 | |
| 729 | def test_update_states_no_changes(model_runner, dist_init): |
| 730 | req_id = "req_0" |
| 731 | |
| 732 | # new req |
| 733 | scheduler_output = _schedule_new_request(req_id) |
| 734 | |
| 735 | model_runner._update_states(scheduler_output) |
| 736 | assert _is_req_added(model_runner, req_id) |
| 737 | assert _is_req_scheduled(model_runner, req_id) |
| 738 | |
| 739 | # schedule req |
| 740 | scheduler_output = SchedulerOutput( |
| 741 | scheduled_new_reqs=[], |
| 742 | scheduled_cached_reqs=CachedRequestData.make_empty(), |
| 743 | num_scheduled_tokens={req_id: 1}, |
| 744 | total_num_scheduled_tokens=1, |
| 745 | scheduled_spec_decode_tokens={}, |
| 746 | scheduled_encoder_inputs={}, |
| 747 | num_common_prefix_blocks=[], |
| 748 | finished_req_ids=set(), |
| 749 | free_encoder_mm_hashes=[], |
| 750 | ) |
| 751 | |
| 752 | metadata_before = model_runner.input_batch.sampling_metadata |
| 753 | model_runner._update_states(scheduler_output) |
| 754 | assert not _is_sampling_metadata_changed(model_runner, metadata_before) |
| 755 | assert _is_req_added(model_runner, req_id) |
| 756 | assert _is_req_scheduled(model_runner, req_id) |
| 757 | assert _is_req_state_block_table_match(model_runner, req_id) |
| 758 | |
| 759 | |
| 760 | def test_update_states_request_unscheduled(model_runner, dist_init): |
| 761 | req_ids = ("req_0", "req_1") |
| 762 | |
| 763 | # new reqs |
| 764 | scheduler_output = _schedule_new_request(*req_ids) |
| 765 | |
| 766 | model_runner._update_states(scheduler_output) |
| 767 | |
| 768 | assert _is_req_added(model_runner, req_ids[0]) |
| 769 | assert _is_req_scheduled(model_runner, req_ids[0]) |
| 770 | |
| 771 | assert _is_req_added(model_runner, req_ids[1]) |
| 772 | assert _is_req_scheduled(model_runner, req_ids[1]) |
| 773 | |
| 774 | # unschedule req_1 |
| 775 | scheduler_output = SchedulerOutput( |
| 776 | scheduled_new_reqs=[], |
| 777 | scheduled_cached_reqs=CachedRequestData.make_empty(), |
| 778 | num_scheduled_tokens={req_ids[0]: 1}, |
| 779 | total_num_scheduled_tokens=1, |
| 780 | scheduled_spec_decode_tokens={}, |
| 781 | scheduled_encoder_inputs={}, |
| 782 | num_common_prefix_blocks=[], |
| 783 | finished_req_ids=set(), |
| 784 | free_encoder_mm_hashes=[], |
| 785 | ) |
| 786 | |
| 787 | metadata_before = model_runner._update_states(scheduler_output) |
| 788 | assert _is_sampling_metadata_changed(model_runner, metadata_before) |
| 789 | |
| 790 | assert _is_req_added(model_runner, req_ids[0]) |
| 791 | assert _is_req_scheduled(model_runner, req_ids[0]) |
| 792 | |
| 793 | assert _is_req_added(model_runner, req_ids[1]) |
| 794 | assert not _is_req_scheduled(model_runner, req_ids[1]) |
| 795 | |
| 796 | |
| 797 | def test_update_states_pp_non_async_multi_request_keeps_token_buffers_consistent( |
| 798 | model_runner, model_runner_2, dist_init, monkeypatch |
| 799 | ): |
| 800 | req_ids = ["req_0", "req_1"] |
| 801 | non_last_runner = model_runner |
| 802 | last_runner = model_runner_2 |
| 803 | non_last_runner.use_async_scheduling = False |
| 804 | last_runner.use_async_scheduling = False |
| 805 | |
| 806 | # Both ranks start from the same request set. |
| 807 | monkeypatch.setattr( |
| 808 | "vllm.v1.worker.gpu_model_runner.get_pp_group", |
| 809 | lambda: SimpleNamespace(is_last_rank=False, world_size=2), |
| 810 | ) |
| 811 | non_last_runner._update_states(_schedule_new_request(*req_ids)) |
| 812 | last_runner._update_states(_schedule_new_request(*req_ids)) |
| 813 | |
| 814 | sampled_by_last_rank = {req_ids[0]: 101, req_ids[1]: 201} |
| 815 | # Emulate last-rank bookkeeping result from previous step: |
| 816 | # sampled tokens already cached in CPU token buffers. |
| 817 | for req_id, token_id in sampled_by_last_rank.items(): |
| 818 | req_index = last_runner.input_batch.req_id_to_index[req_id] |
| 819 | start_idx = int(last_runner.input_batch.num_tokens_no_spec[req_index]) |
| 820 | end_idx = start_idx + 1 |
| 821 | last_runner.input_batch.token_ids_cpu[req_index, start_idx:end_idx] = [token_id] |
| 822 | last_runner.input_batch.is_token_ids[req_index, start_idx:end_idx] = True |
| 823 | last_runner.input_batch.num_tokens_no_spec[req_index] = end_idx |
| 824 | last_runner.requests[req_id].output_token_ids.append(token_id) |
| 825 | |
| 826 | scheduler_output = _schedule_cached_requests( |
| 827 | req_ids=req_ids, |
| 828 | num_scheduled_tokens={req_ids[0]: 1, req_ids[1]: 1}, |
| 829 | new_token_ids=[[101], [201]], |
| 830 | num_computed_tokens=[3, 3], # prompt tokens only |
| 831 | num_output_tokens=[1, 1], |
| 832 | ) |
| 833 | # non-last rank appends new_token_ids in _update_states. |
| 834 | monkeypatch.setattr( |
| 835 | "vllm.v1.worker.gpu_model_runner.get_pp_group", |
| 836 | lambda: SimpleNamespace(is_last_rank=False, world_size=2), |
| 837 | ) |
| 838 | non_last_runner._update_states(scheduler_output) |
| 839 | # last rank should keep its already-bookkept CPU buffers unchanged. |
| 840 | monkeypatch.setattr( |
| 841 | "vllm.v1.worker.gpu_model_runner.get_pp_group", |
| 842 | lambda: SimpleNamespace(is_last_rank=True, world_size=2), |
| 843 | ) |
| 844 | last_runner._update_states(scheduler_output) |
| 845 | |
| 846 | # Verify consistency between PP ranks after _update_states. |
| 847 | for req_id in req_ids: |
| 848 | non_last_idx = non_last_runner.input_batch.req_id_to_index[req_id] |
| 849 | last_idx = last_runner.input_batch.req_id_to_index[req_id] |
| 850 | non_last_len = int(non_last_runner.input_batch.num_tokens_no_spec[non_last_idx]) |
| 851 | last_len = int(last_runner.input_batch.num_tokens_no_spec[last_idx]) |
| 852 | assert non_last_len == last_len |
| 853 | assert ( |
| 854 | non_last_runner.input_batch.token_ids_cpu[ |
| 855 | non_last_idx, :non_last_len |
| 856 | ].tolist() |
| 857 | == last_runner.input_batch.token_ids_cpu[last_idx, :last_len].tolist() |
| 858 | ) |
| 859 | |
| 860 | |
| 861 | def test_update_states_pp_async_multi_request_keeps_rank_state_consistent( |
| 862 | model_runner, model_runner_2, dist_init, monkeypatch |
| 863 | ): |
| 864 | req_ids = ["req_0", "req_1"] |
| 865 | non_last_runner = model_runner |
| 866 | last_runner = model_runner_2 |
| 867 | non_last_runner.use_async_scheduling = True |
| 868 | last_runner.use_async_scheduling = True |
| 869 | |
| 870 | # Both ranks start from the same request set. |
| 871 | monkeypatch.setattr( |
| 872 | "vllm.v1.worker.gpu_model_runner.get_pp_group", |
| 873 | lambda: SimpleNamespace(is_last_rank=False, world_size=2), |
| 874 | ) |
| 875 | non_last_runner._update_states(_schedule_new_request(*req_ids)) |
| 876 | last_runner._update_states(_schedule_new_request(*req_ids)) |
| 877 | |
| 878 | # Simulate async previous-step sampled tokens known on both ranks. |
| 879 | # non-last rank may receive them via PP communication; last rank has |
| 880 | # them from local sampling/bookkeeping. |
| 881 | sampled_by_last_rank = {req_ids[0]: 111, req_ids[1]: 222} |
| 882 | for runner in (non_last_runner, last_runner): |
| 883 | for req_id, token_id in sampled_by_last_rank.items(): |
| 884 | req_index = runner.input_batch.req_id_to_index[req_id] |
| 885 | start_idx = int(runner.input_batch.num_tokens_no_spec[req_index]) |
| 886 | end_idx = start_idx + 1 |
| 887 | runner.input_batch.token_ids_cpu[req_index, start_idx:end_idx] = [token_id] |
| 888 | runner.input_batch.is_token_ids[req_index, start_idx:end_idx] = True |
| 889 | runner.input_batch.num_tokens_no_spec[req_index] = end_idx |
| 890 | runner.requests[req_id].output_token_ids.append(token_id) |
| 891 | |
| 892 | scheduler_output = _schedule_cached_requests( |
| 893 | req_ids=req_ids, |
| 894 | num_scheduled_tokens={req_ids[0]: 1, req_ids[1]: 1}, |
| 895 | new_token_ids=[], |
| 896 | num_computed_tokens=[4, 4], |
| 897 | num_output_tokens=[1, 1], |
| 898 | ) |
| 899 | # non-last rank: async PP branch (new_token_ids empty). |
| 900 | monkeypatch.setattr( |
| 901 | "vllm.v1.worker.gpu_model_runner.get_pp_group", |
| 902 | lambda: SimpleNamespace(is_last_rank=False, world_size=2), |
| 903 | ) |
| 904 | non_last_runner._update_states(scheduler_output) |
| 905 | # last rank: keep already-bookkept state aligned with scheduler view. |
| 906 | monkeypatch.setattr( |
| 907 | "vllm.v1.worker.gpu_model_runner.get_pp_group", |
| 908 | lambda: SimpleNamespace(is_last_rank=True, world_size=2), |
| 909 | ) |
| 910 | last_runner._update_states(scheduler_output) |
| 911 | |
| 912 | for req_id in req_ids: |
| 913 | non_last_idx = non_last_runner.input_batch.req_id_to_index[req_id] |
| 914 | last_idx = last_runner.input_batch.req_id_to_index[req_id] |
| 915 | non_last_len = int(non_last_runner.input_batch.num_tokens_no_spec[non_last_idx]) |
| 916 | last_len = int(last_runner.input_batch.num_tokens_no_spec[last_idx]) |
| 917 | assert non_last_len == last_len |
| 918 | assert ( |
| 919 | non_last_runner.input_batch.token_ids_cpu[ |
| 920 | non_last_idx, :non_last_len |
| 921 | ].tolist() |
| 922 | == last_runner.input_batch.token_ids_cpu[last_idx, :last_len].tolist() |
| 923 | ) |
| 924 | |
| 925 | |
| 926 | def test_update_config(model_runner): |
| 927 | # Simple update |
| 928 | model_runner.update_config({"load_config": {"load_format": "dummy"}}) |
| 929 | assert model_runner.load_config.load_format == "dummy" |
| 930 | # Raise error on non-existing config |
| 931 | with pytest.raises(ValueError, match="do_not_exist_config"): |
| 932 | model_runner.update_config({"do_not_exist_config": "dummy"}) |
| 933 | |
| 934 | |
| 935 | def test_load_model_weights_inplace(dist_init, model_runner, model_runner_2): |
| 936 | # In this test, model_runner loads model + weights in one go, while |
| 937 | # model_runner_2 loads dummy weights first then load real weights inplace |
| 938 | model_runner.load_model() |
| 939 | original_load_format = model_runner_2.load_config.load_format |
| 940 | model_runner_2.update_config({"load_config": {"load_format": "dummy"}}) |
| 941 | model_runner_2.load_model() # Initial model loading with dummy weights |
| 942 | assert str(model_runner.get_model().state_dict()) != str( |
| 943 | model_runner_2.get_model().state_dict() |
| 944 | ) |
| 945 | model_runner_2.update_config({"load_config": {"load_format": original_load_format}}) |
| 946 | model_runner_2.reload_weights() # Load real weights inplace |
| 947 | assert str(model_runner.get_model().state_dict()) == str( |
| 948 | model_runner_2.get_model().state_dict() |
| 949 | ) |
| 950 | |
| 951 | |
| 952 | def test_reload_weights_before_load_model(model_runner): |
| 953 | with pytest.raises(ValueError): |
| 954 | model_runner.reload_weights() |
| 955 | |
| 956 | |
| 957 | def test_sample_passes_reordered_draft_probs_to_rejection_sampler(): |
| 958 | runner = object.__new__(GPUModelRunner) |
| 959 | runner.use_async_scheduling = False |
| 960 | runner.input_batch = SimpleNamespace( |
| 961 | sampling_metadata=Mock(spec=SamplingMetadata), |
| 962 | update_async_output_token_ids=Mock(), |
| 963 | req_ids=["req_a", "req_b", "req_c"], |
| 964 | ) |
| 965 | runner.rejection_sampler = Mock(return_value="sampler_output") |
| 966 | runner.sampler = Mock() |
| 967 | runner._draft_prob_req_ids = ["req_c", "req_a", "req_b"] |
| 968 | runner._draft_probs = torch.arange(3 * 3 * 4, dtype=torch.float32).reshape(3, 3, 4) |
| 969 | |
| 970 | spec_decode_metadata = SpecDecodeMetadata.make_dummy( |
| 971 | [[1, 2], [], [3]], |
| 972 | device=torch.device("cpu"), |
| 973 | ) |
| 974 | logits = torch.randn(6, 4) |
| 975 | |
| 976 | output = GPUModelRunner._sample(runner, logits, spec_decode_metadata) |
| 977 | |
| 978 | assert output == "sampler_output" |
| 979 | passed_draft_probs = runner.rejection_sampler.call_args.args[1] |
| 980 | expected_draft_probs = torch.cat( |
| 981 | [ |
| 982 | runner._draft_probs[1, :2], |
| 983 | runner._draft_probs[0, :1], |
| 984 | ], |
| 985 | dim=0, |
| 986 | ) |
| 987 | assert torch.equal(passed_draft_probs, expected_draft_probs) |
| 988 | |
| 989 | |
| 990 | def test_dummy_sampler_run_warms_all_greedy_rejection_sampler(monkeypatch): |
| 991 | runner = object.__new__(GPUModelRunner) |
| 992 | runner.device = torch.device("cpu") |
| 993 | runner.vllm_config = SimpleNamespace( |
| 994 | model_config=SimpleNamespace(multimodal_config=None) |
| 995 | ) |
| 996 | runner.model = SimpleNamespace(compute_logits=Mock(return_value=torch.randn(3, 8))) |
| 997 | runner.sampler = Mock(return_value="sampler_output") |
| 998 | runner.sampler.logprobs_mode = "processed_logprobs" |
| 999 | runner.speculative_config = SimpleNamespace( |
| 1000 | rejection_sample_method="standard", |
| 1001 | draft_sample_method="greedy", |
| 1002 | ) |
| 1003 | runner.rejection_sampler = Mock() |
| 1004 | synchronize = Mock() |
| 1005 | monkeypatch.setattr(torch.accelerator, "synchronize", synchronize) |
| 1006 | |
| 1007 | output = GPUModelRunner._dummy_sampler_run(runner, torch.randn(3, 4)) |
| 1008 | |
| 1009 | assert output == "sampler_output" |
| 1010 | assert runner.rejection_sampler.call_count == 2 |
| 1011 | mixed_metadata = runner.rejection_sampler.call_args_list[0].args[3] |
| 1012 | all_greedy_metadata = runner.rejection_sampler.call_args_list[1].args[3] |
| 1013 | assert not mixed_metadata.all_greedy |
| 1014 | assert mixed_metadata.temperature is not None |
| 1015 | assert all_greedy_metadata.all_greedy |
| 1016 | assert not all_greedy_metadata.all_random |
| 1017 | assert all_greedy_metadata.temperature is None |
| 1018 | synchronize.assert_called_once_with() |
| 1019 | |
| 1020 | |
| 1021 | def test_invalid_draft_suffixes_remain_rejected_in_metadata(): |
| 1022 | runner = object.__new__(GPUModelRunner) |
| 1023 | runner.device = torch.device("cpu") |
| 1024 | runner.arange_np = np.arange(64, dtype=np.int64) |
| 1025 | runner._arange_scratch = np.empty(64, dtype=np.int64) |
| 1026 | # Placeholder (-1) drafts are kept in input_ids (clamped to 0 only at the |
| 1027 | # embedding boundary). For num_draft_tokens=[2, 1, 2] the draft positions |
| 1028 | # are [1, 2, 4, 6, 7], so the gather carries the -1s straight into the |
| 1029 | # rejection-sampling metadata. |
| 1030 | runner.input_ids = SimpleNamespace( |
| 1031 | gpu=torch.tensor([99, 10, -1, 99, 12, 99, 13, -1], dtype=torch.int32), |
| 1032 | ) |
| 1033 | |
| 1034 | metadata = GPUModelRunner._calc_spec_decode_metadata( |
| 1035 | runner, |
| 1036 | np.array([2, 1, 2], dtype=np.int32), |
| 1037 | np.array([3, 5, 8], dtype=np.int32), |
| 1038 | ) |
| 1039 | |
| 1040 | assert metadata.draft_token_ids.tolist() == [10, -1, 12, 13, -1] |
| 1041 | |
| 1042 | |
| 1043 | def test_init_kv_cache_with_kv_sharing_invalid_target_layer_order(default_vllm_config): |
| 1044 | torch.set_default_dtype(torch.float16) |
| 1045 | layer_0 = "model.layers.0.self_attn.attn" |
| 1046 | layer_1 = "model.layers.1.self_attn.attn" |
| 1047 | error_msg = f"{layer_1} must come before the current layer" |
| 1048 | with pytest.raises(ValueError, match=error_msg): |
| 1049 | fwd_context = { |
| 1050 | # initialization below will fail because target layer is invalid; |
| 1051 | # the target layer needs to come before layer 1 |
| 1052 | layer_0: Attention( |
| 1053 | num_heads=8, |
| 1054 | head_size=64, |
| 1055 | scale=1.0, |
| 1056 | prefix=layer_0, |
| 1057 | kv_sharing_target_layer_name=layer_1, |
| 1058 | ), |
| 1059 | layer_1: Attention( |
| 1060 | num_heads=8, |
| 1061 | head_size=64, |
| 1062 | scale=1.0, |
| 1063 | prefix=layer_1, |
| 1064 | ), |
| 1065 | } |
| 1066 | # suppress var not used error |
| 1067 | assert fwd_context is not None |
| 1068 | |
| 1069 | |
| 1070 | def test_init_kv_cache_with_kv_sharing_target_layer_not_exist(default_vllm_config): |
| 1071 | torch.set_default_dtype(torch.float16) |
| 1072 | layer_0 = "model.layers.0.self_attn.attn" |
| 1073 | layer_1 = "model.layers.1.self_attn.attn" |
| 1074 | invalid_layer = "model.layers.0.cross_attn.attn" |
| 1075 | error_msg = f"{invalid_layer} is not a valid Attention layer in the model" |
| 1076 | with pytest.raises(ValueError, match=error_msg): |
| 1077 | fwd_context = { |
| 1078 | layer_0: Attention( |
| 1079 | num_heads=8, |
| 1080 | head_size=64, |
| 1081 | scale=1.0, |
| 1082 | prefix=layer_0, |
| 1083 | ), |
| 1084 | layer_1: Attention( |
| 1085 | num_heads=8, |
| 1086 | head_size=64, |
| 1087 | scale=1.0, |
| 1088 | prefix=layer_1, |
| 1089 | # invalid layer: cross_attn.atn doesn't exist! |
| 1090 | kv_sharing_target_layer_name=invalid_layer, |
| 1091 | ), |
| 1092 | } |
| 1093 | # suppress var not used error |
| 1094 | assert fwd_context is not None |
| 1095 | |
| 1096 | |
| 1097 | def test_init_kv_cache_with_kv_sharing_target_same_as_current(default_vllm_config): |
| 1098 | torch.set_default_dtype(torch.float16) |
| 1099 | layer_0 = "model.layers.0.self_attn.attn" |
| 1100 | layer_1 = "model.layers.1.self_attn.attn" |
| 1101 | error_msg = f"{layer_1} cannot be the same as the current layer" |
| 1102 | with pytest.raises(ValueError, match=error_msg): |
| 1103 | fwd_context = { |
| 1104 | # initialization below will fail because target layer is invalid; |
| 1105 | # the target layer needs to come before layer 1 |
| 1106 | layer_0: Attention( |
| 1107 | num_heads=8, |
| 1108 | head_size=64, |
| 1109 | scale=1.0, |
| 1110 | prefix=layer_0, |
| 1111 | ), |
| 1112 | layer_1: Attention( |
| 1113 | num_heads=8, |
| 1114 | head_size=64, |
| 1115 | scale=1.0, |
| 1116 | prefix=layer_1, |
| 1117 | kv_sharing_target_layer_name=layer_1, |
| 1118 | ), |
| 1119 | } |
| 1120 | # suppress var not used error |
| 1121 | assert fwd_context is not None |
| 1122 | |
| 1123 | |
| 1124 | def test_init_kv_cache_without_kv_sharing(default_vllm_config): |
| 1125 | torch.set_default_dtype(torch.float16) |
| 1126 | layer_0 = "model.layers.0.self_attn.attn" |
| 1127 | layer_1 = "model.layers.1.self_attn.attn" |
| 1128 | vllm_config = get_vllm_config() |
| 1129 | with set_current_vllm_config(vllm_config): |
| 1130 | fwd_context = { |
| 1131 | layer_0: Attention( |
| 1132 | num_heads=8, |
| 1133 | head_size=64, |
| 1134 | scale=1.0, |
| 1135 | prefix=layer_0, |
| 1136 | ), |
| 1137 | layer_1: Attention( |
| 1138 | num_heads=8, |
| 1139 | head_size=64, |
| 1140 | scale=1.0, |
| 1141 | prefix=layer_1, |
| 1142 | ), |
| 1143 | } |
| 1144 | # suppress var not used error |
| 1145 | assert fwd_context is not None |
| 1146 | # Set high context length to test max context length estimation |
| 1147 | vllm_config.model_config.max_model_len = 3_000_000 |
| 1148 | vllm_ctx = vllm_config.compilation_config.static_forward_context |
| 1149 | runner = GPUModelRunner(vllm_config, DEVICE_TYPE) |
| 1150 | kv_cache_spec = runner.get_kv_cache_spec() |
| 1151 | assert len(kv_cache_spec) == 2 |
| 1152 | assert len(runner.shared_kv_cache_layers) == 0 |
| 1153 | |
| 1154 | available_memory = 20 * GiB_bytes |
| 1155 | # page size for layer 0's kv_cache_spec is 32KB |
| 1156 | num_expected_blocks = 327680 # 20GB / 32KB / 2 (num layers) |
| 1157 | kv_cache_config = get_kv_cache_configs( |
| 1158 | vllm_config, [kv_cache_spec], [available_memory] |
| 1159 | )[0] |
| 1160 | assert kv_cache_config.num_blocks == num_expected_blocks |
| 1161 | assert len(kv_cache_config.kv_cache_tensors) == 1 |
| 1162 | assert kv_cache_config.kv_cache_tensors[0].size == available_memory |
| 1163 | |
| 1164 | max_context_len = estimate_max_model_len(vllm_config, kv_cache_spec, 5 * GiB_bytes) |
| 1165 | # max context len with KV sharing should be 2x as large as without |
| 1166 | assert max_context_len == 1310720 |
| 1167 | |
| 1168 | # important: override tensor size to prevent large mem alloc during test |
| 1169 | # this will only allocate 1 block worth of memory per layer (2 layers * 32kb) |
| 1170 | kv_cache_config.num_blocks = 1 |
| 1171 | for kv_cache_tensor in kv_cache_config.kv_cache_tensors: |
| 1172 | page_size = kv_cache_spec[kv_cache_tensor.layers[0]].page_size_bytes |
| 1173 | kv_cache_tensor.size = page_size * len(kv_cache_tensor.layers) |
| 1174 | # One block per layer: a layer's region is exactly one page. |
| 1175 | kv_cache_tensor.layer_stride = page_size |
| 1176 | kv_cache_tensor.block_stride = page_size |
| 1177 | |
| 1178 | runner.initialize_kv_cache(kv_cache_config) |
| 1179 | |
| 1180 | layer_0_kv = vllm_ctx[layer_0].kv_cache |
| 1181 | layer_1_kv = vllm_ctx[layer_1].kv_cache |
| 1182 | # check layer 1 kv cache does NOT share memory with layer 0 |
| 1183 | assert id(layer_1_kv) != id(layer_0_kv) |
| 1184 | |
| 1185 | # check layer 1 added to kv cache group's layer names |
| 1186 | assert len(kv_cache_config.kv_cache_groups) == 1 |
| 1187 | assert len(kv_cache_config.kv_cache_groups[0].layer_names) == 2 |
| 1188 | assert kv_cache_config.kv_cache_groups[0].layer_names[0] == layer_0 |
| 1189 | assert kv_cache_config.kv_cache_groups[0].layer_names[1] == layer_1 |
| 1190 | |
| 1191 | |
| 1192 | def test_init_kv_cache_with_kv_sharing_valid(default_vllm_config): |
| 1193 | torch.set_default_dtype(torch.float16) |
| 1194 | layer_0 = "model.layers.0.self_attn.attn" |
| 1195 | layer_1 = "model.layers.1.self_attn.attn" |
| 1196 | vllm_config = get_vllm_config() |
| 1197 | with set_current_vllm_config(vllm_config): |
| 1198 | fwd_context = { |
| 1199 | layer_0: Attention( |
| 1200 | num_heads=8, |
| 1201 | head_size=64, |
| 1202 | scale=1.0, |
| 1203 | prefix=layer_0, |
| 1204 | ), |
| 1205 | layer_1: Attention( |
| 1206 | num_heads=8, |
| 1207 | head_size=64, |
| 1208 | scale=1.0, |
| 1209 | prefix=layer_1, |
| 1210 | kv_sharing_target_layer_name="model.layers.0.self_attn.attn", |
| 1211 | ), |
| 1212 | } |
| 1213 | # suppress var not used error |
| 1214 | assert fwd_context is not None |
| 1215 | # Set high context length to test max context length estimation |
| 1216 | vllm_config.model_config.max_model_len = 3_000_000 |
| 1217 | vllm_ctx = vllm_config.compilation_config.static_forward_context |
| 1218 | runner = GPUModelRunner(vllm_config, DEVICE_TYPE) |
| 1219 | kv_cache_spec = runner.get_kv_cache_spec() |
| 1220 | assert len(kv_cache_spec) == 1 |
| 1221 | assert layer_0 in kv_cache_spec |
| 1222 | assert runner.shared_kv_cache_layers[layer_1] == layer_0 |
| 1223 | |
| 1224 | available_memory = 20 * GiB_bytes |
| 1225 | # page size for layer 0's kv_cache_spec is 32KB |
| 1226 | # with KV sharing, we can allocate (available_mem//page_size//1) blocks |
| 1227 | # which is twice as many as without KV sharing |
| 1228 | num_expected_blocks = 655360 # 20GB / 32KB |
| 1229 | kv_cache_config = get_kv_cache_configs( |
| 1230 | vllm_config, [kv_cache_spec], [available_memory] |
| 1231 | )[0] |
| 1232 | assert kv_cache_config.num_blocks == num_expected_blocks |
| 1233 | assert len(kv_cache_config.kv_cache_tensors) == 1 |
| 1234 | # Each layer now has twice the available memory for KV cache |
| 1235 | # compared to no KV sharing |
| 1236 | assert kv_cache_config.kv_cache_tensors[0].size == available_memory |
| 1237 | |
| 1238 | max_context_len = estimate_max_model_len(vllm_config, kv_cache_spec, 5 * GiB_bytes) |
| 1239 | # max context len with KV sharing should be 2x as large as without |
| 1240 | assert max_context_len == 2 * 1310720 |
| 1241 | |
| 1242 | # important: override tensor size to prevent large mem alloc during test |
| 1243 | # this will only allocate 1 block worth of memory (32kb) |
| 1244 | kv_cache_config.num_blocks = 1 |
| 1245 | page_size = kv_cache_spec[layer_0].page_size_bytes |
| 1246 | kv_cache_config.kv_cache_tensors[0].size = page_size |
| 1247 | kv_cache_config.kv_cache_tensors[0].layer_stride = page_size |
| 1248 | |
| 1249 | runner.initialize_kv_cache(kv_cache_config) |
| 1250 | kv_cache_config_after_init = runner.kv_cache_config |
| 1251 | |
| 1252 | layer_0_kv = vllm_ctx[layer_0].kv_cache |
| 1253 | layer_1_kv = vllm_ctx[layer_1].kv_cache |
| 1254 | # check layer 1 kv cache shares memory with layer 0 |
| 1255 | assert id(layer_1_kv) == id(layer_0_kv) |
| 1256 | |
| 1257 | # check layer 1 added to kv cache group's layer names |
| 1258 | assert len(kv_cache_config_after_init.kv_cache_groups) == 1 |
| 1259 | assert len(kv_cache_config_after_init.kv_cache_groups[0].layer_names) == 2 |
| 1260 | assert kv_cache_config_after_init.kv_cache_groups[0].layer_names[0] == layer_0 |
| 1261 | assert kv_cache_config_after_init.kv_cache_groups[0].layer_names[1] == layer_1 |
| 1262 | |
| 1263 | |
| 1264 | @pytest.mark.skipif( |
| 1265 | not current_platform.is_cuda(), |
| 1266 | reason="Attention backend FLASHINFER is only supported on CUDA.", |
| 1267 | ) |
| 1268 | def test_hybrid_attention_mamba_tensor_shapes(): |
| 1269 | """The GPU model runner creates different views into the |
| 1270 | KVCacheTensors for the attention and mamba layers |
| 1271 | (via _allocate_kv_caches). This test verifies |
| 1272 | that the views are compatible: writing a mamba block |
| 1273 | will not corrupt an attention block and vice versa |
| 1274 | """ |
| 1275 | set_random_seed(42) |
| 1276 | |
| 1277 | update_environment_variables( |
| 1278 | { |
| 1279 | "RANK": "0", |
| 1280 | "LOCAL_RANK": "0", |
| 1281 | "WORLD_SIZE": "1", |
| 1282 | "MASTER_ADDR": "localhost", |
| 1283 | "MASTER_PORT": "12345", |
| 1284 | } |
| 1285 | ) |
| 1286 | from tests.utils import ensure_current_vllm_config |
| 1287 | |
| 1288 | with ensure_current_vllm_config(): |
| 1289 | init_distributed_environment() |
| 1290 | initialize_model_parallel(tensor_model_parallel_size=1) |
| 1291 | torch.set_default_dtype(torch.float16) |
| 1292 | |
| 1293 | model_config = ModelConfig( |
| 1294 | model="ibm-granite/granite-4.0-tiny-preview", |
| 1295 | dtype="float16", |
| 1296 | ) |
| 1297 | scheduler_config = SchedulerConfig( |
| 1298 | max_num_seqs=10, |
| 1299 | max_num_batched_tokens=512, |
| 1300 | max_model_len=512, |
| 1301 | is_encoder_decoder=model_config.is_encoder_decoder, |
| 1302 | ) |
| 1303 | cache_config = CacheConfig( |
| 1304 | block_size=BLOCK_SIZE, |
| 1305 | gpu_memory_utilization=0.9, |
| 1306 | cache_dtype="auto", |
| 1307 | ) |
| 1308 | cache_config.kv_cache_layout = "LBNHC" |
| 1309 | parallel_config = ParallelConfig() |
| 1310 | attention_config = AttentionConfig(backend=AttentionBackendEnum.FLASHINFER) |
| 1311 | vllm_config = VllmConfig( |
| 1312 | model_config=model_config, |
| 1313 | cache_config=cache_config, |
| 1314 | scheduler_config=scheduler_config, |
| 1315 | parallel_config=parallel_config, |
| 1316 | attention_config=attention_config, |
| 1317 | ) |
| 1318 | |
| 1319 | layer_0 = "model.layers.0.self_attn.attn" |
| 1320 | layer_1 = "model.layers.1.self_attn.attn" |
| 1321 | layer_2 = "model.layers.2.mixer" |
| 1322 | layer_3 = "model.layers.3.mixer" |
| 1323 | layer_4 = "model.layers.4.mixer" |
| 1324 | layer_5 = "model.layers.5.mixer" |
| 1325 | |
| 1326 | with set_current_vllm_config(vllm_config): |
| 1327 | hf_config = vllm_config.model_config.hf_config |
| 1328 | fwd_context = {} |
| 1329 | for key in [layer_0, layer_1]: |
| 1330 | fwd_context[key] = Attention( |
| 1331 | num_heads=model_config.get_num_attention_heads(parallel_config), |
| 1332 | num_kv_heads=model_config.get_num_kv_heads(parallel_config), |
| 1333 | head_size=model_config.get_head_size(), |
| 1334 | scale=1.0, |
| 1335 | prefix=key, |
| 1336 | ) |
| 1337 | for key in [layer_2, layer_3, layer_4, layer_5]: |
| 1338 | fwd_context[key] = MambaMixer2( |
| 1339 | hidden_size=hf_config.hidden_size, |
| 1340 | ssm_state_size=hf_config.mamba_d_state, |
| 1341 | conv_kernel_size=hf_config.mamba_d_conv, |
| 1342 | intermediate_size=hf_config.mamba_expand * hf_config.hidden_size, |
| 1343 | use_conv_bias=hf_config.mamba_conv_bias, |
| 1344 | use_bias=hf_config.mamba_proj_bias, |
| 1345 | n_groups=hf_config.mamba_n_groups, |
| 1346 | num_heads=hf_config.mamba_n_heads, |
| 1347 | head_dim=hf_config.mamba_d_head, |
| 1348 | rms_norm_eps=hf_config.rms_norm_eps, |
| 1349 | activation=hf_config.hidden_act, |
| 1350 | cache_config=cache_config, |
| 1351 | model_config=model_config, |
| 1352 | prefix=key, |
| 1353 | ) |
| 1354 | # suppress var not used error |
| 1355 | assert fwd_context is not None |
| 1356 | vllm_ctx = vllm_config.compilation_config.static_forward_context |
| 1357 | |
| 1358 | runner = GPUModelRunner(vllm_config, DEVICE_TYPE) |
| 1359 | current_platform.update_block_size_for_backend(vllm_config) |
| 1360 | kv_cache_spec = runner.get_kv_cache_spec() |
| 1361 | |
| 1362 | available_memory = 5 * GiB_bytes |
| 1363 | kv_cache_config = get_kv_cache_configs( |
| 1364 | vllm_config, [kv_cache_spec], [available_memory] |
| 1365 | )[0] |
| 1366 | runner.initialize_kv_cache(kv_cache_config) |
| 1367 | |
| 1368 | # random partition of blocks |
| 1369 | # blocks0 will be assigned to attention layers |
| 1370 | # blocks1 will be assigned to mamba layers |
| 1371 | num_blocks = kv_cache_config.num_blocks |
| 1372 | ind = np.arange(num_blocks) |
| 1373 | np.random.shuffle(ind) |
| 1374 | blocks0, blocks1 = ind[: (num_blocks // 2)], ind[(num_blocks // 2) :] |
| 1375 | |
| 1376 | attn_shape = vllm_ctx[layer_0].kv_cache.shape |
| 1377 | conv_shape = vllm_ctx[layer_2].kv_cache[0].shape |
| 1378 | ssm_shape = vllm_ctx[layer_2].kv_cache[1].shape |
| 1379 | |
| 1380 | # assert we are using FlashInfer |
| 1381 | assert attn_shape[0] % num_blocks == 0 |
| 1382 | block_split_ratio = attn_shape[0] // num_blocks |
| 1383 | |
| 1384 | # use small blocks for testing to avoid memory issues |
| 1385 | test_block_size = min(2, len(blocks0), len(blocks1)) |
| 1386 | |
| 1387 | # use non-overlapping blocks to avoid data contamination |
| 1388 | # Split kernel blocks: first half for attention, second half for mamba |
| 1389 | mid_point = num_blocks // 2 |
| 1390 | |
| 1391 | # attention uses kernel blocks from first half (mapped to logical blocks) |
| 1392 | kv_blocks_for_attention = np.array([0, 1])[:test_block_size] |
| 1393 | |
| 1394 | # mamba uses kernel blocks from second half |
| 1395 | kv_blocks_for_mamba = np.array([mid_point, mid_point + 1])[:test_block_size] |
| 1396 | |
| 1397 | # create small constant tensors for testing with corrected shapes |
| 1398 | # attention: [block_size, ...] starting from dimension 2 |
| 1399 | attn_constant_shape = attn_shape[2:] |
| 1400 | conv_constant_shape = conv_shape[1:] |
| 1401 | ssm_constant_shape = ssm_shape[1:] |
| 1402 | |
| 1403 | attn_blocks_constant = torch.full( |
| 1404 | (test_block_size, *attn_constant_shape), device=DEVICE_TYPE, fill_value=3.33 |
| 1405 | ) |
| 1406 | conv_blocks_constant = torch.full( |
| 1407 | (test_block_size, *conv_constant_shape), device=DEVICE_TYPE, fill_value=6.66 |
| 1408 | ) |
| 1409 | ssm_blocks_constant = torch.full( |
| 1410 | (test_block_size, *ssm_constant_shape), device=DEVICE_TYPE, fill_value=9.99 |
| 1411 | ) |
| 1412 | |
| 1413 | # Fill attention blocks with constants using kv block indices |
| 1414 | kernel_blocks_for_attention = kv_blocks_for_attention * block_split_ratio |
| 1415 | |
| 1416 | for layer in [layer_0, layer_1]: |
| 1417 | # attention: kv_cache[kernel_block_idx, kv_idx, ...] |
| 1418 | for i, kernel_block in enumerate(kernel_blocks_for_attention): |
| 1419 | vllm_ctx[layer].kv_cache[kernel_block, :] = attn_blocks_constant[i] |
| 1420 | |
| 1421 | # fill mamba blocks with constants using kernel block indices |
| 1422 | for layer in [layer_2, layer_3, layer_4, layer_5]: |
| 1423 | # mamba: kv_cache[component][kernel_block_idx, ...] |
| 1424 | for i, kv_block in enumerate(kv_blocks_for_mamba): |
| 1425 | vllm_ctx[layer].kv_cache[0][kv_block, :] = conv_blocks_constant[i] |
| 1426 | vllm_ctx[layer].kv_cache[1][kv_block, :] = ssm_blocks_constant[i] |
| 1427 | |
| 1428 | # verify attention and mamba contents are correct |
| 1429 | for layer in [layer_0, layer_1]: |
| 1430 | for i, kernel_block in enumerate(kernel_blocks_for_attention): |
| 1431 | actual_kv = vllm_ctx[layer].kv_cache[kernel_block, :] |
| 1432 | expected = attn_blocks_constant[i] |
| 1433 | |
| 1434 | # Packed layout: (num_kv_heads, block_size, 2*head_size). Every |
| 1435 | # head in the block was filled with the same constant. |
| 1436 | for head_idx in range(actual_kv.shape[0]): |
| 1437 | assert torch.equal(actual_kv[head_idx], expected) |
| 1438 | |
| 1439 | for layer in [layer_2, layer_3, layer_4, layer_5]: |
| 1440 | for i, kv_block in enumerate(kv_blocks_for_mamba): |
| 1441 | actual_conv = vllm_ctx[layer].kv_cache[0][kv_block, :] |
| 1442 | actual_ssm = vllm_ctx[layer].kv_cache[1][kv_block, :] |
| 1443 | expected_conv = conv_blocks_constant[i] |
| 1444 | expected_ssm = ssm_blocks_constant[i] |
| 1445 | |
| 1446 | assert torch.equal(actual_conv, expected_conv) |
| 1447 | assert torch.equal(actual_ssm, expected_ssm) |
| 1448 | |
| 1449 | for layer in [layer_2, layer_3, layer_4, layer_5]: |
| 1450 | for i, kv_block in enumerate(kv_blocks_for_mamba): |
| 1451 | actual_conv = vllm_ctx[layer].kv_cache[0][kv_block, :] |
| 1452 | actual_ssm = vllm_ctx[layer].kv_cache[1][kv_block, :] |
| 1453 | expected_conv = conv_blocks_constant[i] |
| 1454 | expected_ssm = ssm_blocks_constant[i] |
| 1455 | assert torch.equal(actual_conv, expected_conv) |
| 1456 | assert torch.equal(actual_ssm, expected_ssm) |
| 1457 | |
| 1458 | |
| 1459 | def test_input_batch_reinitialized_after_late_interleave_adjustment(monkeypatch): |
| 1460 | runner = object.__new__(GPUModelRunner) |
| 1461 | runner.vllm_config = SimpleNamespace(reasoning_config=None) |
| 1462 | runner.parallel_config = SimpleNamespace(cp_kv_cache_interleave_size=16) |
| 1463 | runner.cache_config = SimpleNamespace(use_replayssm=False) |
| 1464 | runner.model_config = SimpleNamespace(get_vocab_size=lambda: 32) |
| 1465 | runner.max_model_len = 64 |
| 1466 | runner.max_encoder_len = 0 |
| 1467 | runner.max_num_reqs = 1 |
| 1468 | runner.max_num_tokens = 64 |
| 1469 | runner.num_spec_tokens = 0 |
| 1470 | runner.device = torch.device("cpu") |
| 1471 | runner.is_pooling_model = False |
| 1472 | runner._init_block_sizes = [16] |
| 1473 | runner._init_kernel_block_sizes = [16] |
| 1474 | runner._init_max_num_blocks = [4] |
| 1475 | runner._init_slot_mapping_modes = [ |
| 1476 | gpu_model_runner_module.SlotMappingMode.TOKEN_TO_KV_SLOT |
| 1477 | ] |
| 1478 | runner.cp_kv_cache_interleave_size = 1 |
| 1479 | runner.input_batch = SimpleNamespace( |
| 1480 | logitsprocs=None, |
| 1481 | logitsprocs_need_output_token_ids=False, |
| 1482 | ) |
| 1483 | runner.jit_warmup_registry = Mock() |
| 1484 | runner.jit_warmup_registry.activate.return_value = nullcontext() |
| 1485 | |
| 1486 | spec = SimpleNamespace( |
| 1487 | block_size=16, |
| 1488 | max_num_blocks_per_req=lambda *_: 4, |
| 1489 | ) |
| 1490 | kv_cache_config = SimpleNamespace( |
| 1491 | kv_cache_groups=[SimpleNamespace(kv_cache_spec=spec)] |
| 1492 | ) |
| 1493 | input_batch_cls = Mock(return_value=SimpleNamespace()) |
| 1494 | monkeypatch.setattr(gpu_model_runner_module, "InputBatch", input_batch_cls) |
| 1495 | monkeypatch.setattr( |
| 1496 | gpu_model_runner_module, |
| 1497 | "get_kv_cache_spec_kind", |
| 1498 | lambda _: gpu_model_runner_module.KVCacheSpecKind.FULL_ATTENTION, |
| 1499 | ) |
| 1500 | |
| 1501 | runner.may_reinitialize_input_batch(kv_cache_config, [16]) |
| 1502 | |
| 1503 | assert input_batch_cls.call_count == 1 |
| 1504 | assert input_batch_cls.call_args.kwargs["cp_kv_cache_interleave_size"] == 16 |
| 1505 | |
| 1506 | |
| 1507 | def test_v2_runner_snapshots_late_interleave_adjustment(monkeypatch): |
| 1508 | from vllm.v1.worker.gpu import model_runner as v2_model_runner_module |
| 1509 | |
| 1510 | runner = object.__new__(v2_model_runner_module.GPUModelRunner) |
| 1511 | runner.parallel_config = SimpleNamespace(cp_kv_cache_interleave_size=16) |
| 1512 | runner.cp_interleave = 1 |
| 1513 | |
| 1514 | class StopInitialization(Exception): |
| 1515 | pass |
| 1516 | |
| 1517 | monkeypatch.setattr( |
| 1518 | v2_model_runner_module, |
| 1519 | "deepcopy", |
| 1520 | Mock(side_effect=StopInitialization), |
| 1521 | ) |
| 1522 | |
| 1523 | with pytest.raises(StopInitialization): |
| 1524 | runner.initialize_kv_cache(SimpleNamespace()) |
| 1525 | |
| 1526 | assert runner.cp_interleave == 16 |
| 1527 | |
| 1528 | |
| 1529 | def test_hybrid_block_table_initialization(): |
| 1530 | """Test hybrid block table with different kernel and kvcache_manager block |
| 1531 | sizes.""" |
| 1532 | from vllm.v1.worker.block_table import BlockTable |
| 1533 | |
| 1534 | # Test configuration: kvcache_manager block size = 32, |
| 1535 | # kernel block size = 16 |
| 1536 | block_size = 32 |
| 1537 | kernel_block_sizes = [16] |
| 1538 | max_num_reqs = 10 |
| 1539 | max_num_blocks_per_req = 20 |
| 1540 | max_num_batched_tokens = 512 |
| 1541 | cp_kv_cache_interleave_size = 8 |
| 1542 | |
| 1543 | block_table = BlockTable( |
| 1544 | block_size=block_size, |
| 1545 | max_num_reqs=max_num_reqs, |
| 1546 | max_num_blocks_per_req=max_num_blocks_per_req, |
| 1547 | max_num_batched_tokens=max_num_batched_tokens, |
| 1548 | pin_memory=False, |
| 1549 | device=torch.device(DEVICE_TYPE), |
| 1550 | kernel_block_size=kernel_block_sizes[0], |
| 1551 | cp_kv_cache_interleave_size=cp_kv_cache_interleave_size, |
| 1552 | ) |
| 1553 | |
| 1554 | # Verify hybrid block configuration |
| 1555 | assert block_table.use_hybrid_blocks is True |
| 1556 | assert block_table.block_size == kernel_block_sizes[0] |
| 1557 | assert block_table.blocks_per_kv_block == ( |
| 1558 | block_size // kernel_block_sizes[0] |
| 1559 | ) # Changed to use first element |
| 1560 | |
| 1561 | # Test block table conversion logic |
| 1562 | # One kvcache_manager block should map to multiple kernel blocks |
| 1563 | kvcache_manager_blocks = [0, 1, 2] |
| 1564 | |
| 1565 | # Verify that kvcache_manager blocks can be converted to kernel blocks |
| 1566 | # and that block table operations work correctly. |
| 1567 | req_index = 0 |
| 1568 | block_table.append_row(kvcache_manager_blocks, req_index) |
| 1569 | # Get expected kernel blocks from the implementation for verification. |
| 1570 | expected_kernel_blocks = block_table.map_to_kernel_blocks( |
| 1571 | np.array(kvcache_manager_blocks), |
| 1572 | block_table.blocks_per_kv_block, |
| 1573 | block_table._kernel_block_arange, |
| 1574 | ) |
| 1575 | # Verify block table state |
| 1576 | assert block_table.num_blocks_per_row[req_index] == len(expected_kernel_blocks) |
| 1577 | assert np.array_equal( |
| 1578 | block_table.block_table.np[req_index, : len(expected_kernel_blocks)], |
| 1579 | expected_kernel_blocks, |
| 1580 | ) |
| 1581 | |
| 1582 | |
| 1583 | def test_get_block_table_width_aligns_to_128_tokens(): |
| 1584 | assert get_block_table_width(1875, 64) == 1876 |
| 1585 | |
| 1586 | |
| 1587 | def test_get_block_table_width_splits_virtual_blocks(): |
| 1588 | assert get_block_table_width(235, 256, 64) == 940 |
| 1589 | |
| 1590 | |
| 1591 | def test_mamba_state_table_width_is_not_aligned(): |
| 1592 | block_tables = MultiGroupBlockTable( |
| 1593 | max_num_reqs=1, |
| 1594 | max_num_batched_tokens=1, |
| 1595 | pin_memory=False, |
| 1596 | device=torch.device("cpu"), |
| 1597 | block_sizes=[39664], |
| 1598 | kernel_block_sizes=[39664], |
| 1599 | max_num_blocks=[1], |
| 1600 | slot_mapping_modes=[SlotMappingMode.NONE], |
| 1601 | ) |
| 1602 | |
| 1603 | assert block_tables[0].max_num_blocks_per_req == 1 |
| 1604 | |
| 1605 | |
| 1606 | def test_input_batch_with_kernel_block_sizes(): |
| 1607 | """Test InputBatch initialization with kernel_block_sizes parameter.""" |
| 1608 | max_num_reqs = 10 |
| 1609 | max_model_len = 512 |
| 1610 | max_num_batched_tokens = 512 |
| 1611 | device = torch.device(DEVICE_TYPE) |
| 1612 | vocab_size = 50272 |
| 1613 | |
| 1614 | # Test with different kernel block sizes |
| 1615 | block_sizes = [32, 64] |
| 1616 | kernel_block_sizes = [16, 32] |
| 1617 | |
| 1618 | input_batch = InputBatch( |
| 1619 | max_num_reqs=max_num_reqs, |
| 1620 | max_model_len=max_model_len, |
| 1621 | max_num_batched_tokens=max_num_batched_tokens, |
| 1622 | device=device, |
| 1623 | vocab_size=vocab_size, |
| 1624 | block_sizes=block_sizes, |
| 1625 | kernel_block_sizes=kernel_block_sizes, |
| 1626 | max_num_blocks_per_req=[16, 8], |
| 1627 | ) |
| 1628 | |
| 1629 | # Verify that block tables were created with kernel block sizes |
| 1630 | assert len(input_batch.block_table.block_tables) == len(block_sizes) |
| 1631 | |
| 1632 | for i, (kv_size, kernel_size) in enumerate(zip(block_sizes, kernel_block_sizes)): |
| 1633 | block_table = input_batch.block_table.block_tables[i] |
| 1634 | if kv_size != kernel_size: |
| 1635 | assert block_table.use_hybrid_blocks is True |
| 1636 | assert block_table.block_size == kernel_size |
| 1637 | else: |
| 1638 | assert block_table.use_hybrid_blocks is False |
| 1639 | assert block_table.block_size == kernel_size |
| 1640 | |
| 1641 | |
| 1642 | def test_hybrid_cache_integration(default_vllm_config, dist_init): |
| 1643 | """Test hybrid cache architecture integration with GPUModelRunner.""" |
| 1644 | # Create a new model runner with hybrid cache configuration |
| 1645 | vllm_config = get_vllm_config() |
| 1646 | |
| 1647 | # Configure hybrid cache with different kvcache_manager block size |
| 1648 | vllm_config.cache_config.block_size = 32 |
| 1649 | |
| 1650 | model_config = vllm_config.model_config |
| 1651 | num_heads = model_config.get_num_kv_heads(vllm_config.parallel_config) |
| 1652 | head_size = model_config.get_head_size() |
| 1653 | vllm_config.compilation_config.static_forward_context["layer.0"] = Attention( |
| 1654 | num_heads, head_size, 0.1 |
| 1655 | ) |
| 1656 | |
| 1657 | runner = GPUModelRunner(vllm_config, DEVICE_TYPE) |
| 1658 | |
| 1659 | # Initialize KV cache with configuration |
| 1660 | attn_spec = FullAttentionSpec( |
| 1661 | block_size=16, # Use kernel block size directly |
| 1662 | num_kv_heads=runner.model_config.get_num_kv_heads(runner.parallel_config), |
| 1663 | head_size=runner.model_config.get_head_size(), |
| 1664 | dtype=runner.kv_cache_dtype, |
| 1665 | ) |
| 1666 | tensor_size = attn_spec.page_size_bytes * NUM_BLOCKS |
| 1667 | kv_cache_config = KVCacheConfig( |
| 1668 | num_blocks=NUM_BLOCKS, |
| 1669 | kv_cache_tensors=[ |
| 1670 | KVCacheTensor( |
| 1671 | size=tensor_size, |
| 1672 | layers=["layer.0"], |
| 1673 | layer_stride=tensor_size, |
| 1674 | block_stride=attn_spec.page_size_bytes, |
| 1675 | ), |
| 1676 | ], |
| 1677 | kv_cache_groups=[ |
| 1678 | KVCacheGroupSpec(layer_names=["layer.0"], kv_cache_spec=attn_spec) |
| 1679 | ], |
| 1680 | ) |
| 1681 | runner.kv_cache_config = kv_cache_config |
| 1682 | |
| 1683 | # Initialize input batch with kernel block sizes |
| 1684 | runner.input_batch = InputBatch( |
| 1685 | max_num_reqs=runner.max_num_reqs, |
| 1686 | max_model_len=runner.max_model_len, |
| 1687 | max_num_batched_tokens=runner.max_num_tokens, |
| 1688 | device=runner.device, |
| 1689 | vocab_size=runner.model_config.get_vocab_size(), |
| 1690 | block_sizes=[kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size], |
| 1691 | kernel_block_sizes=[16], |
| 1692 | max_num_blocks_per_req=[NUM_BLOCKS], |
| 1693 | ) # Use kernel block size |
| 1694 | |
| 1695 | runner.initialize_attn_backend(kv_cache_config) |
| 1696 | |
| 1697 | # Verify hybrid block table configuration |
| 1698 | block_table = runner.input_batch.block_table.block_tables[0] |
| 1699 | assert block_table.block_size == ( |
| 1700 | kv_cache_config.kv_cache_groups[0].kv_cache_spec.block_size |
| 1701 | ) |
| 1702 | |
| 1703 | # Test request processing with hybrid blocks |
| 1704 | req_id = "hybrid_req_0" |
| 1705 | scheduler_output = _schedule_new_request(req_id) |
| 1706 | |
| 1707 | # Update states should work with hybrid blocks |
| 1708 | runner._update_states(scheduler_output) |
| 1709 | assert _is_req_scheduled(runner, req_id) |
| 1710 | assert _is_req_state_block_table_match(runner, req_id) |
| 1711 | |
| 1712 | |
| 1713 | def test_is_uniform_decode() -> None: |
| 1714 | # Normal |
| 1715 | assert GPUModelRunner._is_uniform_decode( |
| 1716 | max_num_scheduled_tokens=1, |
| 1717 | uniform_decode_query_len=1, |
| 1718 | num_tokens=16, |
| 1719 | num_reqs=16, |
| 1720 | ) |
| 1721 | assert not GPUModelRunner._is_uniform_decode( |
| 1722 | max_num_scheduled_tokens=2, |
| 1723 | uniform_decode_query_len=1, |
| 1724 | num_tokens=16, |
| 1725 | num_reqs=16, |
| 1726 | ) |
| 1727 | assert not GPUModelRunner._is_uniform_decode( |
| 1728 | max_num_scheduled_tokens=1, |
| 1729 | uniform_decode_query_len=1, |
| 1730 | num_tokens=16, |
| 1731 | num_reqs=15, |
| 1732 | ) |
| 1733 | # Spec decoding |
| 1734 | assert GPUModelRunner._is_uniform_decode( |
| 1735 | max_num_scheduled_tokens=5, |
| 1736 | uniform_decode_query_len=5, |
| 1737 | num_tokens=30, |
| 1738 | num_reqs=6, |
| 1739 | ) |
| 1740 | assert not GPUModelRunner._is_uniform_decode( |
| 1741 | max_num_scheduled_tokens=5, |
| 1742 | uniform_decode_query_len=4, |
| 1743 | num_tokens=30, |
| 1744 | num_reqs=6, |
| 1745 | ) |
| 1746 | assert not GPUModelRunner._is_uniform_decode( |
| 1747 | max_num_scheduled_tokens=5, |
| 1748 | uniform_decode_query_len=5, |
| 1749 | num_tokens=30, |
| 1750 | num_reqs=7, |
| 1751 | ) |
| 1752 | # Force uniform decode |
| 1753 | assert GPUModelRunner._is_uniform_decode( |
| 1754 | max_num_scheduled_tokens=1, |
| 1755 | uniform_decode_query_len=1, |
| 1756 | num_tokens=16, |
| 1757 | num_reqs=16, |
| 1758 | force_uniform_decode=True, |
| 1759 | ) |
| 1760 | assert GPUModelRunner._is_uniform_decode( |
| 1761 | max_num_scheduled_tokens=2, |
| 1762 | uniform_decode_query_len=1, |
| 1763 | num_tokens=16, |
| 1764 | num_reqs=16, |
| 1765 | force_uniform_decode=True, |
| 1766 | ) |
| 1767 | assert GPUModelRunner._is_uniform_decode( |
| 1768 | max_num_scheduled_tokens=1, |
| 1769 | uniform_decode_query_len=1, |
| 1770 | num_tokens=16, |
| 1771 | num_reqs=15, |
| 1772 | force_uniform_decode=True, |
| 1773 | ) |
| 1774 | assert not GPUModelRunner._is_uniform_decode( |
| 1775 | max_num_scheduled_tokens=1, |
| 1776 | uniform_decode_query_len=1, |
| 1777 | num_tokens=16, |
| 1778 | num_reqs=16, |
| 1779 | force_uniform_decode=False, |
| 1780 | ) |
| 1781 | assert not GPUModelRunner._is_uniform_decode( |
| 1782 | max_num_scheduled_tokens=2, |
| 1783 | uniform_decode_query_len=1, |
| 1784 | num_tokens=16, |
| 1785 | num_reqs=16, |
| 1786 | force_uniform_decode=False, |
| 1787 | ) |
| 1788 | assert not GPUModelRunner._is_uniform_decode( |
| 1789 | max_num_scheduled_tokens=1, |
| 1790 | uniform_decode_query_len=1, |
| 1791 | num_tokens=16, |
| 1792 | num_reqs=15, |
| 1793 | force_uniform_decode=False, |
| 1794 | ) |
| 1795 | |
| 1796 | |
| 1797 | @pytest.mark.skipif( |
| 1798 | not current_platform.is_cuda(), |
| 1799 | reason="Attention backend FLASHINFER is only supported on CUDA.", |
| 1800 | ) |
| 1801 | def test_mamba_cache_raises_when_max_num_seqs_exceeds_blocks(): |
| 1802 | """Test that a ValueError is raised when max_num_seqs exceeds the |
| 1803 | available Mamba cache blocks for hybrid models with FULL cudagraphs. |
| 1804 | |
| 1805 | See: https://github.com/vllm-project/vllm/issues/34094 |
| 1806 | """ |
| 1807 | set_random_seed(42) |
| 1808 | |
| 1809 | update_environment_variables( |
| 1810 | { |
| 1811 | "RANK": "0", |
| 1812 | "LOCAL_RANK": "0", |
| 1813 | "WORLD_SIZE": "1", |
| 1814 | "MASTER_ADDR": "localhost", |
| 1815 | "MASTER_PORT": "12345", |
| 1816 | } |
| 1817 | ) |
| 1818 | from tests.utils import ensure_current_vllm_config |
| 1819 | |
| 1820 | with ensure_current_vllm_config(): |
| 1821 | init_distributed_environment() |
| 1822 | initialize_model_parallel(tensor_model_parallel_size=1) |
| 1823 | torch.set_default_dtype(torch.float16) |
| 1824 | |
| 1825 | model_config = ModelConfig( |
| 1826 | model="ibm-granite/granite-4.0-tiny-preview", |
| 1827 | dtype="float16", |
| 1828 | ) |
| 1829 | scheduler_config = SchedulerConfig( |
| 1830 | max_num_seqs=10, |
| 1831 | max_num_batched_tokens=512, |
| 1832 | max_model_len=512, |
| 1833 | is_encoder_decoder=model_config.is_encoder_decoder, |
| 1834 | ) |
| 1835 | cache_config = CacheConfig( |
| 1836 | block_size=BLOCK_SIZE, |
| 1837 | gpu_memory_utilization=0.9, |
| 1838 | cache_dtype="auto", |
| 1839 | ) |
| 1840 | cache_config.kv_cache_layout = "LBNHC" |
| 1841 | parallel_config = ParallelConfig() |
| 1842 | attention_config = AttentionConfig(backend=AttentionBackendEnum.FLASHINFER) |
| 1843 | vllm_config = VllmConfig( |
| 1844 | model_config=model_config, |
| 1845 | cache_config=cache_config, |
| 1846 | scheduler_config=scheduler_config, |
| 1847 | parallel_config=parallel_config, |
| 1848 | attention_config=attention_config, |
| 1849 | ) |
| 1850 | |
| 1851 | with set_current_vllm_config(vllm_config): |
| 1852 | hf_config = vllm_config.model_config.hf_config |
| 1853 | fwd_context = {} |
| 1854 | for key in ["model.layers.0.self_attn.attn", "model.layers.1.self_attn.attn"]: |
| 1855 | fwd_context[key] = Attention( |
| 1856 | num_heads=model_config.get_num_attention_heads(parallel_config), |
| 1857 | num_kv_heads=model_config.get_num_kv_heads(parallel_config), |
| 1858 | head_size=model_config.get_head_size(), |
| 1859 | scale=1.0, |
| 1860 | prefix=key, |
| 1861 | ) |
| 1862 | for key in [ |
| 1863 | "model.layers.2.mixer", |
| 1864 | "model.layers.3.mixer", |
| 1865 | "model.layers.4.mixer", |
| 1866 | "model.layers.5.mixer", |
| 1867 | ]: |
| 1868 | fwd_context[key] = MambaMixer2( |
| 1869 | hidden_size=hf_config.hidden_size, |
| 1870 | ssm_state_size=hf_config.mamba_d_state, |
| 1871 | conv_kernel_size=hf_config.mamba_d_conv, |
| 1872 | intermediate_size=hf_config.mamba_expand * hf_config.hidden_size, |
| 1873 | use_conv_bias=hf_config.mamba_conv_bias, |
| 1874 | use_bias=hf_config.mamba_proj_bias, |
| 1875 | n_groups=hf_config.mamba_n_groups, |
| 1876 | num_heads=hf_config.mamba_n_heads, |
| 1877 | head_dim=hf_config.mamba_d_head, |
| 1878 | rms_norm_eps=hf_config.rms_norm_eps, |
| 1879 | activation=hf_config.hidden_act, |
| 1880 | cache_config=cache_config, |
| 1881 | model_config=model_config, |
| 1882 | prefix=key, |
| 1883 | ) |
| 1884 | assert fwd_context is not None |
| 1885 | |
| 1886 | runner = GPUModelRunner(vllm_config, DEVICE_TYPE) |
| 1887 | current_platform.update_block_size_for_backend(vllm_config) |
| 1888 | kv_cache_spec = runner.get_kv_cache_spec() |
| 1889 | |
| 1890 | available_memory = 5 * GiB_bytes |
| 1891 | kv_cache_config = get_kv_cache_configs( |
| 1892 | vllm_config, [kv_cache_spec], [available_memory] |
| 1893 | )[0] |
| 1894 | num_blocks = kv_cache_config.num_blocks |
| 1895 | |
| 1896 | # Force max_num_seqs to exceed num_blocks so the check triggers. |
| 1897 | runner.max_num_reqs = num_blocks + 100 |
| 1898 | |
| 1899 | with pytest.raises(ValueError, match="max_num_seqs"): |
| 1900 | runner.initialize_kv_cache(kv_cache_config) |