| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for Helion kernel registration. |
| 4 | |
| 5 | Tests ConfiguredHelionKernel, HelionKernelWrapper, and PresetConfigSearch |
| 6 | including config picker registration and custom autotuner integration. |
| 7 | """ |
| 8 | |
| 9 | from unittest.mock import Mock, patch |
| 10 | |
| 11 | import pytest |
| 12 | import torch |
| 13 | |
| 14 | from vllm.utils.import_utils import has_helion |
| 15 | |
| 16 | if not has_helion(): |
| 17 | pytest.skip( |
| 18 | "Helion is not installed. Install with: pip install vllm[helion]", |
| 19 | allow_module_level=True, |
| 20 | ) |
| 21 | |
| 22 | import helion |
| 23 | import helion.language as hl |
| 24 | |
| 25 | from tests.kernels.helion.helpers import dummy_kernel_registry |
| 26 | from vllm.kernels.helion.case_key import CaseKey |
| 27 | from vllm.kernels.helion.config_manager import ConfigManager |
| 28 | from vllm.kernels.helion.register import ( |
| 29 | _HOP_AVAILABLE, |
| 30 | ConfiguredHelionKernel, |
| 31 | HelionKernelWrapper, |
| 32 | get_kernel_by_name, |
| 33 | get_registered_kernels, |
| 34 | register_kernel, |
| 35 | validate_helion_settings, |
| 36 | ) |
| 37 | |
| 38 | if _HOP_AVAILABLE: |
| 39 | from helion._compat import supports_torch_compile_fusion |
| 40 | from helion._compiler._dynamo.higher_order_ops import ( |
| 41 | helion_kernel_wrapper_mutation, |
| 42 | ) |
| 43 | from torch._inductor.utils import run_and_get_code |
| 44 | |
| 45 | |
| 46 | def _add_kernel(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor: |
| 47 | out = torch.empty_like(x) |
| 48 | for tile in hl.tile(x.size()): |
| 49 | out[tile] = x[tile] + y[tile] |
| 50 | return out |
| 51 | |
| 52 | |
| 53 | @pytest.fixture |
| 54 | def sample_configs(): |
| 55 | """Create real Helion config objects for testing.""" |
| 56 | return { |
| 57 | CaseKey({"batchsize": 32, "hiddensize": 4096}): helion.Config( |
| 58 | block_sizes=[128], |
| 59 | num_warps=4, |
| 60 | num_stages=3, |
| 61 | ), |
| 62 | CaseKey({"batchsize": 64, "hiddensize": 4096}): helion.Config( |
| 63 | block_sizes=[256], |
| 64 | num_warps=8, |
| 65 | num_stages=4, |
| 66 | ), |
| 67 | CaseKey({"batchsize": 128, "hiddensize": 4096}): helion.Config( |
| 68 | block_sizes=[512], |
| 69 | num_warps=16, |
| 70 | num_stages=2, |
| 71 | ), |
| 72 | CaseKey.default(): helion.Config( |
| 73 | block_sizes=[64], |
| 74 | num_warps=2, |
| 75 | num_stages=2, |
| 76 | ), |
| 77 | } |
| 78 | |
| 79 | |
| 80 | @pytest.fixture |
| 81 | def sample_kernel(): |
| 82 | """Create a simple test kernel function.""" |
| 83 | |
| 84 | def test_kernel(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor: |
| 85 | """Simple test kernel that adds two tensors.""" |
| 86 | return x + y |
| 87 | |
| 88 | return test_kernel |
| 89 | |
| 90 | |
| 91 | @pytest.fixture |
| 92 | def config_manager_with_test_configs(sample_configs): |
| 93 | """Set up ConfigManager with test configs for nvidia_h200 platform.""" |
| 94 | mock_config_manager = Mock(spec=ConfigManager) |
| 95 | mock_config_manager.get_platform_configs = Mock(return_value=sample_configs) |
| 96 | return mock_config_manager |
| 97 | |
| 98 | |
| 99 | @pytest.fixture |
| 100 | def configured_kernel(sample_kernel, sample_configs, config_manager_with_test_configs): |
| 101 | """Create a ConfiguredHelionKernel for testing.""" |
| 102 | |
| 103 | def test_config_picker(args, config_keys): |
| 104 | return None |
| 105 | |
| 106 | with ( |
| 107 | patch( |
| 108 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 109 | return_value=config_manager_with_test_configs, |
| 110 | ), |
| 111 | patch( |
| 112 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 113 | return_value="nvidia_h200", |
| 114 | ), |
| 115 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 116 | ): |
| 117 | mock_decorated = Mock() |
| 118 | mock_kernel.return_value = Mock(return_value=mock_decorated) |
| 119 | |
| 120 | return ConfiguredHelionKernel( |
| 121 | op_name="test_kernel", |
| 122 | config_picker=test_config_picker, |
| 123 | raw_kernel_func=sample_kernel, |
| 124 | helion_settings=None, |
| 125 | ) |
| 126 | |
| 127 | |
| 128 | class TestValidateHelionSettings: |
| 129 | """Test suite for validate_helion_settings utility function.""" |
| 130 | |
| 131 | def test_accepts_none_settings(self): |
| 132 | """Test that None settings are accepted without error.""" |
| 133 | validate_helion_settings(None, "test_kernel") # Should not raise |
| 134 | |
| 135 | def test_accepts_valid_settings(self): |
| 136 | """Test that valid settings without conflicts are accepted.""" |
| 137 | settings = helion.Settings() |
| 138 | settings.static_shapes = False |
| 139 | settings.print_output_code = True |
| 140 | validate_helion_settings(settings, "test_kernel") # Should not raise |
| 141 | |
| 142 | def test_rejects_autotuner_fn(self): |
| 143 | """Test that settings with custom autotuner_fn raise ValueError.""" |
| 144 | settings = helion.Settings() |
| 145 | settings.autotuner_fn = lambda *args: None # Set custom autotuner function |
| 146 | |
| 147 | with pytest.raises(ValueError, match="uses a custom autotuner"): |
| 148 | validate_helion_settings(settings, "test_kernel") |
| 149 | |
| 150 | def test_warns_on_static_shapes_true(self): |
| 151 | """Test that static_shapes=True emits a warning about being overridden.""" |
| 152 | settings = helion.Settings() |
| 153 | settings.static_shapes = True |
| 154 | |
| 155 | with patch("vllm.kernels.helion.register.logger") as mock_logger: |
| 156 | validate_helion_settings(settings, "test_kernel") |
| 157 | mock_logger.warning.assert_called_once() |
| 158 | assert "overridden to False" in mock_logger.warning.call_args[0][0] |
| 159 | |
| 160 | |
| 161 | def create_configured_kernel_with_configs( |
| 162 | op_name, |
| 163 | config_picker, |
| 164 | kernel_func, |
| 165 | configs, |
| 166 | platform="nvidia_h200", |
| 167 | helion_settings=None, |
| 168 | ): |
| 169 | """Helper to create ConfiguredHelionKernel with real config objects.""" |
| 170 | mock_config_manager = Mock(spec=ConfigManager) |
| 171 | mock_config_manager.get_platform_configs = Mock(return_value=configs) |
| 172 | |
| 173 | with ( |
| 174 | patch( |
| 175 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 176 | return_value=mock_config_manager, |
| 177 | ), |
| 178 | patch( |
| 179 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 180 | return_value=platform, |
| 181 | ), |
| 182 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 183 | ): |
| 184 | mock_decorated = Mock() |
| 185 | mock_kernel.return_value = Mock(return_value=mock_decorated) |
| 186 | |
| 187 | return ConfiguredHelionKernel( |
| 188 | op_name=op_name, |
| 189 | config_picker=config_picker, |
| 190 | raw_kernel_func=kernel_func, |
| 191 | helion_settings=helion_settings, |
| 192 | ) |
| 193 | |
| 194 | |
| 195 | class TestConfiguredHelionKernel: |
| 196 | """Test suite for ConfiguredHelionKernel.""" |
| 197 | |
| 198 | def test_init_raises_without_picker(self, sample_kernel, sample_configs): |
| 199 | """Test that __init__ raises when no picker registered.""" |
| 200 | configs: dict[CaseKey, helion.Config] = { |
| 201 | CaseKey.default(): sample_configs[CaseKey.default()] |
| 202 | } |
| 203 | mock_config_manager = Mock(spec=ConfigManager) |
| 204 | mock_config_manager.get_platform_configs = Mock(return_value=configs) |
| 205 | |
| 206 | with ( |
| 207 | patch( |
| 208 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 209 | return_value=mock_config_manager, |
| 210 | ), |
| 211 | patch( |
| 212 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 213 | return_value="nvidia_h200", |
| 214 | ), |
| 215 | pytest.raises(RuntimeError, match="No config picker registered"), |
| 216 | ): |
| 217 | ConfiguredHelionKernel( |
| 218 | op_name="test_kernel", |
| 219 | config_picker=None, # No picker registered |
| 220 | raw_kernel_func=sample_kernel, |
| 221 | helion_settings=None, |
| 222 | ) |
| 223 | |
| 224 | def test_config_selector_validates_picker_result( |
| 225 | self, sample_kernel, sample_configs |
| 226 | ): |
| 227 | """Test that config selector validates picker returns valid key.""" |
| 228 | |
| 229 | def invalid_picker(args, config_keys): |
| 230 | return {"invalid": 999} |
| 231 | |
| 232 | kernel = create_configured_kernel_with_configs( |
| 233 | op_name="test_kernel", |
| 234 | config_picker=invalid_picker, |
| 235 | kernel_func=sample_kernel, |
| 236 | configs=sample_configs, |
| 237 | ) |
| 238 | |
| 239 | key_computer = kernel._create_key_computer() |
| 240 | selector = kernel._create_config_selector(key_computer) |
| 241 | |
| 242 | with pytest.raises( |
| 243 | ValueError, match="Config picker returned invalid config key" |
| 244 | ): |
| 245 | selector((torch.randn(32, 4096),)) |
| 246 | |
| 247 | def test_config_selector_handles_none_from_picker( |
| 248 | self, sample_kernel, sample_configs |
| 249 | ): |
| 250 | """Test that config selector falls back to 'default' on None.""" |
| 251 | |
| 252 | def none_picker(args, config_keys): |
| 253 | return None |
| 254 | |
| 255 | kernel = create_configured_kernel_with_configs( |
| 256 | op_name="test_kernel", |
| 257 | config_picker=none_picker, |
| 258 | kernel_func=sample_kernel, |
| 259 | configs=sample_configs, |
| 260 | ) |
| 261 | |
| 262 | key_computer = kernel._create_key_computer() |
| 263 | selector = kernel._create_config_selector(key_computer) |
| 264 | |
| 265 | result = selector((torch.randn(32, 4096),)) |
| 266 | assert result is kernel.configs[CaseKey.default()] |
| 267 | |
| 268 | def test_create_decorated_kernel_passes_helion_settings( |
| 269 | self, sample_kernel, sample_configs |
| 270 | ): |
| 271 | """Test that _create_decorated_kernel passes helion_settings.""" |
| 272 | |
| 273 | def default_picker(args, config_keys): |
| 274 | return None |
| 275 | |
| 276 | settings = helion.Settings() |
| 277 | settings.print_output_code = True |
| 278 | |
| 279 | mock_config_manager = Mock(spec=ConfigManager) |
| 280 | mock_config_manager.get_platform_configs = Mock(return_value=sample_configs) |
| 281 | |
| 282 | with ( |
| 283 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 284 | patch( |
| 285 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 286 | return_value=mock_config_manager, |
| 287 | ), |
| 288 | patch( |
| 289 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 290 | return_value="nvidia_h200", |
| 291 | ), |
| 292 | ): |
| 293 | mock_decorated = Mock() |
| 294 | mock_kernel.return_value = Mock(return_value=mock_decorated) |
| 295 | |
| 296 | ConfiguredHelionKernel( |
| 297 | op_name="test_kernel", |
| 298 | config_picker=default_picker, |
| 299 | raw_kernel_func=sample_kernel, |
| 300 | helion_settings=settings, |
| 301 | ) |
| 302 | |
| 303 | call_kwargs = mock_kernel.call_args[1] |
| 304 | assert "print_output_code" in call_kwargs |
| 305 | assert call_kwargs["print_output_code"] is True |
| 306 | # static_shapes is always forced to False by vLLM |
| 307 | assert call_kwargs["static_shapes"] is False |
| 308 | |
| 309 | def test_key_and_config_selector_use_same_logic( |
| 310 | self, sample_kernel, sample_configs |
| 311 | ): |
| 312 | """Test that key and config_selector produce identical results.""" |
| 313 | |
| 314 | def tracking_picker(args, config_keys): |
| 315 | x = args[0] |
| 316 | batch_size = x.shape[0] |
| 317 | if batch_size <= 32: |
| 318 | return CaseKey({"batchsize": 32, "hiddensize": 4096}) |
| 319 | elif batch_size <= 64: |
| 320 | return CaseKey({"batchsize": 64, "hiddensize": 4096}) |
| 321 | return CaseKey({"batchsize": 128, "hiddensize": 4096}) |
| 322 | |
| 323 | mock_config_manager = Mock(spec=ConfigManager) |
| 324 | mock_config_manager.get_platform_configs = Mock(return_value=sample_configs) |
| 325 | |
| 326 | with ( |
| 327 | patch("vllm.kernels.helion.register.helion.kernel") as mock_helion_kernel, |
| 328 | patch( |
| 329 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 330 | return_value=mock_config_manager, |
| 331 | ), |
| 332 | patch( |
| 333 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 334 | return_value="nvidia_h200", |
| 335 | ), |
| 336 | ): |
| 337 | mock_decorated = Mock() |
| 338 | mock_helion_kernel.return_value = Mock(return_value=mock_decorated) |
| 339 | |
| 340 | kernel = ConfiguredHelionKernel( |
| 341 | op_name="test_kernel", |
| 342 | config_picker=tracking_picker, |
| 343 | raw_kernel_func=sample_kernel, |
| 344 | helion_settings=None, |
| 345 | ) |
| 346 | |
| 347 | call_kwargs = mock_helion_kernel.call_args[1] |
| 348 | key_fn = call_kwargs["key"] |
| 349 | autotuner_fn = call_kwargs["autotuner_fn"] |
| 350 | |
| 351 | tensor = torch.randn(50, 4096) # batch=50, should select batchsize_64 |
| 352 | |
| 353 | key_result = key_fn(tensor) |
| 354 | autotuner = autotuner_fn(None, (tensor,)) |
| 355 | config = autotuner.autotune() |
| 356 | |
| 357 | expected_key = CaseKey({"batchsize": 64, "hiddensize": 4096}) |
| 358 | assert key_result == str(expected_key) |
| 359 | assert config is kernel.configs[expected_key] |
| 360 | |
| 361 | |
| 362 | class TestHelionKernelWrapper: |
| 363 | """Test suite for HelionKernelWrapper.""" |
| 364 | |
| 365 | def test_init_disables_on_missing_configs(self, sample_kernel): |
| 366 | """Test __init__ marks wrapper as disabled when configs are missing.""" |
| 367 | |
| 368 | def fake_impl(*args, **kwargs): |
| 369 | return torch.zeros_like(args[0]) |
| 370 | |
| 371 | def default_picker(args, config_keys): |
| 372 | return None |
| 373 | |
| 374 | mock_config_manager = Mock(spec=ConfigManager) |
| 375 | mock_config_manager.get_platform_configs = Mock( |
| 376 | return_value={} |
| 377 | ) # Empty configs |
| 378 | |
| 379 | with ( |
| 380 | patch( |
| 381 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 382 | return_value=mock_config_manager, |
| 383 | ), |
| 384 | patch( |
| 385 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 386 | return_value="nvidia_h200", |
| 387 | ), |
| 388 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 389 | ): |
| 390 | mock_kernel.return_value = Mock(return_value=sample_kernel) |
| 391 | |
| 392 | wrapper = HelionKernelWrapper( |
| 393 | raw_kernel_func=sample_kernel, |
| 394 | op_name="test_kernel", |
| 395 | fake_impl=fake_impl, |
| 396 | config_picker=default_picker, |
| 397 | ) |
| 398 | |
| 399 | assert wrapper._disabled is True |
| 400 | assert wrapper._disabled_reason is not None |
| 401 | assert "No configs available" in wrapper._disabled_reason |
| 402 | |
| 403 | def test_disabled_wrapper_raises_on_call(self, sample_kernel): |
| 404 | """Test __call__ raises RuntimeError on a disabled wrapper.""" |
| 405 | |
| 406 | def fake_impl(*args, **kwargs): |
| 407 | return torch.zeros_like(args[0]) |
| 408 | |
| 409 | def default_picker(args, config_keys): |
| 410 | return None |
| 411 | |
| 412 | mock_config_manager = Mock(spec=ConfigManager) |
| 413 | mock_config_manager.get_platform_configs = Mock(return_value={}) |
| 414 | |
| 415 | with ( |
| 416 | patch( |
| 417 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 418 | return_value=mock_config_manager, |
| 419 | ), |
| 420 | patch( |
| 421 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 422 | return_value="nvidia_h200", |
| 423 | ), |
| 424 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 425 | ): |
| 426 | mock_kernel.return_value = Mock(return_value=sample_kernel) |
| 427 | |
| 428 | wrapper = HelionKernelWrapper( |
| 429 | raw_kernel_func=sample_kernel, |
| 430 | op_name="test_kernel", |
| 431 | fake_impl=fake_impl, |
| 432 | config_picker=default_picker, |
| 433 | ) |
| 434 | |
| 435 | with pytest.raises(RuntimeError, match="is disabled"): |
| 436 | wrapper(torch.randn(4, 4), torch.randn(4, 4)) |
| 437 | |
| 438 | def test_disabled_wrapper_get_configured_op_raises(self, sample_kernel): |
| 439 | """Test get_configured_op raises RuntimeError on a disabled wrapper.""" |
| 440 | |
| 441 | def fake_impl(*args, **kwargs): |
| 442 | return torch.zeros_like(args[0]) |
| 443 | |
| 444 | def default_picker(args, config_keys): |
| 445 | return None |
| 446 | |
| 447 | mock_config_manager = Mock(spec=ConfigManager) |
| 448 | mock_config_manager.get_platform_configs = Mock(return_value={}) |
| 449 | |
| 450 | with ( |
| 451 | patch( |
| 452 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 453 | return_value=mock_config_manager, |
| 454 | ), |
| 455 | patch( |
| 456 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 457 | return_value="nvidia_h200", |
| 458 | ), |
| 459 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 460 | ): |
| 461 | mock_kernel.return_value = Mock(return_value=sample_kernel) |
| 462 | |
| 463 | wrapper = HelionKernelWrapper( |
| 464 | raw_kernel_func=sample_kernel, |
| 465 | op_name="test_kernel", |
| 466 | fake_impl=fake_impl, |
| 467 | config_picker=default_picker, |
| 468 | ) |
| 469 | |
| 470 | with pytest.raises(RuntimeError, match="is disabled"): |
| 471 | wrapper.get_configured_op() |
| 472 | |
| 473 | def test_disabled_wrapper_supports_get_inputs(self, sample_kernel): |
| 474 | """Test get_inputs works on a disabled wrapper.""" |
| 475 | |
| 476 | def fake_impl(*args, **kwargs): |
| 477 | return torch.zeros_like(args[0]) |
| 478 | |
| 479 | def default_picker(args, config_keys): |
| 480 | return None |
| 481 | |
| 482 | expected_inputs = {"key1": (torch.randn(4),)} |
| 483 | input_gen = Mock(return_value=expected_inputs) |
| 484 | |
| 485 | mock_config_manager = Mock(spec=ConfigManager) |
| 486 | mock_config_manager.get_platform_configs = Mock(return_value={}) |
| 487 | |
| 488 | with ( |
| 489 | patch( |
| 490 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 491 | return_value=mock_config_manager, |
| 492 | ), |
| 493 | patch( |
| 494 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 495 | return_value="nvidia_h200", |
| 496 | ), |
| 497 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 498 | ): |
| 499 | mock_kernel.return_value = Mock(return_value=sample_kernel) |
| 500 | |
| 501 | wrapper = HelionKernelWrapper( |
| 502 | raw_kernel_func=sample_kernel, |
| 503 | op_name="test_kernel", |
| 504 | fake_impl=fake_impl, |
| 505 | config_picker=default_picker, |
| 506 | input_generator=input_gen, |
| 507 | ) |
| 508 | |
| 509 | assert wrapper._disabled is True |
| 510 | result = wrapper.get_inputs() |
| 511 | assert result is expected_inputs |
| 512 | |
| 513 | def test_disabled_wrapper_supports_run_autotune(self, sample_kernel): |
| 514 | """Test run_autotune works on a disabled wrapper.""" |
| 515 | |
| 516 | def fake_impl(*args, **kwargs): |
| 517 | return torch.zeros_like(args[0]) |
| 518 | |
| 519 | def default_picker(args, config_keys): |
| 520 | return None |
| 521 | |
| 522 | mock_config_manager = Mock(spec=ConfigManager) |
| 523 | mock_config_manager.get_platform_configs = Mock(return_value={}) |
| 524 | |
| 525 | mock_config = Mock() |
| 526 | |
| 527 | with ( |
| 528 | patch( |
| 529 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 530 | return_value=mock_config_manager, |
| 531 | ), |
| 532 | patch( |
| 533 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 534 | return_value="nvidia_h200", |
| 535 | ), |
| 536 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 537 | ): |
| 538 | mock_kernel.return_value = Mock(return_value=sample_kernel) |
| 539 | |
| 540 | wrapper = HelionKernelWrapper( |
| 541 | raw_kernel_func=sample_kernel, |
| 542 | op_name="test_kernel", |
| 543 | fake_impl=fake_impl, |
| 544 | config_picker=default_picker, |
| 545 | ) |
| 546 | |
| 547 | assert wrapper._disabled is True |
| 548 | |
| 549 | with patch( |
| 550 | "vllm.kernels.helion.register.create_helion_decorated_kernel" |
| 551 | ) as mock_create: |
| 552 | mock_autotune_kernel = Mock() |
| 553 | mock_autotune_kernel.autotune.return_value = mock_config |
| 554 | mock_create.return_value = mock_autotune_kernel |
| 555 | |
| 556 | inputs = (torch.randn(4, 4),) |
| 557 | result = wrapper.run_autotune(inputs) |
| 558 | assert result is mock_config |
| 559 | |
| 560 | def test_init_caches_configured_kernel(self, sample_kernel, sample_configs): |
| 561 | """Test __init__ eagerly builds and caches ConfiguredHelionKernel.""" |
| 562 | |
| 563 | def fake_impl(*args, **kwargs): |
| 564 | return torch.zeros_like(args[0]) |
| 565 | |
| 566 | def default_picker(args, config_keys): |
| 567 | return None |
| 568 | |
| 569 | mock_config_manager = Mock(spec=ConfigManager) |
| 570 | mock_config_manager.get_platform_configs = Mock(return_value=sample_configs) |
| 571 | |
| 572 | with ( |
| 573 | patch( |
| 574 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 575 | return_value=mock_config_manager, |
| 576 | ), |
| 577 | patch( |
| 578 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 579 | return_value="nvidia_h200", |
| 580 | ), |
| 581 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 582 | ): |
| 583 | mock_kernel.return_value = Mock(return_value=sample_kernel) |
| 584 | |
| 585 | wrapper = HelionKernelWrapper( |
| 586 | raw_kernel_func=sample_kernel, |
| 587 | op_name="test_kernel", |
| 588 | fake_impl=fake_impl, |
| 589 | config_picker=default_picker, |
| 590 | ) |
| 591 | |
| 592 | assert wrapper._configured_kernel is not None |
| 593 | result1 = wrapper.get_configured_op() |
| 594 | result2 = wrapper.get_configured_op() |
| 595 | assert result1 is result2 |
| 596 | |
| 597 | @pytest.mark.skipif( |
| 598 | not _HOP_AVAILABLE, reason="HOP path only used when HOP available" |
| 599 | ) |
| 600 | def test_init_eagerly_initializes_hop_path(self): |
| 601 | """Test that register_kernel eagerly builds the configured kernel |
| 602 | on the HOP path (no custom op registration needed).""" |
| 603 | from vllm.kernels.helion.utils import get_canonical_gpu_name |
| 604 | |
| 605 | configs: dict[CaseKey, helion.Config] = { |
| 606 | CaseKey.default(): helion.Config(block_sizes=[4, 4]) |
| 607 | } |
| 608 | with ( |
| 609 | dummy_kernel_registry(configs=configs) as register, |
| 610 | patch( |
| 611 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 612 | wraps=get_canonical_gpu_name, |
| 613 | ) as mock_gpu, |
| 614 | ): |
| 615 | wrapper = register( |
| 616 | config_picker=lambda args, keys: None, |
| 617 | )(_add_kernel) |
| 618 | |
| 619 | mock_gpu.assert_called_once() |
| 620 | assert wrapper._configured_kernel is not None |
| 621 | |
| 622 | with patch( |
| 623 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 624 | side_effect=AssertionError("get_canonical_gpu_name called during __call__"), |
| 625 | ): |
| 626 | x = torch.randn(4, 4, device="cuda") |
| 627 | y = torch.randn(4, 4, device="cuda") |
| 628 | result = wrapper(x, y) |
| 629 | expected = x + y |
| 630 | assert torch.allclose(result, expected) |
| 631 | |
| 632 | @pytest.mark.skipif( |
| 633 | _HOP_AVAILABLE, reason="CustomOp path not used when HOP available" |
| 634 | ) |
| 635 | def test_init_eagerly_initializes(self): |
| 636 | """Test that register_kernel eagerly loads configs and detects GPU |
| 637 | during construction so __call__ needs no further initialization.""" |
| 638 | from vllm.kernels.helion.utils import get_canonical_gpu_name |
| 639 | |
| 640 | with ( |
| 641 | dummy_kernel_registry() as register, |
| 642 | patch( |
| 643 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 644 | wraps=get_canonical_gpu_name, |
| 645 | ) as mock_gpu, |
| 646 | ): |
| 647 | wrapper = register( |
| 648 | config_picker=lambda args, keys: None, |
| 649 | )(_add_kernel) |
| 650 | |
| 651 | # Init must have detected GPU and built the kernel |
| 652 | mock_gpu.assert_called_once() |
| 653 | assert wrapper._configured_kernel is not None |
| 654 | assert hasattr(torch.ops.vllm_helion, wrapper.op_name) |
| 655 | |
| 656 | @pytest.mark.skipif( |
| 657 | _HOP_AVAILABLE, reason="CustomOp path not used when HOP available" |
| 658 | ) |
| 659 | def test_get_or_register_custom_op_returns_cached_op( |
| 660 | self, sample_kernel, sample_configs |
| 661 | ): |
| 662 | def fake_impl(*args, **kwargs): |
| 663 | return torch.zeros_like(args[0]) |
| 664 | |
| 665 | def default_picker(args, config_keys): |
| 666 | return None |
| 667 | |
| 668 | mock_config_manager = Mock(spec=ConfigManager) |
| 669 | mock_config_manager.get_platform_configs = Mock(return_value=sample_configs) |
| 670 | |
| 671 | existing_op = Mock() |
| 672 | mock_namespace = Mock() |
| 673 | mock_namespace.test_kernel = existing_op |
| 674 | |
| 675 | with ( |
| 676 | patch( |
| 677 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 678 | return_value=mock_config_manager, |
| 679 | ), |
| 680 | patch( |
| 681 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 682 | return_value="nvidia_h200", |
| 683 | ), |
| 684 | patch.object(torch.ops, "vllm_helion", mock_namespace), |
| 685 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 686 | ): |
| 687 | mock_decorated = Mock() |
| 688 | mock_kernel.return_value = Mock(return_value=mock_decorated) |
| 689 | |
| 690 | wrapper = HelionKernelWrapper( |
| 691 | raw_kernel_func=sample_kernel, |
| 692 | op_name="test_kernel", |
| 693 | fake_impl=fake_impl, |
| 694 | config_picker=default_picker, |
| 695 | ) |
| 696 | result = wrapper._get_or_register_custom_op() |
| 697 | assert result is existing_op |
| 698 | |
| 699 | @pytest.mark.skipif( |
| 700 | _HOP_AVAILABLE, reason="CustomOp path not used when HOP available" |
| 701 | ) |
| 702 | def test_get_or_register_custom_op_registers_new_op( |
| 703 | self, sample_kernel, sample_configs |
| 704 | ): |
| 705 | def fake_impl(*args, **kwargs): |
| 706 | return torch.zeros_like(args[0]) |
| 707 | |
| 708 | def default_picker(args, config_keys): |
| 709 | return None |
| 710 | |
| 711 | mock_config_manager = Mock(spec=ConfigManager) |
| 712 | mock_config_manager.get_platform_configs = Mock(return_value=sample_configs) |
| 713 | |
| 714 | new_op = Mock() |
| 715 | registered_ops: dict[str, Mock] = {} |
| 716 | mutates_args = ["y"] |
| 717 | |
| 718 | class MockNamespace: |
| 719 | def __getattr__(self, name): |
| 720 | if name in registered_ops: |
| 721 | return registered_ops[name] |
| 722 | raise AttributeError(name) |
| 723 | |
| 724 | mock_namespace = MockNamespace() |
| 725 | |
| 726 | def register_side_effect(op_name, op_func, **kwargs): |
| 727 | registered_ops[op_name] = new_op |
| 728 | |
| 729 | with ( |
| 730 | patch( |
| 731 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 732 | return_value=mock_config_manager, |
| 733 | ), |
| 734 | patch( |
| 735 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 736 | return_value="nvidia_h200", |
| 737 | ), |
| 738 | patch.object(torch.ops, "vllm_helion", mock_namespace), |
| 739 | patch( |
| 740 | "vllm.kernels.helion.register.direct_register_custom_op", |
| 741 | side_effect=register_side_effect, |
| 742 | ) as mock_register, |
| 743 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 744 | ): |
| 745 | mock_decorated = Mock() |
| 746 | mock_kernel.return_value = Mock(return_value=mock_decorated) |
| 747 | |
| 748 | wrapper = HelionKernelWrapper( |
| 749 | raw_kernel_func=sample_kernel, |
| 750 | op_name="test_kernel", |
| 751 | fake_impl=fake_impl, |
| 752 | mutates_args=mutates_args, |
| 753 | config_picker=default_picker, |
| 754 | ) |
| 755 | result = wrapper._get_or_register_custom_op() |
| 756 | |
| 757 | mock_register.assert_called_once() |
| 758 | assert result is new_op |
| 759 | assert mock_register.call_args[1]["op_func"] is mock_decorated |
| 760 | assert mock_register.call_args[1]["mutates_args"] is mutates_args |
| 761 | |
| 762 | |
| 763 | class TestKernelRegistry: |
| 764 | """Test suite for kernel registry functionality.""" |
| 765 | |
| 766 | def setup_method(self): |
| 767 | """Save and clear the registry before each test.""" |
| 768 | from vllm.kernels.helion.register import _REGISTERED_KERNELS |
| 769 | |
| 770 | self._saved_registry = dict(_REGISTERED_KERNELS) |
| 771 | _REGISTERED_KERNELS.clear() |
| 772 | |
| 773 | def teardown_method(self): |
| 774 | """Restore the registry after each test.""" |
| 775 | from vllm.kernels.helion.register import _REGISTERED_KERNELS |
| 776 | |
| 777 | _REGISTERED_KERNELS.clear() |
| 778 | _REGISTERED_KERNELS.update(self._saved_registry) |
| 779 | |
| 780 | def test_get_registered_kernels_returns_copy(self): |
| 781 | """Test get_registered_kernels returns copy of registry.""" |
| 782 | result1 = get_registered_kernels() |
| 783 | result2 = get_registered_kernels() |
| 784 | |
| 785 | # Should be separate objects |
| 786 | assert result1 is not result2 |
| 787 | # Should have same content |
| 788 | assert result1 == result2 |
| 789 | |
| 790 | def test_get_kernel_by_name_returns_kernel(self): |
| 791 | """Test get_kernel_by_name returns registered kernel.""" |
| 792 | with dummy_kernel_registry() as register: |
| 793 | wrapper = register("test_kernel", config_picker=lambda args, keys: None)( |
| 794 | _add_kernel |
| 795 | ) |
| 796 | |
| 797 | from vllm.kernels.helion.register import _REGISTERED_KERNELS |
| 798 | |
| 799 | _REGISTERED_KERNELS["test_kernel"] = wrapper |
| 800 | |
| 801 | result = get_kernel_by_name("test_kernel") |
| 802 | assert result is wrapper |
| 803 | |
| 804 | def test_get_kernel_by_name_returns_none_for_missing(self): |
| 805 | """Test get_kernel_by_name returns None for missing kernel.""" |
| 806 | result = get_kernel_by_name("nonexistent") |
| 807 | assert result is None |
| 808 | |
| 809 | def test_register_kernel_auto_generates_fake_impl(self): |
| 810 | """Test register_kernel auto-generates fake_impl when not provided.""" |
| 811 | with ( |
| 812 | dummy_kernel_registry() as register, |
| 813 | patch("vllm.kernels.helion.register.infer_fake_impl") as mock_infer, |
| 814 | ): |
| 815 | mock_fake = Mock() |
| 816 | mock_infer.return_value = mock_fake |
| 817 | wrapper = register( |
| 818 | config_picker=lambda args, keys: None, |
| 819 | )(_add_kernel) |
| 820 | |
| 821 | mock_infer.assert_called_once_with(_add_kernel, None) |
| 822 | assert wrapper._fake_impl is mock_fake |
| 823 | |
| 824 | def test_register_kernel_creates_wrapper(self): |
| 825 | """Test register_kernel creates HelionKernelWrapper.""" |
| 826 | with dummy_kernel_registry() as register: |
| 827 | result = register("test_name", config_picker=lambda args, keys: None)( |
| 828 | _add_kernel |
| 829 | ) |
| 830 | |
| 831 | assert isinstance(result, HelionKernelWrapper) |
| 832 | assert result.op_name == "test_name" |
| 833 | assert result.raw_kernel_func is _add_kernel |
| 834 | |
| 835 | def test_register_kernel_auto_detects_name(self): |
| 836 | """Test register_kernel uses function name when no name provided.""" |
| 837 | with dummy_kernel_registry() as register: |
| 838 | wrapper = register(config_picker=lambda args, keys: None)(_add_kernel) |
| 839 | |
| 840 | assert wrapper.op_name == "_add_kernel" |
| 841 | |
| 842 | def test_register_kernel_registers_in_global_registry(self): |
| 843 | """Test register_kernel adds wrapper to global registry.""" |
| 844 | with dummy_kernel_registry() as register: |
| 845 | wrapper = register("test_kernel", config_picker=lambda args, keys: None)( |
| 846 | _add_kernel |
| 847 | ) |
| 848 | |
| 849 | registered_kernels = get_registered_kernels() |
| 850 | assert "test_kernel" in registered_kernels |
| 851 | assert registered_kernels["test_kernel"] is wrapper |
| 852 | |
| 853 | def test_register_kernel_passes_helion_settings(self): |
| 854 | """Test register_kernel passes helion_settings to wrapper.""" |
| 855 | settings = helion.Settings() |
| 856 | settings.print_output_code = True |
| 857 | |
| 858 | with dummy_kernel_registry() as register: |
| 859 | result = register( |
| 860 | "test_name", |
| 861 | config_picker=lambda args, keys: None, |
| 862 | helion_settings=settings, |
| 863 | )(_add_kernel) |
| 864 | |
| 865 | assert result.helion_settings is settings |
| 866 | |
| 867 | def test_register_kernel_supports_decorator_syntax(self): |
| 868 | """Test register_kernel works with decorator arguments.""" |
| 869 | mock_fake = Mock() |
| 870 | |
| 871 | with dummy_kernel_registry() as register: |
| 872 | result = register( |
| 873 | "custom_name", |
| 874 | config_picker=lambda args, keys: None, |
| 875 | fake_impl=mock_fake, |
| 876 | )(_add_kernel) |
| 877 | |
| 878 | assert result.op_name == "custom_name" |
| 879 | assert result._fake_impl is mock_fake |
| 880 | |
| 881 | def test_register_kernel_raises_on_duplicate_registration(self): |
| 882 | """Test register_kernel raises error on duplicate names.""" |
| 883 | with dummy_kernel_registry() as register: |
| 884 | register("duplicate_name", config_picker=lambda args, keys: None)( |
| 885 | _add_kernel |
| 886 | ) |
| 887 | |
| 888 | with pytest.raises(ValueError, match="already registered"): |
| 889 | register("duplicate_name", config_picker=lambda args, keys: None)( |
| 890 | _add_kernel |
| 891 | ) |
| 892 | |
| 893 | def test_register_kernel_rejects_autotuner_fn_in_settings(self): |
| 894 | """Test register_kernel rejects conflicting autotuner_fn.""" |
| 895 | mock_settings = Mock() |
| 896 | mock_settings.to_dict.return_value = {"autotuner_fn": Mock()} |
| 897 | |
| 898 | with pytest.raises(ValueError, match="uses a custom autotuner"): |
| 899 | |
| 900 | @register_kernel( |
| 901 | "test", |
| 902 | config_picker=lambda args, keys: None, |
| 903 | helion_settings=mock_settings, |
| 904 | ) |
| 905 | def test_kernel(x): |
| 906 | return x |
| 907 | |
| 908 | def test_register_kernel_no_warning_with_static_shapes_false(self): |
| 909 | """Test register_kernel doesn't warn with static_shapes=False.""" |
| 910 | mock_settings = Mock() |
| 911 | mock_settings.to_dict.return_value = {"static_shapes": False} |
| 912 | |
| 913 | with ( |
| 914 | dummy_kernel_registry() as register, |
| 915 | patch("vllm.kernels.helion.register.logger") as mock_logger, |
| 916 | ): |
| 917 | register( |
| 918 | "test", |
| 919 | config_picker=lambda args, keys: None, |
| 920 | helion_settings=mock_settings, |
| 921 | )(_add_kernel) |
| 922 | |
| 923 | mock_logger.warning.assert_not_called() |
| 924 | |
| 925 | def test_disabled_kernel_appears_in_registry(self): |
| 926 | """Test that a disabled wrapper is still in the global registry.""" |
| 927 | |
| 928 | def fake_impl(*args, **kwargs): |
| 929 | return torch.zeros_like(args[0]) |
| 930 | |
| 931 | mock_config_manager = Mock(spec=ConfigManager) |
| 932 | mock_config_manager.get_platform_configs = Mock(return_value={}) |
| 933 | |
| 934 | with ( |
| 935 | patch( |
| 936 | "vllm.kernels.helion.config_manager.ConfigManager", |
| 937 | return_value=mock_config_manager, |
| 938 | ), |
| 939 | patch( |
| 940 | "vllm.kernels.helion.utils.get_canonical_gpu_name", |
| 941 | return_value="nvidia_h200", |
| 942 | ), |
| 943 | patch("vllm.kernels.helion.register.helion.kernel") as mock_kernel, |
| 944 | ): |
| 945 | mock_kernel.return_value = Mock(return_value=_add_kernel) |
| 946 | |
| 947 | wrapper = register_kernel( |
| 948 | "disabled_kernel", |
| 949 | config_picker=lambda args, keys: None, |
| 950 | fake_impl=fake_impl, |
| 951 | )(_add_kernel) |
| 952 | |
| 953 | assert wrapper._disabled is True |
| 954 | registered = get_registered_kernels() |
| 955 | assert "disabled_kernel" in registered |
| 956 | assert registered["disabled_kernel"] is wrapper |
| 957 | |
| 958 | |
| 959 | @pytest.mark.skipif(not _HOP_AVAILABLE, reason="Requires PyTorch >= 2.11 for HOP") |
| 960 | class TestTorchCompileHOP: |
| 961 | """Test that HelionKernelWrapper emits the correct HOP under torch.compile.""" |
| 962 | |
| 963 | def test_compiled_graph_contains_helion_hop(self): |
| 964 | """Verify torch.compile on a HelionKernelWrapper emits a |
| 965 | helion_kernel_wrapper_mutation HOP node in the FX graph.""" |
| 966 | configs: dict[CaseKey, helion.Config] = { |
| 967 | CaseKey.default(): helion.Config(block_sizes=[4, 4]) |
| 968 | } |
| 969 | |
| 970 | with dummy_kernel_registry(configs=configs) as register: |
| 971 | add_helion_kernel = register( |
| 972 | op_name="test_torch_compile_add_kernel", |
| 973 | config_picker=lambda args, keys: None, |
| 974 | )(_add_kernel) |
| 975 | |
| 976 | captured_graph: torch.fx.GraphModule | None = None |
| 977 | |
| 978 | def capturing_backend(gm, example_inputs): |
| 979 | nonlocal captured_graph |
| 980 | assert captured_graph is None, "Backend called multiple times" |
| 981 | captured_graph = gm |
| 982 | return gm.forward |
| 983 | |
| 984 | def f(x, y): |
| 985 | return add_helion_kernel(x, y) |
| 986 | |
| 987 | torch._dynamo.reset() |
| 988 | compiled_f = torch.compile(f, backend=capturing_backend, fullgraph=True) |
| 989 | |
| 990 | x = torch.randn(4, 4, device="cuda") |
| 991 | y = torch.randn(4, 4, device="cuda") |
| 992 | |
| 993 | # Run compiled version and capture graph |
| 994 | compiled_result = compiled_f(x, y) |
| 995 | |
| 996 | assert captured_graph is not None |
| 997 | hop_nodes = [ |
| 998 | node |
| 999 | for node in captured_graph.graph.nodes |
| 1000 | if node.op == "call_function" |
| 1001 | and node.target is helion_kernel_wrapper_mutation |
| 1002 | ] |
| 1003 | assert len(hop_nodes) > 0, ( |
| 1004 | "Expected helion_kernel_wrapper_mutation HOP node in compiled graph, " |
| 1005 | f"but found none. Graph nodes: " |
| 1006 | f"{[(n.op, n.target) for n in captured_graph.graph.nodes]}" |
| 1007 | ) |
| 1008 | |
| 1009 | # Verify compiled result matches eager execution |
| 1010 | eager_result = f(x, y) # Run in eager mode |
| 1011 | |
| 1012 | assert torch.allclose(compiled_result, eager_result, atol=1e-5, rtol=1e-5), ( |
| 1013 | "Compiled execution result doesn't match eager execution. " |
| 1014 | f"Max difference: {torch.max(torch.abs(compiled_result - eager_result))}" |
| 1015 | ) |
| 1016 | |
| 1017 | @pytest.mark.skipif( |
| 1018 | not (_HOP_AVAILABLE and supports_torch_compile_fusion()), |
| 1019 | reason="Requires PyTorch with Helion inductor fusion support", |
| 1020 | ) |
| 1021 | def test_inductor_backend_compiles_helion_hop(self): |
| 1022 | """Test torch.compile with inductor backend and Helion fusion enabled.""" |
| 1023 | configs: dict[CaseKey, helion.Config] = { |
| 1024 | CaseKey.default(): helion.Config(block_sizes=[4, 4]) |
| 1025 | } |
| 1026 | |
| 1027 | with dummy_kernel_registry(configs=configs) as register: |
| 1028 | add_helion_kernel = register( |
| 1029 | op_name="test_inductor_add_kernel", |
| 1030 | config_picker=lambda args, keys: None, |
| 1031 | helion_settings=helion.Settings( |
| 1032 | torch_compile_fusion=True, static_shapes=False |
| 1033 | ), |
| 1034 | )(_add_kernel) |
| 1035 | |
| 1036 | def f(x, y): |
| 1037 | x = x * 2.0 |
| 1038 | y = y + 1.0 |
| 1039 | out = add_helion_kernel(x, y) |
| 1040 | return out.relu() |
| 1041 | |
| 1042 | torch._dynamo.reset() |
| 1043 | compiled_f = torch.compile(f, backend="inductor", fullgraph=True) |
| 1044 | |
| 1045 | x = torch.randn(4, 4, device="cuda") |
| 1046 | y = torch.randn(4, 4, device="cuda") |
| 1047 | |
| 1048 | compiled_result, source_codes = run_and_get_code(compiled_f, x, y) |
| 1049 | eager_result = f(x, y) |
| 1050 | |
| 1051 | assert torch.allclose(compiled_result, eager_result, atol=1e-5, rtol=1e-5), ( |
| 1052 | "Inductor-compiled result doesn't match eager execution. " |
| 1053 | f"Max difference: {torch.max(torch.abs(compiled_result - eager_result))}" |
| 1054 | ) |
| 1055 | |
| 1056 | # With fusion enabled, prologue/epilogue ops should be fused into |
| 1057 | # a single triton kernel rather than generating separate kernels. |
| 1058 | kernel_count = sum(code.count("@triton.jit") for code in source_codes) |
| 1059 | assert kernel_count == 1, ( |
| 1060 | f"Expected 1 fused triton kernel, got {kernel_count}. " |
| 1061 | "Prologue/epilogue ops were not fused into the Helion kernel." |
| 1062 | ) |