| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for MLA prefill backend selector.""" |
| 4 | |
| 5 | from unittest.mock import MagicMock, patch |
| 6 | |
| 7 | import pytest |
| 8 | import torch |
| 9 | |
| 10 | from vllm.config import AttentionConfig, ModelConfig, VllmConfig |
| 11 | from vllm.platforms import current_platform |
| 12 | from vllm.platforms.interface import DeviceCapability |
| 13 | from vllm.v1.attention.backends.mla.prefill.base import MLADimensions |
| 14 | from vllm.v1.attention.backends.mla.prefill.registry import MLAPrefillBackendEnum |
| 15 | from vllm.v1.attention.backends.mla.prefill.selector import ( |
| 16 | MLAPrefillSelectorConfig, |
| 17 | _auto_select_mla_prefill_backend, |
| 18 | _get_mla_prefill_backend_priorities, |
| 19 | get_mla_prefill_backend, |
| 20 | ) |
| 21 | |
| 22 | |
| 23 | @pytest.fixture(autouse=True) |
| 24 | def clear_cache(): |
| 25 | """Clear lru cache to ensure each test case runs without caching.""" |
| 26 | _auto_select_mla_prefill_backend.cache_clear() |
| 27 | |
| 28 | |
| 29 | def _make_mock_model_config( |
| 30 | qk_nope_head_dim: int = 128, |
| 31 | qk_rope_head_dim: int = 64, |
| 32 | v_head_dim: int = 128, |
| 33 | dtype: torch.dtype = torch.bfloat16, |
| 34 | ) -> ModelConfig: |
| 35 | mock_config = MagicMock(spec=ModelConfig) |
| 36 | mock_config.dtype = dtype |
| 37 | mock_config.hf_text_config = MagicMock() |
| 38 | mock_config.hf_text_config.qk_nope_head_dim = qk_nope_head_dim |
| 39 | mock_config.hf_text_config.qk_rope_head_dim = qk_rope_head_dim |
| 40 | mock_config.hf_text_config.v_head_dim = v_head_dim |
| 41 | return mock_config |
| 42 | |
| 43 | |
| 44 | def _make_vllm_config( |
| 45 | model_config: ModelConfig | None = None, |
| 46 | mla_prefill_backend: MLAPrefillBackendEnum | None = None, |
| 47 | ) -> VllmConfig: |
| 48 | if model_config is None: |
| 49 | model_config = _make_mock_model_config() |
| 50 | |
| 51 | attention_config = AttentionConfig(mla_prefill_backend=mla_prefill_backend) |
| 52 | mock_vllm_config = MagicMock(spec=VllmConfig) |
| 53 | mock_vllm_config.model_config = model_config |
| 54 | mock_vllm_config.attention_config = attention_config |
| 55 | return mock_vllm_config |
| 56 | |
| 57 | |
| 58 | class TestGetMLAPrefillBackend: |
| 59 | """Tests for get_mla_prefill_backend (public API).""" |
| 60 | |
| 61 | def test_cpu_uses_sdpa_prefill(self): |
| 62 | vllm_config = _make_vllm_config() |
| 63 | zen_cls = MLAPrefillBackendEnum.ZEN_CPU.get_class() |
| 64 | |
| 65 | # A Zen CPU with zentorch selects the Zen backend, so expect whichever |
| 66 | # one this host offers. |
| 67 | expected = ( |
| 68 | zen_cls if zen_cls.is_available() else MLAPrefillBackendEnum.CPU.get_class() |
| 69 | ) |
| 70 | |
| 71 | with patch("vllm.platforms.current_platform") as mock_platform: |
| 72 | mock_platform.is_cpu.return_value = True |
| 73 | |
| 74 | backend = get_mla_prefill_backend(vllm_config) |
| 75 | assert backend is expected |
| 76 | |
| 77 | def test_zen_cpu_prefers_zentorch_prefill(self): |
| 78 | vllm_config = _make_vllm_config() |
| 79 | zen_cls = MLAPrefillBackendEnum.ZEN_CPU.get_class() |
| 80 | |
| 81 | with ( |
| 82 | patch("vllm.platforms.current_platform") as mock_platform, |
| 83 | patch.object(zen_cls, "is_available", return_value=True), |
| 84 | ): |
| 85 | mock_platform.is_cpu.return_value = True |
| 86 | |
| 87 | backend = get_mla_prefill_backend(vllm_config) |
| 88 | assert backend is zen_cls |
| 89 | |
| 90 | def test_zen_cpu_falls_back_when_zentorch_unavailable(self): |
| 91 | vllm_config = _make_vllm_config() |
| 92 | zen_cls = MLAPrefillBackendEnum.ZEN_CPU.get_class() |
| 93 | |
| 94 | with ( |
| 95 | patch("vllm.platforms.current_platform") as mock_platform, |
| 96 | patch.object(zen_cls, "is_available", return_value=False), |
| 97 | ): |
| 98 | mock_platform.is_cpu.return_value = True |
| 99 | |
| 100 | backend = get_mla_prefill_backend(vllm_config) |
| 101 | assert backend is MLAPrefillBackendEnum.CPU.get_class() |
| 102 | |
| 103 | def test_no_device_capability_returns_flash_attn(self): |
| 104 | vllm_config = _make_vllm_config() |
| 105 | |
| 106 | class FlashAttnBackend: |
| 107 | @staticmethod |
| 108 | def get_name(): |
| 109 | return "FLASH_ATTN" |
| 110 | |
| 111 | with ( |
| 112 | patch("vllm.platforms.current_platform") as mock_platform, |
| 113 | patch.object( |
| 114 | MLAPrefillBackendEnum.FLASH_ATTN, |
| 115 | "get_class", |
| 116 | return_value=FlashAttnBackend, |
| 117 | ), |
| 118 | ): |
| 119 | mock_platform.get_device_capability.return_value = None |
| 120 | mock_platform.is_cpu.return_value = False |
| 121 | |
| 122 | backend = get_mla_prefill_backend(vllm_config) |
| 123 | assert backend.get_name() == "FLASH_ATTN" |
| 124 | |
| 125 | def test_explicit_flash_attn_selection(self): |
| 126 | try: |
| 127 | flash_attn_cls = MLAPrefillBackendEnum.FLASH_ATTN.get_class() |
| 128 | except ImportError: |
| 129 | pytest.skip("FLASH_ATTN backend not available") |
| 130 | return |
| 131 | |
| 132 | vllm_config = _make_vllm_config( |
| 133 | mla_prefill_backend=MLAPrefillBackendEnum.FLASH_ATTN, |
| 134 | ) |
| 135 | |
| 136 | with patch("vllm.platforms.current_platform") as mock_platform: |
| 137 | mock_platform.is_cpu.return_value = False |
| 138 | mock_platform.get_device_capability.return_value = DeviceCapability( |
| 139 | major=9, minor=0 |
| 140 | ) |
| 141 | |
| 142 | with patch.object( |
| 143 | flash_attn_cls, |
| 144 | "validate_configuration", |
| 145 | return_value=[], |
| 146 | ): |
| 147 | backend = get_mla_prefill_backend(vllm_config) |
| 148 | assert backend.get_name() == "FLASH_ATTN" |
| 149 | |
| 150 | def test_explicit_backend_invalid_raises_error(self): |
| 151 | vllm_config = _make_vllm_config( |
| 152 | mla_prefill_backend=MLAPrefillBackendEnum.FLASHINFER, |
| 153 | ) |
| 154 | |
| 155 | with patch("vllm.platforms.current_platform") as mock_platform: |
| 156 | mock_platform.is_cpu.return_value = False |
| 157 | mock_platform.get_device_capability.return_value = DeviceCapability( |
| 158 | major=9, minor=0 |
| 159 | ) |
| 160 | |
| 161 | with pytest.raises(ValueError, match="is not valid"): |
| 162 | get_mla_prefill_backend(vllm_config) |
| 163 | |
| 164 | def test_explicit_backend_import_error_raises(self): |
| 165 | vllm_config = _make_vllm_config( |
| 166 | mla_prefill_backend=MLAPrefillBackendEnum.TRTLLM_RAGGED, |
| 167 | ) |
| 168 | |
| 169 | with patch("vllm.platforms.current_platform") as mock_platform: |
| 170 | mock_platform.is_cpu.return_value = False |
| 171 | mock_platform.get_device_capability.return_value = DeviceCapability( |
| 172 | major=10, minor=0 |
| 173 | ) |
| 174 | |
| 175 | with ( |
| 176 | patch.object( |
| 177 | MLAPrefillBackendEnum.TRTLLM_RAGGED, |
| 178 | "get_class", |
| 179 | side_effect=ImportError("trtllm not installed"), |
| 180 | ), |
| 181 | pytest.raises(ValueError, match="is not valid"), |
| 182 | ): |
| 183 | get_mla_prefill_backend(vllm_config) |
| 184 | |
| 185 | @pytest.mark.parametrize( |
| 186 | ("qk_nope_head_dim", "v_head_dim"), |
| 187 | [(128, 128), (192, 256), (64, 128)], |
| 188 | ids=["deepseek", "glm", "mistral_s4"], |
| 189 | ) |
| 190 | def test_auto_selection_on_hopper(self, qk_nope_head_dim: int, v_head_dim: int): |
| 191 | try: |
| 192 | flash_attn_cls = MLAPrefillBackendEnum.FLASH_ATTN.get_class() |
| 193 | except ImportError: |
| 194 | pytest.skip("FLASH_ATTN backend not available") |
| 195 | return |
| 196 | |
| 197 | vllm_config = _make_vllm_config( |
| 198 | _make_mock_model_config( |
| 199 | qk_nope_head_dim=qk_nope_head_dim, |
| 200 | v_head_dim=v_head_dim, |
| 201 | ) |
| 202 | ) |
| 203 | |
| 204 | with ( |
| 205 | patch("vllm.platforms.current_platform") as mock_platform, |
| 206 | patch.object(flash_attn_cls, "is_available", return_value=True), |
| 207 | patch( |
| 208 | "vllm.v1.attention.backends.mla.prefill.flash_attn." |
| 209 | "get_flash_attn_version", |
| 210 | return_value=3, |
| 211 | ), |
| 212 | ): |
| 213 | mock_platform.is_cpu.return_value = False |
| 214 | mock_platform.get_device_capability.return_value = DeviceCapability( |
| 215 | major=9, minor=0 |
| 216 | ) |
| 217 | mock_platform.is_rocm.return_value = False |
| 218 | |
| 219 | backend = get_mla_prefill_backend(vllm_config) |
| 220 | assert backend.get_name() == "FLASH_ATTN" |
| 221 | |
| 222 | |
| 223 | class TestAutoSelectMLAPrefillBackend: |
| 224 | """Tests for fallback and error paths in auto-selection.""" |
| 225 | |
| 226 | def test_blackwell_glm_dimensions_use_trtllm(self): |
| 227 | capability = DeviceCapability(major=10, minor=0) |
| 228 | selector_config = MLAPrefillSelectorConfig( |
| 229 | dtype=torch.bfloat16, |
| 230 | mla_dimensions=MLADimensions( |
| 231 | qk_nope_head_dim=192, |
| 232 | qk_rope_head_dim=64, |
| 233 | v_head_dim=256, |
| 234 | ), |
| 235 | ) |
| 236 | |
| 237 | try: |
| 238 | flash_attn_cls = MLAPrefillBackendEnum.FLASH_ATTN.get_class() |
| 239 | trtllm_cls = MLAPrefillBackendEnum.TRTLLM_RAGGED.get_class() |
| 240 | except ImportError: |
| 241 | pytest.skip("MLA prefill backend not available") |
| 242 | return |
| 243 | |
| 244 | with ( |
| 245 | patch("vllm.platforms.current_platform") as mock_platform, |
| 246 | patch.object(flash_attn_cls, "is_available", return_value=True), |
| 247 | patch.object(trtllm_cls, "validate_configuration", return_value=[]), |
| 248 | ): |
| 249 | # Force the non-ROCm priority on the Blackwell. |
| 250 | mock_platform.is_rocm.return_value = False |
| 251 | backend = _auto_select_mla_prefill_backend( |
| 252 | capability, |
| 253 | selector_config, |
| 254 | ) |
| 255 | assert backend.get_name() == "TRTLLM_RAGGED" |
| 256 | |
| 257 | def test_all_fail_raises_error(self): |
| 258 | capability = DeviceCapability(major=10, minor=0) |
| 259 | selector_config = MLAPrefillSelectorConfig( |
| 260 | dtype=torch.bfloat16, |
| 261 | mla_dimensions=MLADimensions( |
| 262 | qk_nope_head_dim=128, |
| 263 | qk_rope_head_dim=64, |
| 264 | v_head_dim=128, |
| 265 | ), |
| 266 | ) |
| 267 | |
| 268 | def mock_get_class(backend_enum): # noqa: ARG001 |
| 269 | cls = MagicMock() |
| 270 | cls.validate_configuration.return_value = ["not available"] |
| 271 | return cls |
| 272 | |
| 273 | with patch.object(MLAPrefillBackendEnum, "get_class", mock_get_class): |
| 274 | _auto_select_mla_prefill_backend.cache_clear() |
| 275 | with pytest.raises(ValueError, match="No valid MLA"): |
| 276 | _auto_select_mla_prefill_backend( |
| 277 | capability, |
| 278 | selector_config, |
| 279 | ) |
| 280 | |
| 281 | |
| 282 | class TestBackendValidation: |
| 283 | """Tests for backend validation logic.""" |
| 284 | |
| 285 | def test_backend_supported_dimension_validation(self): |
| 286 | try: |
| 287 | from vllm.v1.attention.backends.mla.prefill.flashinfer import ( |
| 288 | FlashInferPrefillBackend, |
| 289 | ) |
| 290 | from vllm.v1.attention.backends.mla.prefill.trtllm_ragged import ( |
| 291 | TrtllmRaggedPrefillBackend, |
| 292 | ) |
| 293 | except ImportError: |
| 294 | pytest.skip("MLA prefill backend not available") |
| 295 | return |
| 296 | |
| 297 | capability = DeviceCapability(major=10, minor=0) |
| 298 | selector_config = MLAPrefillSelectorConfig( |
| 299 | dtype=torch.bfloat16, |
| 300 | mla_dimensions=MLADimensions( |
| 301 | qk_nope_head_dim=128, |
| 302 | qk_rope_head_dim=64, |
| 303 | v_head_dim=128, |
| 304 | ), |
| 305 | ) |
| 306 | |
| 307 | with patch.object(FlashInferPrefillBackend, "is_available", return_value=True): |
| 308 | invalid_reasons = FlashInferPrefillBackend.validate_configuration( |
| 309 | capability, |
| 310 | selector_config, |
| 311 | ) |
| 312 | assert len(invalid_reasons) == 0 |
| 313 | |
| 314 | selector_config_invalid = MLAPrefillSelectorConfig( |
| 315 | dtype=torch.bfloat16, |
| 316 | mla_dimensions=MLADimensions( |
| 317 | qk_nope_head_dim=64, |
| 318 | qk_rope_head_dim=64, |
| 319 | v_head_dim=128, |
| 320 | ), |
| 321 | ) |
| 322 | |
| 323 | with patch.object(FlashInferPrefillBackend, "is_available", return_value=True): |
| 324 | invalid_reasons = FlashInferPrefillBackend.validate_configuration( |
| 325 | capability, |
| 326 | selector_config_invalid, |
| 327 | ) |
| 328 | assert len(invalid_reasons) == 1 |
| 329 | assert "supported MLA dimensions" in invalid_reasons[0] |
| 330 | |
| 331 | selector_config_glm5 = MLAPrefillSelectorConfig( |
| 332 | dtype=torch.bfloat16, |
| 333 | mla_dimensions=MLADimensions( |
| 334 | qk_nope_head_dim=192, |
| 335 | qk_rope_head_dim=64, |
| 336 | v_head_dim=256, |
| 337 | ), |
| 338 | ) |
| 339 | |
| 340 | with patch.object( |
| 341 | TrtllmRaggedPrefillBackend, "is_available", return_value=True |
| 342 | ): |
| 343 | invalid_reasons = TrtllmRaggedPrefillBackend.validate_configuration( |
| 344 | capability, |
| 345 | selector_config_glm5, |
| 346 | ) |
| 347 | assert invalid_reasons == [] |
| 348 | |
| 349 | def test_flash_attn_accepts_glm53_flash_nope_dimensions(self): |
| 350 | """(256, 0, 256) runs the same FA kernels as GLM-5's (192, 64, 256); a |
| 351 | RoPE-carrying 320-wide query does not.""" |
| 352 | try: |
| 353 | from vllm.v1.attention.backends.mla.prefill.flash_attn import ( |
| 354 | FlashAttnPrefillBackend, |
| 355 | ) |
| 356 | except ImportError: |
| 357 | pytest.skip("MLA prefill backend not available") |
| 358 | return |
| 359 | |
| 360 | capability = DeviceCapability(major=10, minor=0) |
| 361 | |
| 362 | def validate(qk_rope_head_dim: int) -> list[str]: |
| 363 | selector_config = MLAPrefillSelectorConfig( |
| 364 | dtype=torch.bfloat16, |
| 365 | mla_dimensions=MLADimensions( |
| 366 | qk_nope_head_dim=256, |
| 367 | qk_rope_head_dim=qk_rope_head_dim, |
| 368 | v_head_dim=256, |
| 369 | ), |
| 370 | ) |
| 371 | with patch.object( |
| 372 | FlashAttnPrefillBackend, "is_available", return_value=True |
| 373 | ): |
| 374 | return FlashAttnPrefillBackend.validate_configuration( |
| 375 | capability, selector_config |
| 376 | ) |
| 377 | |
| 378 | assert validate(qk_rope_head_dim=0) == [] |
| 379 | invalid_reasons = validate(qk_rope_head_dim=64) |
| 380 | assert len(invalid_reasons) == 1 |
| 381 | assert "supported MLA dimensions" in invalid_reasons[0] |
| 382 | |
| 383 | |
| 384 | @pytest.mark.skipif( |
| 385 | not current_platform.is_cuda_alike(), |
| 386 | reason="Imports vllm.platforms.rocm, whose module init requires a CUDA or " |
| 387 | "ROCm torch build; not importable on XPU/CPU/TPU.", |
| 388 | ) |
| 389 | class TestROCmAiterFAPrefillSelection: |
| 390 | """Tests for the ROCm AITER FlashAttention MLA prefill backend.""" |
| 391 | |
| 392 | def test_rocm_priorities_prefer_aiter_fa(self): |
| 393 | """On ROCm, ROCM_AITER_FA is tried first, FLASH_ATTN as fallback.""" |
| 394 | with patch("vllm.platforms.current_platform") as mock_platform: |
| 395 | mock_platform.is_rocm.return_value = True |
| 396 | priorities = _get_mla_prefill_backend_priorities( |
| 397 | DeviceCapability(major=9, minor=5), |
| 398 | MLADimensions( |
| 399 | qk_nope_head_dim=128, |
| 400 | qk_rope_head_dim=64, |
| 401 | v_head_dim=128, |
| 402 | ), |
| 403 | ) |
| 404 | |
| 405 | assert priorities == [ |
| 406 | MLAPrefillBackendEnum.ROCM_AITER_FA, |
| 407 | MLAPrefillBackendEnum.FLASH_ATTN, |
| 408 | ] |
| 409 | |
| 410 | def test_supported_dtypes_are_fp16_bf16_only(self): |
| 411 | from vllm.v1.attention.backends.mla.prefill.aiter_flash_attn import ( |
| 412 | AiterFlashAttnPrefillBackend, |
| 413 | ) |
| 414 | |
| 415 | assert AiterFlashAttnPrefillBackend.supports_dtype(torch.bfloat16) |
| 416 | assert AiterFlashAttnPrefillBackend.supports_dtype(torch.float16) |
| 417 | # FP8 is served by the separate AITER ASM backend, not this one. |
| 418 | assert not AiterFlashAttnPrefillBackend.supports_dtype(torch.float8_e4m3fn) |
| 419 | |
| 420 | def test_supports_compute_capability_on_rocm(self): |
| 421 | from vllm.v1.attention.backends.mla.prefill import aiter_flash_attn as mod |
| 422 | |
| 423 | # Gating is decided by on_mi3xx(), not by capability |
| 424 | capability = MagicMock() |
| 425 | |
| 426 | with patch.object(mod.current_platform, "is_rocm", return_value=False): |
| 427 | assert not mod.AiterFlashAttnPrefillBackend.supports_compute_capability( |
| 428 | capability |
| 429 | ) |
| 430 | |
| 431 | with ( |
| 432 | patch.object(mod.current_platform, "is_rocm", return_value=True), |
| 433 | patch("vllm.platforms.rocm.on_mi3xx", return_value=False), |
| 434 | ): |
| 435 | assert not mod.AiterFlashAttnPrefillBackend.supports_compute_capability( |
| 436 | capability |
| 437 | ) |
| 438 | |
| 439 | with ( |
| 440 | patch.object(mod.current_platform, "is_rocm", return_value=True), |
| 441 | patch("vllm.platforms.rocm.on_mi3xx", return_value=True), |
| 442 | ): |
| 443 | assert mod.AiterFlashAttnPrefillBackend.supports_compute_capability( |
| 444 | capability |
| 445 | ) |
| 446 | |
| 447 | def test_is_available_delegates_to_rocm_aiter_ops(self): |
| 448 | from vllm._aiter_ops import rocm_aiter_ops |
| 449 | from vllm.v1.attention.backends.mla.prefill import aiter_flash_attn as mod |
| 450 | |
| 451 | with patch.object(rocm_aiter_ops, "is_enabled", return_value=False): |
| 452 | assert not mod.AiterFlashAttnPrefillBackend.is_available() |
| 453 | |
| 454 | with patch.object(rocm_aiter_ops, "is_enabled", return_value=True): |
| 455 | assert mod.AiterFlashAttnPrefillBackend.is_available() |
| 456 | |
| 457 | def test_auto_select_prefers_aiter_fa_on_rocm(self): |
| 458 | from vllm.v1.attention.backends.mla.prefill.aiter_flash_attn import ( |
| 459 | AiterFlashAttnPrefillBackend, |
| 460 | ) |
| 461 | |
| 462 | # gfx gating is simulated via the mocked validate_configuration, |
| 463 | # not the capability. |
| 464 | capability = MagicMock() |
| 465 | selector_config = MLAPrefillSelectorConfig(dtype=torch.bfloat16) |
| 466 | |
| 467 | with ( |
| 468 | patch("vllm.platforms.current_platform") as mock_platform, |
| 469 | patch.object( |
| 470 | AiterFlashAttnPrefillBackend, |
| 471 | "validate_configuration", |
| 472 | return_value=[], |
| 473 | ), |
| 474 | ): |
| 475 | mock_platform.is_rocm.return_value = True |
| 476 | backend = _auto_select_mla_prefill_backend(capability, selector_config) |
| 477 | assert backend.get_name() == "ROCM_AITER_FA" |
| 478 | |
| 479 | def test_auto_select_falls_back_to_flash_attn_when_aiter_invalid(self): |
| 480 | from vllm.v1.attention.backends.mla.prefill.aiter_flash_attn import ( |
| 481 | AiterFlashAttnPrefillBackend, |
| 482 | ) |
| 483 | |
| 484 | try: |
| 485 | flash_attn_cls = MLAPrefillBackendEnum.FLASH_ATTN.get_class() |
| 486 | except ImportError: |
| 487 | pytest.skip("FLASH_ATTN backend not available") |
| 488 | return |
| 489 | |
| 490 | # the fallback is forced by the mocked validate_configuration, |
| 491 | # not the capability. |
| 492 | capability = MagicMock() |
| 493 | selector_config = MLAPrefillSelectorConfig(dtype=torch.bfloat16) |
| 494 | |
| 495 | with ( |
| 496 | patch("vllm.platforms.current_platform") as mock_platform, |
| 497 | patch.object( |
| 498 | AiterFlashAttnPrefillBackend, |
| 499 | "validate_configuration", |
| 500 | return_value=["compute capability not supported"], |
| 501 | ), |
| 502 | patch.object(flash_attn_cls, "validate_configuration", return_value=[]), |
| 503 | ): |
| 504 | mock_platform.is_rocm.return_value = True |
| 505 | backend = _auto_select_mla_prefill_backend(capability, selector_config) |
| 506 | assert backend.get_name() == "FLASH_ATTN" |
| 507 | |
| 508 | |
| 509 | class TestMLAPrefillBackendParsing: |
| 510 | """Tests for string-based mla_prefill_backend parsing from CLI args.""" |
| 511 | |
| 512 | def test_valid_string_parses_to_enum(self): |
| 513 | config = AttentionConfig( |
| 514 | mla_prefill_backend="FLASH_ATTN", |
| 515 | ) |
| 516 | assert config.mla_prefill_backend == MLAPrefillBackendEnum.FLASH_ATTN |
| 517 | |
| 518 | def test_invalid_string_raises_error(self): |
| 519 | with pytest.raises(ValueError, match="Unknown MLA prefill backend"): |
| 520 | AttentionConfig( |
| 521 | mla_prefill_backend="NONEXISTENT", |
| 522 | ) |
| 523 | |
| 524 | |
| 525 | class TestMLAPrefillBackendConfig: |
| 526 | """Tests for mla_prefill_backend configuration in AttentionConfig.""" |
| 527 | |
| 528 | def test_default_backend_is_none(self): |
| 529 | config = AttentionConfig() |
| 530 | assert config.mla_prefill_backend is None |
| 531 | |
| 532 | def test_explicit_flash_attn_backend(self): |
| 533 | config = AttentionConfig( |
| 534 | mla_prefill_backend=MLAPrefillBackendEnum.FLASH_ATTN, |
| 535 | ) |
| 536 | assert config.mla_prefill_backend == MLAPrefillBackendEnum.FLASH_ATTN |
| 537 | |
| 538 | def test_explicit_trtllm_ragged_backend(self): |
| 539 | config = AttentionConfig( |
| 540 | mla_prefill_backend=MLAPrefillBackendEnum.TRTLLM_RAGGED, |
| 541 | ) |
| 542 | assert config.mla_prefill_backend == MLAPrefillBackendEnum.TRTLLM_RAGGED |