| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from dataclasses import replace |
| 4 | from unittest.mock import MagicMock, patch |
| 5 | |
| 6 | import pytest |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | |
| 10 | from tests.utils import create_new_process_for_each_test |
| 11 | from vllm.compilation.cuda_graph import CUDAGraphWrapper |
| 12 | from vllm.compilation.monitor import set_cudagraph_capturing_enabled |
| 13 | from vllm.config import ( |
| 14 | CompilationConfig, |
| 15 | CompilationMode, |
| 16 | CUDAGraphMode, |
| 17 | ParallelConfig, |
| 18 | SchedulerConfig, |
| 19 | VllmConfig, |
| 20 | ) |
| 21 | from vllm.config.lora import LoRAConfig |
| 22 | from vllm.forward_context import BatchDescriptor, set_forward_context |
| 23 | from vllm.platforms import current_platform |
| 24 | from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher |
| 25 | |
| 26 | DEVICE_TYPE = current_platform.device_type |
| 27 | |
| 28 | |
| 29 | # Helper MLP for testing |
| 30 | class SimpleMLP(nn.Module): |
| 31 | def __init__(self): |
| 32 | super().__init__() |
| 33 | self.fc1 = nn.Linear(10, 10) |
| 34 | self.fc2 = nn.Linear(10, 10) |
| 35 | |
| 36 | def forward(self, x): |
| 37 | return self.fc2(self.fc1(x)) |
| 38 | |
| 39 | |
| 40 | def _create_vllm_config( |
| 41 | compilation_config: CompilationConfig, |
| 42 | max_num_seqs: int = 8, |
| 43 | lora_config: bool = False, |
| 44 | ) -> MagicMock: |
| 45 | mock_config = MagicMock(spec=VllmConfig) |
| 46 | mock_config.compilation_config = compilation_config |
| 47 | mock_config.scheduler_config = SchedulerConfig.default_factory( |
| 48 | max_num_seqs=max_num_seqs, |
| 49 | ) |
| 50 | mock_config.parallel_config = ParallelConfig() |
| 51 | mock_config.speculative_config = None # No speculative decoding |
| 52 | mock_config.num_speculative_tokens = 0 |
| 53 | if not lora_config: |
| 54 | mock_config.lora_config = None |
| 55 | else: |
| 56 | # Create a real LoRAConfig with specialize_active_lora enabled |
| 57 | mock_config.lora_config = LoRAConfig( |
| 58 | max_loras=4, |
| 59 | specialize_active_lora=True, |
| 60 | ) |
| 61 | # Mimic the behavior of VllmConfig.__post_init__() |
| 62 | if compilation_config.mode == CompilationMode.VLLM_COMPILE: |
| 63 | compilation_config.set_splitting_ops_for_v1( |
| 64 | all2all_backend=mock_config.parallel_config.all2all_backend, |
| 65 | data_parallel_size=mock_config.parallel_config.data_parallel_size, |
| 66 | ) |
| 67 | |
| 68 | # mimic VllmConfig.__post_init__ |
| 69 | if compilation_config.cudagraph_capture_sizes: |
| 70 | compilation_config.max_cudagraph_capture_size = ( |
| 71 | compilation_config.cudagraph_capture_sizes[-1] |
| 72 | ) |
| 73 | |
| 74 | compilation_config.post_init_cudagraph_sizes() |
| 75 | |
| 76 | return mock_config |
| 77 | |
| 78 | |
| 79 | class TestCudagraphDispatcher: |
| 80 | @pytest.mark.parametrize( |
| 81 | "cudagraph_mode_str,compilation_mode,lora_config", |
| 82 | [ |
| 83 | # Test case 0: Full CG for mixed batches, no separate routine |
| 84 | ("FULL", CompilationMode.NONE, False), |
| 85 | # Test case 1: Full CG for uniform batches, piecewise for mixed |
| 86 | ("FULL_AND_PIECEWISE", CompilationMode.NONE, False), |
| 87 | # Test case 2: Full CG for uniform batches, no CG for mixed |
| 88 | ("FULL_DECODE_ONLY", CompilationMode.NONE, False), |
| 89 | # Test case 3: PIECEWISE for all |
| 90 | ("PIECEWISE", CompilationMode.VLLM_COMPILE, False), |
| 91 | # Test case 4: PIECEWISE for all, specialize LoRA cases |
| 92 | ("PIECEWISE", CompilationMode.VLLM_COMPILE, True), |
| 93 | ], |
| 94 | ) |
| 95 | def test_dispatcher(self, cudagraph_mode_str, compilation_mode, lora_config): |
| 96 | # Setup dispatcher |
| 97 | comp_config = CompilationConfig( |
| 98 | cudagraph_mode=cudagraph_mode_str, |
| 99 | mode=compilation_mode, |
| 100 | cudagraph_capture_sizes=[1, 8], |
| 101 | ) |
| 102 | |
| 103 | config = _create_vllm_config( |
| 104 | comp_config, max_num_seqs=8, lora_config=lora_config |
| 105 | ) |
| 106 | if ( |
| 107 | cudagraph_mode_str == "FULL_AND_PIECEWISE" |
| 108 | and compilation_mode == CompilationMode.NONE |
| 109 | ): |
| 110 | with pytest.raises(AssertionError): |
| 111 | dispatcher = CudagraphDispatcher(config) |
| 112 | return |
| 113 | |
| 114 | dispatcher = CudagraphDispatcher(config) |
| 115 | dispatcher.initialize_cudagraph_keys( |
| 116 | cudagraph_mode=comp_config.cudagraph_mode, uniform_decode_query_len=1 |
| 117 | ) |
| 118 | |
| 119 | # Verify the key is initialized correctly |
| 120 | # With LoRA specialization (max_loras=4, specialize_active_lora=True): |
| 121 | # - lora_cases = [0, 1, 2, 4, 5] (no-lora + powers of 2 up to 4 + max_loras+1) |
| 122 | # - capture_sizes = [1, 8] |
| 123 | # - Total keys = 2 sizes × 5 lora_cases = 10 |
| 124 | if cudagraph_mode_str in ["FULL_AND_PIECEWISE", "PIECEWISE"]: |
| 125 | assert len(dispatcher.cudagraph_keys[CUDAGraphMode.PIECEWISE]) == ( |
| 126 | 10 if lora_config else 2 |
| 127 | ) |
| 128 | else: |
| 129 | assert len(dispatcher.cudagraph_keys[CUDAGraphMode.PIECEWISE]) == 0 |
| 130 | if cudagraph_mode_str not in ["NONE", "PIECEWISE"]: |
| 131 | assert len(dispatcher.cudagraph_keys[CUDAGraphMode.FULL]) == ( |
| 132 | 10 if lora_config else 2 |
| 133 | ) |
| 134 | else: |
| 135 | assert len(dispatcher.cudagraph_keys[CUDAGraphMode.FULL]) == 0 |
| 136 | |
| 137 | # Test dispatch logic |
| 138 | # 1. non-uniform batch, size in cudagraph size list |
| 139 | # FULL mode uses exact keys with num_reqs set |
| 140 | desc_full_with_reqs = BatchDescriptor(num_tokens=8, num_reqs=8, uniform=False) |
| 141 | # PIECEWISE mode uses relaxed keys with num_reqs=None |
| 142 | desc_piecewise = BatchDescriptor(num_tokens=8, num_reqs=None, uniform=False) |
| 143 | rt_mode, key = dispatcher.dispatch( |
| 144 | num_tokens=8, uniform_decode=False, has_lora=False |
| 145 | ) |
| 146 | if cudagraph_mode_str == "FULL": |
| 147 | assert rt_mode == CUDAGraphMode.FULL |
| 148 | assert key == desc_full_with_reqs |
| 149 | elif cudagraph_mode_str in ["FULL_AND_PIECEWISE", "PIECEWISE"]: |
| 150 | assert rt_mode == CUDAGraphMode.PIECEWISE |
| 151 | assert key == desc_piecewise |
| 152 | else: |
| 153 | assert rt_mode == CUDAGraphMode.NONE |
| 154 | |
| 155 | # 2. uniform decode batch, size in cudagraph size list |
| 156 | desc_uniform_exact = BatchDescriptor(num_tokens=8, num_reqs=8, uniform=True) |
| 157 | desc_non_uniform = BatchDescriptor(num_tokens=8, num_reqs=8, uniform=False) |
| 158 | rt_mode, key = dispatcher.dispatch( |
| 159 | num_tokens=8, uniform_decode=True, has_lora=False |
| 160 | ) |
| 161 | if cudagraph_mode_str == "FULL": |
| 162 | # Pure FULL mode uses non-uniform keys for all batches |
| 163 | assert rt_mode == CUDAGraphMode.FULL |
| 164 | assert key == desc_non_uniform |
| 165 | elif cudagraph_mode_str in ["FULL_DECODE_ONLY", "FULL_AND_PIECEWISE"]: |
| 166 | # These modes have separate uniform decode keys |
| 167 | assert rt_mode == CUDAGraphMode.FULL |
| 168 | assert key == desc_uniform_exact |
| 169 | elif cudagraph_mode_str == "PIECEWISE": |
| 170 | assert rt_mode == CUDAGraphMode.PIECEWISE |
| 171 | assert key == replace(desc_uniform_exact, num_reqs=None, uniform=False) |
| 172 | else: |
| 173 | assert rt_mode == CUDAGraphMode.NONE |
| 174 | |
| 175 | # 3. No key match |
| 176 | rt_mode, key = dispatcher.dispatch( |
| 177 | num_tokens=15, uniform_decode=False, has_lora=False |
| 178 | ) |
| 179 | assert rt_mode == CUDAGraphMode.NONE |
| 180 | assert key == BatchDescriptor(num_tokens=15) |
| 181 | |
| 182 | # 4. invalid_modes={FULL} should have a fall back mode |
| 183 | # (e.g., cascade attention) |
| 184 | desc_full_exact = BatchDescriptor(num_tokens=8, uniform=False) |
| 185 | rt_mode, key = dispatcher.dispatch( |
| 186 | num_tokens=8, |
| 187 | uniform_decode=False, |
| 188 | has_lora=False, |
| 189 | invalid_modes={CUDAGraphMode.FULL}, |
| 190 | ) |
| 191 | |
| 192 | if "PIECEWISE" in cudagraph_mode_str: # string contains check |
| 193 | assert rt_mode == CUDAGraphMode.PIECEWISE |
| 194 | assert key == replace(desc_full_exact, num_reqs=None, uniform=False) |
| 195 | else: |
| 196 | assert rt_mode == CUDAGraphMode.NONE |
| 197 | |
| 198 | # 5. valid_modes={NONE} always returns NONE even when keys exist |
| 199 | rt_mode, key = dispatcher.dispatch( |
| 200 | num_tokens=8, |
| 201 | uniform_decode=False, |
| 202 | has_lora=False, |
| 203 | valid_modes={CUDAGraphMode.NONE}, |
| 204 | ) |
| 205 | assert rt_mode == CUDAGraphMode.NONE |
| 206 | assert key == BatchDescriptor(num_tokens=8) |
| 207 | |
| 208 | @pytest.mark.parametrize( |
| 209 | "cudagraph_mode_str,compilation_mode,expected_modes", |
| 210 | [ |
| 211 | # FULL mode: only FULL keys, no PIECEWISE |
| 212 | ("FULL", CompilationMode.NONE, [CUDAGraphMode.FULL]), |
| 213 | # PIECEWISE mode: only PIECEWISE keys |
| 214 | ("PIECEWISE", CompilationMode.VLLM_COMPILE, [CUDAGraphMode.PIECEWISE]), |
| 215 | # FULL_DECODE_ONLY: only FULL keys for uniform decode |
| 216 | ("FULL_DECODE_ONLY", CompilationMode.NONE, [CUDAGraphMode.FULL]), |
| 217 | # NONE mode: no keys |
| 218 | ("NONE", CompilationMode.NONE, []), |
| 219 | ], |
| 220 | ) |
| 221 | def test_get_capture_descs( |
| 222 | self, cudagraph_mode_str, compilation_mode, expected_modes |
| 223 | ): |
| 224 | """Test get_capture_descs returns correctly grouped and ordered descs.""" |
| 225 | comp_config = CompilationConfig( |
| 226 | cudagraph_mode=cudagraph_mode_str, |
| 227 | mode=compilation_mode, |
| 228 | cudagraph_capture_sizes=[1, 4, 8, 16], |
| 229 | ) |
| 230 | |
| 231 | config = _create_vllm_config(comp_config, max_num_seqs=16) |
| 232 | dispatcher = CudagraphDispatcher(config) |
| 233 | dispatcher.initialize_cudagraph_keys( |
| 234 | cudagraph_mode=comp_config.cudagraph_mode, uniform_decode_query_len=1 |
| 235 | ) |
| 236 | |
| 237 | capture_descs = dispatcher.get_capture_descs() |
| 238 | |
| 239 | # Verify we get the expected modes |
| 240 | actual_modes = [mode for mode, _ in capture_descs] |
| 241 | assert actual_modes == expected_modes |
| 242 | |
| 243 | # Verify each group is sorted largest-first |
| 244 | for mode, descs in capture_descs: |
| 245 | assert len(descs) > 0, "Each group should have at least one descriptor" |
| 246 | num_tokens_list = [d.num_tokens for d in descs] |
| 247 | assert num_tokens_list == sorted(num_tokens_list, reverse=True), ( |
| 248 | f"Descriptors for {mode} should be sorted largest-first" |
| 249 | ) |
| 250 | |
| 251 | # All descriptors in a group should have same uniform value |
| 252 | uniform_values = [d.uniform for d in descs] |
| 253 | assert len(set(uniform_values)) == 1, ( |
| 254 | "All descriptors in a group should have the same uniform value" |
| 255 | ) |
| 256 | |
| 257 | def test_get_capture_descs_empty_when_not_initialized(self): |
| 258 | """Test that get_capture_descs returns empty list when keys not initialized.""" |
| 259 | comp_config = CompilationConfig( |
| 260 | cudagraph_mode=CUDAGraphMode.FULL, |
| 261 | mode=CompilationMode.NONE, |
| 262 | cudagraph_capture_sizes=[1, 8], |
| 263 | ) |
| 264 | config = _create_vllm_config(comp_config, max_num_seqs=8) |
| 265 | dispatcher = CudagraphDispatcher(config) |
| 266 | # Don't initialize keys |
| 267 | |
| 268 | assert dispatcher.get_capture_descs() == [] |
| 269 | |
| 270 | |
| 271 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="Skip if not cuda") |
| 272 | class TestCUDAGraphWrapper: |
| 273 | def setup_method(self): |
| 274 | self.vllm_config = _create_vllm_config(CompilationConfig()) |
| 275 | self.model = SimpleMLP().to(DEVICE_TYPE) |
| 276 | self.persistent_input_buffer = torch.zeros(1, 10, device=DEVICE_TYPE) |
| 277 | self.input_tensor = torch.randn(1, 10, device=DEVICE_TYPE) |
| 278 | |
| 279 | def test_capture_and_replay(self): |
| 280 | wrapper = CUDAGraphWrapper( |
| 281 | self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL |
| 282 | ) |
| 283 | batch_descriptor = BatchDescriptor(num_tokens=10) |
| 284 | |
| 285 | # 0. global warmup |
| 286 | with set_forward_context( |
| 287 | attn_metadata=None, |
| 288 | vllm_config=self.vllm_config, |
| 289 | cudagraph_runtime_mode=CUDAGraphMode.NONE, |
| 290 | batch_descriptor=None, |
| 291 | ): |
| 292 | wrapper(self.input_tensor) |
| 293 | |
| 294 | # 1. Capture |
| 295 | with ( |
| 296 | set_forward_context( |
| 297 | attn_metadata=None, |
| 298 | vllm_config=self.vllm_config, |
| 299 | cudagraph_runtime_mode=CUDAGraphMode.FULL, |
| 300 | batch_descriptor=batch_descriptor, |
| 301 | ), |
| 302 | patch("torch.cuda.graph", wraps=torch.cuda.graph) as mock_cuda_graph, |
| 303 | ): |
| 304 | output1 = wrapper(self.input_tensor) |
| 305 | # capturing phase should generate a zero output |
| 306 | assert torch.allclose(output1, torch.zeros_like(output1)) |
| 307 | mock_cuda_graph.assert_called_once() |
| 308 | |
| 309 | assert batch_descriptor in wrapper.concrete_cudagraph_entries |
| 310 | entry = wrapper.concrete_cudagraph_entries[batch_descriptor] |
| 311 | assert entry.cudagraph is not None |
| 312 | |
| 313 | # 2. Replay |
| 314 | with ( |
| 315 | set_forward_context( |
| 316 | attn_metadata=None, |
| 317 | vllm_config=self.vllm_config, |
| 318 | cudagraph_runtime_mode=CUDAGraphMode.FULL, |
| 319 | batch_descriptor=batch_descriptor, |
| 320 | ), |
| 321 | patch.object( |
| 322 | entry.cudagraph, "replay", wraps=entry.cudagraph.replay |
| 323 | ) as mock_replay, |
| 324 | ): |
| 325 | output2 = wrapper(self.input_tensor) |
| 326 | mock_replay.assert_called_once() |
| 327 | |
| 328 | # Compare with eager output |
| 329 | eager_output = self.model(self.input_tensor) |
| 330 | torch.testing.assert_close(eager_output, output2) |
| 331 | |
| 332 | def test_bypass_on_mode_mismatch(self): |
| 333 | wrapper = CUDAGraphWrapper( |
| 334 | self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL |
| 335 | ) |
| 336 | batch_descriptor = BatchDescriptor(num_tokens=10) |
| 337 | |
| 338 | with ( |
| 339 | set_forward_context( |
| 340 | attn_metadata=None, |
| 341 | vllm_config=self.vllm_config, |
| 342 | cudagraph_runtime_mode=CUDAGraphMode.PIECEWISE, |
| 343 | batch_descriptor=batch_descriptor, |
| 344 | ), |
| 345 | patch("torch.cuda.graph", wraps=torch.cuda.graph) as mock_cuda_graph, |
| 346 | patch.object( |
| 347 | self.model, "forward", wraps=self.model.forward |
| 348 | ) as mock_forward, |
| 349 | ): |
| 350 | wrapper(self.input_tensor) |
| 351 | mock_cuda_graph.assert_not_called() |
| 352 | mock_forward.assert_called_once() |
| 353 | assert not wrapper.concrete_cudagraph_entries |
| 354 | |
| 355 | def test_bypass_on_mode_none(self): |
| 356 | wrapper = CUDAGraphWrapper( |
| 357 | self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL |
| 358 | ) |
| 359 | batch_descriptor = BatchDescriptor(num_tokens=10) |
| 360 | |
| 361 | with ( |
| 362 | set_forward_context( |
| 363 | attn_metadata=None, |
| 364 | vllm_config=self.vllm_config, |
| 365 | cudagraph_runtime_mode=CUDAGraphMode.NONE, |
| 366 | batch_descriptor=batch_descriptor, |
| 367 | ), |
| 368 | patch("torch.cuda.graph", wraps=torch.cuda.graph) as mock_cuda_graph, |
| 369 | ): |
| 370 | wrapper(self.input_tensor) |
| 371 | mock_cuda_graph.assert_not_called() |
| 372 | assert not wrapper.concrete_cudagraph_entries |
| 373 | |
| 374 | |
| 375 | def _run_and_monitor_call( |
| 376 | wrapper, input_tensor, runtime_mode, batch_descriptor, vllm_config |
| 377 | ): |
| 378 | """Helper to run a single call and monitor the action.""" |
| 379 | with ( |
| 380 | patch("torch.cuda.graph", wraps=torch.cuda.graph) as mock_graph_context, |
| 381 | patch.object(wrapper, "runnable", wraps=wrapper.runnable) as mock_runnable, |
| 382 | ): |
| 383 | entry = wrapper.concrete_cudagraph_entries.get(batch_descriptor, None) |
| 384 | |
| 385 | context = set_forward_context( |
| 386 | attn_metadata=None, |
| 387 | vllm_config=vllm_config, |
| 388 | cudagraph_runtime_mode=runtime_mode, |
| 389 | batch_descriptor=batch_descriptor, |
| 390 | ) |
| 391 | mock_replay = MagicMock() |
| 392 | if entry and entry.cudagraph: |
| 393 | with ( |
| 394 | context, |
| 395 | patch.object( |
| 396 | entry.cudagraph, "replay", new_callable=MagicMock |
| 397 | ) as mock_replay, |
| 398 | ): |
| 399 | wrapper(input_tensor) |
| 400 | else: |
| 401 | with context: |
| 402 | wrapper(input_tensor) |
| 403 | |
| 404 | if mock_graph_context.called: |
| 405 | # note that this is globally mocked, so it will be detected |
| 406 | # even whether called by the inner or outer wrapper |
| 407 | return "capture_global" |
| 408 | if mock_replay.called: |
| 409 | # only for outer wrapper |
| 410 | return "replay" |
| 411 | if mock_runnable.call_count > 0: |
| 412 | # only for outer wrapper |
| 413 | return "bypass" |
| 414 | return "unknown" |
| 415 | |
| 416 | |
| 417 | @create_new_process_for_each_test("spawn") |
| 418 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="Skip if not cuda") |
| 419 | def test_capture_replay_bypass_logic(): |
| 420 | comp_config = CompilationConfig( |
| 421 | mode=CompilationMode.VLLM_COMPILE, |
| 422 | cudagraph_mode=CUDAGraphMode.FULL, |
| 423 | cudagraph_capture_sizes=[1, 2], |
| 424 | ) |
| 425 | vllm_config = _create_vllm_config(comp_config) |
| 426 | dispatcher = CudagraphDispatcher(vllm_config) |
| 427 | dispatcher.initialize_cudagraph_keys( |
| 428 | comp_config.cudagraph_mode, uniform_decode_query_len=1 |
| 429 | ) |
| 430 | model = SimpleMLP().to(DEVICE_TYPE) |
| 431 | full_wrapper = CUDAGraphWrapper(model, vllm_config, CUDAGraphMode.FULL) |
| 432 | max_bs = 16 |
| 433 | persistent_input_buffer = torch.zeros(max_bs, 10, device=DEVICE_TYPE) |
| 434 | input_1 = persistent_input_buffer[:1] |
| 435 | input_2 = persistent_input_buffer[:2] |
| 436 | input_3 = persistent_input_buffer[:3] |
| 437 | |
| 438 | desc_1 = BatchDescriptor(num_tokens=1) |
| 439 | desc_2 = BatchDescriptor(num_tokens=2) |
| 440 | desc_3_unseen = BatchDescriptor(num_tokens=3) |
| 441 | |
| 442 | # 0. global warmup |
| 443 | with set_forward_context( |
| 444 | attn_metadata=None, |
| 445 | vllm_config=vllm_config, |
| 446 | cudagraph_runtime_mode=CUDAGraphMode.NONE, |
| 447 | batch_descriptor=None, |
| 448 | ): |
| 449 | full_wrapper(input_1) |
| 450 | |
| 451 | rt_mode, key = dispatcher.dispatch(num_tokens=desc_1.num_tokens) |
| 452 | # 1. Capture first shape |
| 453 | action = _run_and_monitor_call(full_wrapper, input_1, rt_mode, key, vllm_config) |
| 454 | assert action == "capture_global" |
| 455 | |
| 456 | # 2. Replay first shape |
| 457 | action = _run_and_monitor_call(full_wrapper, input_1, rt_mode, key, vllm_config) |
| 458 | assert action == "replay" |
| 459 | |
| 460 | rt_mode, key = dispatcher.dispatch(num_tokens=desc_2.num_tokens) |
| 461 | # 3. Capture second shape |
| 462 | action = _run_and_monitor_call(full_wrapper, input_2, rt_mode, key, vllm_config) |
| 463 | assert action == "capture_global" |
| 464 | |
| 465 | # 4. Replay second shape |
| 466 | action = _run_and_monitor_call( |
| 467 | full_wrapper, input_2, CUDAGraphMode.FULL, key, vllm_config |
| 468 | ) |
| 469 | assert action == "replay" |
| 470 | |
| 471 | # 5. Bypass if no key match |
| 472 | rt_mode, key = dispatcher.dispatch(num_tokens=desc_3_unseen.num_tokens) |
| 473 | assert rt_mode == CUDAGraphMode.NONE |
| 474 | action = _run_and_monitor_call(full_wrapper, input_3, rt_mode, key, vllm_config) |
| 475 | assert action == "bypass" |
| 476 | |
| 477 | # capture unseen shape is not allowed after disable |
| 478 | set_cudagraph_capturing_enabled(False) |
| 479 | with pytest.raises(RuntimeError): |
| 480 | _run_and_monitor_call( |
| 481 | full_wrapper, input_3, CUDAGraphMode.FULL, desc_3_unseen, vllm_config |
| 482 | ) |
| 483 | set_cudagraph_capturing_enabled(True) |
| 484 | |
| 485 | |
| 486 | @create_new_process_for_each_test("spawn") |
| 487 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="Skip if not cuda") |
| 488 | def test_nested_wrappers(): |
| 489 | """Tests a scenario with a PIECEWISE wrapper inside a FULL one.""" |
| 490 | comp_config = CompilationConfig( |
| 491 | mode=CompilationMode.VLLM_COMPILE, |
| 492 | cudagraph_mode=CUDAGraphMode.FULL, |
| 493 | cudagraph_capture_sizes=[1], |
| 494 | ) |
| 495 | vllm_config = _create_vllm_config(comp_config) |
| 496 | dispatcher = CudagraphDispatcher(vllm_config) |
| 497 | dispatcher.initialize_cudagraph_keys( |
| 498 | comp_config.cudagraph_mode, uniform_decode_query_len=1 |
| 499 | ) |
| 500 | model = SimpleMLP().to(DEVICE_TYPE) |
| 501 | full_wrapper = CUDAGraphWrapper(model, vllm_config, CUDAGraphMode.FULL) |
| 502 | input_1 = torch.randn(1, 10, device=DEVICE_TYPE) |
| 503 | |
| 504 | # Setup: Inner model is wrapped with PIECEWISE, outer with FULL |
| 505 | inner_model = SimpleMLP().to(DEVICE_TYPE) |
| 506 | piecewise_wrapper = CUDAGraphWrapper( |
| 507 | inner_model, vllm_config, CUDAGraphMode.PIECEWISE |
| 508 | ) |
| 509 | inner_model.forward = MagicMock(wraps=inner_model.forward) |
| 510 | outer_model = SimpleMLP().to(DEVICE_TYPE) |
| 511 | # When outer model is called, it calls the piecewise_wrapper |
| 512 | outer_model.forward = MagicMock( |
| 513 | wraps=outer_model.forward, side_effect=piecewise_wrapper |
| 514 | ) |
| 515 | full_wrapper = CUDAGraphWrapper(outer_model, vllm_config, CUDAGraphMode.FULL) |
| 516 | |
| 517 | desc_1 = BatchDescriptor(num_tokens=1) |
| 518 | |
| 519 | # 0. global warmup |
| 520 | with set_forward_context( |
| 521 | attn_metadata=None, |
| 522 | vllm_config=vllm_config, |
| 523 | cudagraph_runtime_mode=CUDAGraphMode.NONE, |
| 524 | batch_descriptor=None, |
| 525 | ): |
| 526 | full_wrapper(input_1) |
| 527 | |
| 528 | # --- Test runtime mode FULL--- |
| 529 | # Run with FULL mode context. Expect outer wrapper to capture. |
| 530 | # The inner mock should be called once inside the graph capture. |
| 531 | outer_model.forward.reset_mock() |
| 532 | inner_model.forward.reset_mock() |
| 533 | action = _run_and_monitor_call( |
| 534 | full_wrapper, input_1, CUDAGraphMode.FULL, desc_1, vllm_config |
| 535 | ) |
| 536 | assert action == "capture_global" |
| 537 | assert outer_model.forward.call_count == 1 |
| 538 | assert inner_model.forward.call_count == 1 |
| 539 | |
| 540 | # Run again. Expect outer wrapper to replay. |
| 541 | # The outer model should NOT be called because the whole graph |
| 542 | # is replayed. |
| 543 | action = _run_and_monitor_call( |
| 544 | full_wrapper, input_1, CUDAGraphMode.FULL, desc_1, vllm_config |
| 545 | ) |
| 546 | assert action == "replay" |
| 547 | assert outer_model.forward.call_count == 1 # No new call |
| 548 | assert inner_model.forward.call_count == 1 |
| 549 | |
| 550 | # --- Test runtime mode PIECEWISE --- |
| 551 | outer_model.forward.reset_mock() |
| 552 | inner_model.forward.reset_mock() |
| 553 | # Run with PIECEWISE mode context. |
| 554 | # Expect outer wrapper to bypass and call inner wrapper. |
| 555 | # Inner wrapper should capture. |
| 556 | action = _run_and_monitor_call( |
| 557 | full_wrapper, input_1, CUDAGraphMode.PIECEWISE, desc_1, vllm_config |
| 558 | ) |
| 559 | assert action == "capture_global" |
| 560 | assert outer_model.forward.call_count == 1 |
| 561 | assert inner_model.forward.call_count == 1 |
| 562 | |
| 563 | # Run again with PIECEWISE. |
| 564 | # Outer bypasses, inner replays. |
| 565 | action = _run_and_monitor_call( |
| 566 | full_wrapper, input_1, CUDAGraphMode.PIECEWISE, desc_1, vllm_config |
| 567 | ) |
| 568 | assert action == "bypass" |
| 569 | assert outer_model.forward.call_count == 2 |
| 570 | assert inner_model.forward.call_count == 1 |