| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections import defaultdict |
| 5 | from contextlib import contextmanager |
| 6 | from types import SimpleNamespace |
| 7 | from unittest.mock import MagicMock, patch |
| 8 | |
| 9 | import pytest |
| 10 | import torch |
| 11 | |
| 12 | from vllm.config import ( |
| 13 | CompilationConfig, |
| 14 | CUDAGraphMode, |
| 15 | ParallelConfig, |
| 16 | SchedulerConfig, |
| 17 | VllmConfig, |
| 18 | ) |
| 19 | from vllm.distributed.device_communicators import pynccl_allocator |
| 20 | from vllm.v1.kv_cache_interface import KVCacheConfig |
| 21 | from vllm.v1.worker.gpu import cudagraph_utils as gpu_cudagraph_utils |
| 22 | from vllm.v1.worker.gpu.cudagraph_utils import BatchExecutionDescriptor |
| 23 | from vllm.v1.worker.gpu.input_batch import InputBuffers |
| 24 | from vllm.v1.worker.gpu.pcp_manager import PCPManager |
| 25 | from vllm.v1.worker.gpu.spec_decode.autoregressive.cudagraph_utils import ( |
| 26 | SpeculatorCudaGraphManager, |
| 27 | ) |
| 28 | |
| 29 | pytestmark = pytest.mark.cpu_test |
| 30 | |
| 31 | |
| 32 | @pytest.fixture(autouse=True) |
| 33 | def _reset_graph_pool_id(): |
| 34 | pynccl_allocator._graph_pool_id = None |
| 35 | yield |
| 36 | pynccl_allocator._graph_pool_id = None |
| 37 | |
| 38 | |
| 39 | def _create_vllm_config() -> MagicMock: |
| 40 | compilation_config = CompilationConfig( |
| 41 | cudagraph_mode=CUDAGraphMode.FULL, |
| 42 | cudagraph_capture_sizes=[4], |
| 43 | ) |
| 44 | compilation_config.max_cudagraph_capture_size = 4 |
| 45 | compilation_config.post_init_cudagraph_sizes() |
| 46 | |
| 47 | vllm_config = MagicMock(spec=VllmConfig) |
| 48 | vllm_config.compilation_config = compilation_config |
| 49 | vllm_config.scheduler_config = SchedulerConfig.default_factory(max_num_seqs=4) |
| 50 | vllm_config.parallel_config = ParallelConfig() |
| 51 | vllm_config.speculative_config = None |
| 52 | vllm_config.num_speculative_tokens = 0 |
| 53 | return vllm_config |
| 54 | |
| 55 | |
| 56 | def test_full_capture_sets_graph_pool_id_before_cuda_graph(monkeypatch): |
| 57 | """FULL capture must set graph_pool_id before entering torch.cuda.graph(). |
| 58 | |
| 59 | NCCL symmetric memory checks this global during graph capture; without |
| 60 | it, capture fails with: |
| 61 | AssertionError: graph_pool_id is not set under graph capture |
| 62 | """ |
| 63 | graph_pool = object() |
| 64 | monkeypatch.setattr( |
| 65 | gpu_cudagraph_utils, |
| 66 | "get_pp_group", |
| 67 | lambda: SimpleNamespace(is_first_rank=True, is_last_rank=True), |
| 68 | ) |
| 69 | monkeypatch.setattr( |
| 70 | gpu_cudagraph_utils.current_platform, |
| 71 | "get_global_graph_pool", |
| 72 | lambda: graph_pool, |
| 73 | ) |
| 74 | |
| 75 | manager = gpu_cudagraph_utils.CudaGraphManager( |
| 76 | vllm_config=_create_vllm_config(), |
| 77 | device=torch.device("cpu"), |
| 78 | cudagraph_mode=CUDAGraphMode.FULL, |
| 79 | decode_query_len=1, |
| 80 | ) |
| 81 | |
| 82 | desc = BatchExecutionDescriptor( |
| 83 | cg_mode=CUDAGraphMode.FULL, |
| 84 | num_tokens=4, |
| 85 | num_reqs=4, |
| 86 | uniform_token_count=1, |
| 87 | ) |
| 88 | manager._capture_descs[CUDAGraphMode.FULL] = [desc] |
| 89 | |
| 90 | def create_forward_fn(desc, warmup): |
| 91 | return lambda _mode: None |
| 92 | |
| 93 | @contextmanager |
| 94 | def fake_graph_capture(*args, **kwargs): |
| 95 | yield SimpleNamespace(stream=MagicMock()) |
| 96 | |
| 97 | fake_offloader = MagicMock() |
| 98 | |
| 99 | def cuda_graph_enter(*args, **kwargs): |
| 100 | assert pynccl_allocator._graph_pool_id is graph_pool |
| 101 | |
| 102 | mock_cuda_graph_ctx = MagicMock() |
| 103 | mock_cuda_graph_ctx.__enter__ = cuda_graph_enter |
| 104 | mock_cuda_graph_ctx.__exit__ = MagicMock(return_value=False) |
| 105 | |
| 106 | with ( |
| 107 | patch.object(gpu_cudagraph_utils, "graph_capture", fake_graph_capture), |
| 108 | patch.object(gpu_cudagraph_utils, "get_offloader", lambda: fake_offloader), |
| 109 | patch.object(gpu_cudagraph_utils.torch.cuda, "CUDAGraph"), |
| 110 | patch.object( |
| 111 | gpu_cudagraph_utils.torch.cuda, |
| 112 | "graph", |
| 113 | return_value=mock_cuda_graph_ctx, |
| 114 | ) as mock_cuda_graph, |
| 115 | ): |
| 116 | manager.capture(create_forward_fn) |
| 117 | |
| 118 | mock_cuda_graph.assert_called_once() |
| 119 | |
| 120 | |
| 121 | def test_piecewise_capture_uses_pcp_dummy_slot_mappings(): |
| 122 | num_reqs = 32 |
| 123 | num_tokens = 56 |
| 124 | pcp_world_size = 2 |
| 125 | input_buffers = InputBuffers(num_reqs, num_tokens, torch.device("cpu")) |
| 126 | |
| 127 | pcp_block_tables = SimpleNamespace( |
| 128 | num_kv_cache_groups=1, |
| 129 | input_block_tables=(torch.zeros(num_reqs * 2, 1, dtype=torch.int32),), |
| 130 | ) |
| 131 | pcp_manager = PCPManager( |
| 132 | pcp_world_size=pcp_world_size, |
| 133 | pcp_rank=0, |
| 134 | device=torch.device("cpu"), |
| 135 | shard_decode_requests=False, |
| 136 | max_num_reqs=num_reqs, |
| 137 | max_num_tokens=num_tokens, |
| 138 | block_tables=pcp_block_tables, |
| 139 | ) |
| 140 | |
| 141 | block_tables = MagicMock() |
| 142 | block_tables.cp_size = 1 |
| 143 | block_tables.get_dummy_block_tables.return_value = () |
| 144 | block_tables.get_dummy_slot_mappings.return_value = torch.zeros( |
| 145 | 1, num_tokens, dtype=torch.int64 |
| 146 | ) |
| 147 | model_state = MagicMock() |
| 148 | model_state.prepare_attn.return_value = {} |
| 149 | kv_cache_config = KVCacheConfig( |
| 150 | num_blocks=0, |
| 151 | kv_cache_tensors=[], |
| 152 | kv_cache_groups=[], |
| 153 | ) |
| 154 | |
| 155 | gpu_cudagraph_utils.prepare_inputs_to_capture( |
| 156 | num_reqs, |
| 157 | num_tokens, |
| 158 | model_state, |
| 159 | input_buffers, |
| 160 | block_tables, |
| 161 | [], |
| 162 | kv_cache_config, |
| 163 | full_cudagraph=False, |
| 164 | pcp_manager=pcp_manager, |
| 165 | ) |
| 166 | |
| 167 | slot_mappings = model_state.prepare_attn.call_args.args[3] |
| 168 | assert slot_mappings.shape == (1, num_tokens * pcp_world_size) |
| 169 | block_tables.get_dummy_slot_mappings.assert_not_called() |
| 170 | |
| 171 | |
| 172 | @pytest.mark.parametrize( |
| 173 | "full_cudagraph,max_query_len,expected_max_query_len", |
| 174 | [(False, None, None), (True, None, 8), (True, 1, 1), (True, 3, 3)], |
| 175 | ) |
| 176 | def test_capture_query_length_covers_replay_batches( |
| 177 | full_cudagraph, max_query_len, expected_max_query_len |
| 178 | ): |
| 179 | """Mixed FULL graphs must capture prefill kernels as well as decode kernels.""" |
| 180 | num_tokens = num_reqs = 8 |
| 181 | buffers = InputBuffers(num_reqs, num_tokens, torch.device("cpu")) |
| 182 | block_tables = MagicMock() |
| 183 | block_tables.cp_size = 1 |
| 184 | block_tables.get_dummy_block_tables.return_value = () |
| 185 | block_tables.get_dummy_slot_mappings.return_value = torch.empty( |
| 186 | 0, num_tokens, dtype=torch.int64 |
| 187 | ) |
| 188 | model_state = MagicMock() |
| 189 | kv_cache_config = KVCacheConfig( |
| 190 | num_blocks=0, kv_cache_tensors=[], kv_cache_groups=[] |
| 191 | ) |
| 192 | |
| 193 | gpu_cudagraph_utils.prepare_inputs_to_capture( |
| 194 | num_reqs, |
| 195 | num_tokens, |
| 196 | model_state, |
| 197 | buffers, |
| 198 | block_tables, |
| 199 | [], |
| 200 | kv_cache_config, |
| 201 | full_cudagraph=full_cudagraph, |
| 202 | max_query_len=max_query_len, |
| 203 | ) |
| 204 | |
| 205 | batch = model_state.prepare_attn.call_args.args[0] |
| 206 | assert batch.num_scheduled_tokens.max() == 1 |
| 207 | assert batch.max_query_len == expected_max_query_len |
| 208 | |
| 209 | |
| 210 | @pytest.mark.parametrize( |
| 211 | "uniform_token_count,max_query_len,expected_max_query_len", |
| 212 | [(1, None, 1), (None, 3, 3), (None, None, 8)], |
| 213 | ) |
| 214 | def test_speculator_capture_preserves_decode_query_bounds( |
| 215 | monkeypatch, uniform_token_count, max_query_len, expected_max_query_len |
| 216 | ): |
| 217 | """Draft decode graphs keep narrow queries while mixed graphs cover prefill.""" |
| 218 | num_tokens = num_reqs = 8 |
| 219 | desc = BatchExecutionDescriptor( |
| 220 | CUDAGraphMode.FULL, |
| 221 | num_tokens, |
| 222 | num_reqs, |
| 223 | uniform_token_count=uniform_token_count, |
| 224 | max_query_len=max_query_len, |
| 225 | ) |
| 226 | manager = SpeculatorCudaGraphManager.__new__(SpeculatorCudaGraphManager) |
| 227 | manager.max_num_reqs = num_reqs |
| 228 | manager.vllm_config = SimpleNamespace( |
| 229 | parallel_config=SimpleNamespace(data_parallel_size=1) |
| 230 | ) |
| 231 | buffers = InputBuffers(num_reqs, num_tokens, torch.device("cpu")) |
| 232 | block_tables = MagicMock() |
| 233 | block_tables.cp_size = 1 |
| 234 | block_tables.get_dummy_block_tables.return_value = () |
| 235 | block_tables.get_dummy_slot_mappings.return_value = torch.empty( |
| 236 | 0, num_tokens, dtype=torch.int64 |
| 237 | ) |
| 238 | model_state = MagicMock() |
| 239 | kv_cache_config = KVCacheConfig( |
| 240 | num_blocks=0, kv_cache_tensors=[], kv_cache_groups=[] |
| 241 | ) |
| 242 | |
| 243 | def capture(self, create_forward_fn, progress_bar_desc): |
| 244 | for warmup in (True, False): |
| 245 | create_forward_fn(desc, warmup)(desc.cg_mode) |
| 246 | |
| 247 | monkeypatch.setattr(gpu_cudagraph_utils.CudaGraphManager, "capture", capture) |
| 248 | manager.capture( |
| 249 | MagicMock(), model_state, buffers, block_tables, [], kv_cache_config |
| 250 | ) |
| 251 | |
| 252 | assert model_state.prepare_attn.call_count == 2 |
| 253 | for call in model_state.prepare_attn.call_args_list: |
| 254 | assert call.args[0].max_query_len == expected_max_query_len |
| 255 | |
| 256 | |
| 257 | _DECODE_QUERY_LEN = 3 |
| 258 | |
| 259 | |
| 260 | def _create_decode_vllm_config( |
| 261 | capture_sizes: list[int], |
| 262 | num_speculative_tokens: int = 0, |
| 263 | dynamic_spec_schedule: list[tuple[int, int, int]] | None = None, |
| 264 | max_num_seqs: int = 8, |
| 265 | use_kda_recoverssm: bool = False, |
| 266 | ) -> MagicMock: |
| 267 | compilation_config = CompilationConfig( |
| 268 | cudagraph_mode=CUDAGraphMode.FULL_AND_PIECEWISE, |
| 269 | cudagraph_capture_sizes=capture_sizes, |
| 270 | ) |
| 271 | compilation_config.max_cudagraph_capture_size = capture_sizes[-1] |
| 272 | compilation_config.post_init_cudagraph_sizes() |
| 273 | |
| 274 | vllm_config = MagicMock(spec=VllmConfig) |
| 275 | vllm_config.compilation_config = compilation_config |
| 276 | vllm_config.scheduler_config = SchedulerConfig.default_factory( |
| 277 | max_num_seqs=max_num_seqs |
| 278 | ) |
| 279 | vllm_config.parallel_config = ParallelConfig() |
| 280 | vllm_config.cache_config = SimpleNamespace(use_kda_recoverssm=use_kda_recoverssm) |
| 281 | vllm_config.num_speculative_tokens = num_speculative_tokens |
| 282 | if dynamic_spec_schedule is None: |
| 283 | vllm_config.speculative_config = None |
| 284 | else: |
| 285 | speculative_config = MagicMock() |
| 286 | speculative_config.uses_dynamic_speculative_decoding.return_value = True |
| 287 | speculative_config.num_speculative_tokens_per_batch_size = dynamic_spec_schedule |
| 288 | vllm_config.speculative_config = speculative_config |
| 289 | return vllm_config |
| 290 | |
| 291 | |
| 292 | def _make_spec_decode_manager( |
| 293 | monkeypatch, |
| 294 | decode_query_len: int = _DECODE_QUERY_LEN, |
| 295 | capture_sizes: list[int] | None = None, |
| 296 | num_speculative_tokens: int = 0, |
| 297 | dynamic_spec_schedule: list[tuple[int, int, int]] | None = None, |
| 298 | max_num_seqs: int = 8, |
| 299 | use_kda_recoverssm: bool = False, |
| 300 | varlen_decode: bool = False, |
| 301 | ) -> gpu_cudagraph_utils.CudaGraphManager: |
| 302 | monkeypatch.setattr( |
| 303 | gpu_cudagraph_utils, |
| 304 | "get_pp_group", |
| 305 | lambda: SimpleNamespace(is_first_rank=True, is_last_rank=True), |
| 306 | ) |
| 307 | monkeypatch.setattr( |
| 308 | gpu_cudagraph_utils.current_platform, |
| 309 | "get_global_graph_pool", |
| 310 | lambda: object(), |
| 311 | ) |
| 312 | manager = gpu_cudagraph_utils.CudaGraphManager( |
| 313 | vllm_config=_create_decode_vllm_config( |
| 314 | capture_sizes or [1, 2, 4, 8, 16, 24], |
| 315 | num_speculative_tokens=num_speculative_tokens, |
| 316 | dynamic_spec_schedule=dynamic_spec_schedule, |
| 317 | max_num_seqs=max_num_seqs, |
| 318 | use_kda_recoverssm=use_kda_recoverssm, |
| 319 | ), |
| 320 | device=torch.device("cpu"), |
| 321 | cudagraph_mode=CUDAGraphMode.FULL_AND_PIECEWISE, |
| 322 | decode_query_len=decode_query_len, |
| 323 | varlen_decode=varlen_decode, |
| 324 | ) |
| 325 | manager._graphs_captured = True |
| 326 | return manager |
| 327 | |
| 328 | |
| 329 | def test_uniform_decode_pads_up_to_full_graph(monkeypatch): |
| 330 | manager = _make_spec_decode_manager(monkeypatch) |
| 331 | assert [ |
| 332 | (desc.cg_mode, desc.num_tokens) for desc in manager._candidates[(12, 0)] |
| 333 | ] == [ |
| 334 | (CUDAGraphMode.FULL, 18), |
| 335 | (CUDAGraphMode.PIECEWISE, 16), |
| 336 | ] |
| 337 | |
| 338 | desc = manager.dispatch( |
| 339 | num_reqs=4, |
| 340 | num_tokens=12, |
| 341 | uniform_token_count=_DECODE_QUERY_LEN, |
| 342 | num_active_loras=0, |
| 343 | ) |
| 344 | |
| 345 | assert desc.cg_mode == CUDAGraphMode.FULL |
| 346 | assert desc.uniform_token_count == _DECODE_QUERY_LEN |
| 347 | assert desc.num_tokens == 18 |
| 348 | assert desc.num_reqs == 6 |
| 349 | |
| 350 | |
| 351 | def test_uniform_decode_exact_match_is_not_over_padded(monkeypatch): |
| 352 | manager = _make_spec_decode_manager(monkeypatch) |
| 353 | |
| 354 | desc = manager.dispatch( |
| 355 | num_reqs=3, |
| 356 | num_tokens=9, |
| 357 | uniform_token_count=_DECODE_QUERY_LEN, |
| 358 | num_active_loras=0, |
| 359 | ) |
| 360 | |
| 361 | assert desc.cg_mode == CUDAGraphMode.FULL |
| 362 | assert desc.num_tokens == 9 |
| 363 | assert desc.num_reqs == 3 |
| 364 | |
| 365 | |
| 366 | def test_mixed_batch_never_selects_a_uniform_decode_graph(monkeypatch): |
| 367 | manager = _make_spec_decode_manager(monkeypatch) |
| 368 | |
| 369 | desc = manager.dispatch( |
| 370 | num_reqs=2, |
| 371 | num_tokens=12, |
| 372 | uniform_token_count=None, |
| 373 | num_active_loras=0, |
| 374 | ) |
| 375 | |
| 376 | assert desc.cg_mode == CUDAGraphMode.PIECEWISE |
| 377 | assert desc.uniform_token_count is None |
| 378 | assert desc.num_tokens == 16 |
| 379 | |
| 380 | |
| 381 | def test_varlen_decode_graph_takes_bounded_decode_batches_only(monkeypatch): |
| 382 | """A varlen decode graph replays any decode batch within its query-length |
| 383 | bound. The runner passes no bound for a batch with a prefill, which must not |
| 384 | replay one however short the prefill is.""" |
| 385 | manager = _make_spec_decode_manager(monkeypatch, varlen_decode=True) |
| 386 | |
| 387 | def dispatch(max_query_len): |
| 388 | return manager.dispatch( |
| 389 | num_reqs=3, |
| 390 | num_tokens=6, |
| 391 | uniform_token_count=None, |
| 392 | num_active_loras=0, |
| 393 | max_query_len=max_query_len, |
| 394 | ) |
| 395 | |
| 396 | desc = dispatch(_DECODE_QUERY_LEN) |
| 397 | assert desc.cg_mode == CUDAGraphMode.FULL |
| 398 | assert desc.max_query_len == _DECODE_QUERY_LEN |
| 399 | assert dispatch(None).cg_mode == CUDAGraphMode.PIECEWISE |
| 400 | assert dispatch(_DECODE_QUERY_LEN + 1).cg_mode == CUDAGraphMode.PIECEWISE |
| 401 | |
| 402 | |
| 403 | def test_mixed_batch_at_decode_only_token_count_still_gets_a_graph(monkeypatch): |
| 404 | """A mixed batch must not fall to eager where only decode graphs are staged. |
| 405 | |
| 406 | ``round_up(size, decode_query_len)`` lands FULL decode graphs on token counts |
| 407 | the PIECEWISE ladder never uses -- with capture sizes [1, 2, 4, 8, 16, 24] |
| 408 | and query length 3, FULL covers 3, 6, 9 and 18 while PIECEWISE has only |
| 409 | 1, 2, 4, 8, 16, 24. Building each mode's candidate ranges independently keeps |
| 410 | a PIECEWISE graph reachable there. Deriving the ranges from the staged sizes |
| 411 | instead offers a mixed batch nothing but decode descriptors, whose |
| 412 | ``uniform_token_count`` it can never match, so it runs fully eager. |
| 413 | """ |
| 414 | manager = _make_spec_decode_manager(monkeypatch) |
| 415 | |
| 416 | decode_only_token_counts = sorted( |
| 417 | {desc.num_tokens for desc in manager._capture_descs[CUDAGraphMode.FULL]} |
| 418 | - {desc.num_tokens for desc in manager._capture_descs[CUDAGraphMode.PIECEWISE]} |
| 419 | ) |
| 420 | # Guard the premise: with no such token counts this test would cover nothing. |
| 421 | assert decode_only_token_counts |
| 422 | |
| 423 | for num_tokens in decode_only_token_counts: |
| 424 | desc = manager.dispatch( |
| 425 | num_reqs=1, |
| 426 | num_tokens=num_tokens, |
| 427 | uniform_token_count=None, |
| 428 | num_active_loras=0, |
| 429 | ) |
| 430 | assert desc.cg_mode == CUDAGraphMode.PIECEWISE, num_tokens |
| 431 | assert desc.uniform_token_count is None, num_tokens |
| 432 | assert desc.num_tokens >= num_tokens, num_tokens |
| 433 | |
| 434 | |
| 435 | def test_uniform_decode_beyond_capture_ladder_falls_back(monkeypatch): |
| 436 | manager = _make_spec_decode_manager(monkeypatch) |
| 437 | |
| 438 | desc = manager.dispatch( |
| 439 | num_reqs=9, |
| 440 | num_tokens=27, |
| 441 | uniform_token_count=_DECODE_QUERY_LEN, |
| 442 | num_active_loras=0, |
| 443 | ) |
| 444 | |
| 445 | assert desc.cg_mode == CUDAGraphMode.NONE |
| 446 | |
| 447 | |
| 448 | def test_dynamic_spec_decode_shared_token_count_stays_reachable(monkeypatch): |
| 449 | """Every captured decode graph must remain reachable from dispatch(). |
| 450 | |
| 451 | Under dynamic speculative decoding ``decode_query_lens`` is a list, and the |
| 452 | staging loop rounds each capture size up to *every* query length. Several |
| 453 | decode graphs therefore land on the same ``num_tokens`` -- e.g. capture size |
| 454 | 2 stages both ``round_up(2, 1) == 2`` and ``round_up(2, 2) == 2``. Building |
| 455 | the candidate ranges one descriptor at a time would give all but the first |
| 456 | of those an empty range, so they would never enter a candidate list and |
| 457 | their batches would silently fall through to PIECEWISE. |
| 458 | """ |
| 459 | manager = _make_spec_decode_manager( |
| 460 | monkeypatch, |
| 461 | decode_query_len=4, |
| 462 | capture_sizes=[2, 4, 6, 8], |
| 463 | num_speculative_tokens=3, |
| 464 | # max_num_seqs is 8, and K narrows as the batch grows, so the schedule |
| 465 | # covers K 3, 2, 1 and 0, giving decode_query_lens [1, 2, 3, 4]. |
| 466 | dynamic_spec_schedule=[(1, 2, 3), (3, 4, 2), (5, 6, 1), (7, 8, 0)], |
| 467 | ) |
| 468 | |
| 469 | full_descs = manager._capture_descs[CUDAGraphMode.FULL] |
| 470 | by_num_tokens: dict[int, list] = defaultdict(list) |
| 471 | for desc in full_descs: |
| 472 | by_num_tokens[desc.num_tokens].append(desc) |
| 473 | # Guard the premise: without collisions this test would not cover the bug. |
| 474 | assert any(len(descs) > 1 for descs in by_num_tokens.values()) |
| 475 | |
| 476 | for desc in full_descs: |
| 477 | assert desc.num_reqs is not None |
| 478 | assert desc in manager._candidates[(desc.num_tokens, 0)], desc |
| 479 | assert ( |
| 480 | manager.dispatch( |
| 481 | num_reqs=desc.num_reqs, |
| 482 | num_tokens=desc.num_tokens, |
| 483 | uniform_token_count=desc.uniform_token_count, |
| 484 | num_active_loras=0, |
| 485 | ) |
| 486 | == desc |
| 487 | ), desc |
| 488 | |
| 489 | |
| 490 | @pytest.mark.parametrize( |
| 491 | "decode_query_len,capture_sizes", |
| 492 | [(1, [1, 2, 4, 8]), (2, [2, 4, 8, 16]), (8, [8, 16, 32, 64])], |
| 493 | ) |
| 494 | def test_divisor_query_len_dispatch_is_unchanged( |
| 495 | monkeypatch, decode_query_len, capture_sizes |
| 496 | ): |
| 497 | manager = _make_spec_decode_manager( |
| 498 | monkeypatch, |
| 499 | decode_query_len=decode_query_len, |
| 500 | capture_sizes=capture_sizes, |
| 501 | ) |
| 502 | |
| 503 | max_reqs = capture_sizes[-1] // decode_query_len |
| 504 | for num_reqs in range(1, max_reqs + 1): |
| 505 | num_tokens = num_reqs * decode_query_len |
| 506 | desc = manager.dispatch( |
| 507 | num_reqs=num_reqs, |
| 508 | num_tokens=num_tokens, |
| 509 | uniform_token_count=decode_query_len, |
| 510 | num_active_loras=0, |
| 511 | ) |
| 512 | expected = min(s for s in capture_sizes if s >= num_tokens) |
| 513 | assert desc.cg_mode == CUDAGraphMode.FULL, num_tokens |
| 514 | assert desc.num_tokens == expected, num_tokens |
| 515 | assert desc.num_reqs == expected // decode_query_len, num_tokens |