| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from unittest.mock import patch |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | |
| 8 | from tests.kernels.moe.utils import make_dummy_moe_config |
| 9 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 10 | from vllm.model_executor.layers.fused_moe.config import RoutingMethodType |
| 11 | from vllm.model_executor.layers.fused_moe.oracle.unquantized import ( |
| 12 | UnquantizedMoeBackend, |
| 13 | backend_to_kernel_cls, |
| 14 | select_unquantized_moe_backend, |
| 15 | ) |
| 16 | from vllm.platforms import CpuArchEnum, current_platform |
| 17 | |
| 18 | skipif_not_cuda_rocm = pytest.mark.skipif( |
| 19 | not (current_platform.is_cuda() or current_platform.is_rocm()), |
| 20 | reason="Only supported on CUDA/ROCm platforms.", |
| 21 | ) |
| 22 | |
| 23 | |
| 24 | @pytest.mark.parametrize( |
| 25 | ("amx_supported", "in_dtype", "expect_amx_kernel"), |
| 26 | [ |
| 27 | (True, torch.bfloat16, True), |
| 28 | (False, torch.bfloat16, False), |
| 29 | (True, torch.float16, False), |
| 30 | ], |
| 31 | ) |
| 32 | def test_x86_cpu_unquantized_kernel_selection( |
| 33 | amx_supported: bool, |
| 34 | in_dtype: torch.dtype, |
| 35 | expect_amx_kernel: bool, |
| 36 | ): |
| 37 | from vllm.model_executor.layers.fused_moe.experts.cpu_moe import ( |
| 38 | CPUUnquantizedExperts, |
| 39 | X86CPUUnquantizedExperts, |
| 40 | ) |
| 41 | |
| 42 | with ( |
| 43 | patch.object(current_platform, "is_cpu", return_value=True), |
| 44 | patch.object( |
| 45 | current_platform, |
| 46 | "get_cpu_architecture", |
| 47 | return_value=CpuArchEnum.X86, |
| 48 | ), |
| 49 | patch("torch.cpu._is_amx_tile_supported", return_value=amx_supported), |
| 50 | ): |
| 51 | moe_config = make_dummy_moe_config( |
| 52 | hidden_dim=128, |
| 53 | intermediate_size=128, |
| 54 | in_dtype=in_dtype, |
| 55 | ) |
| 56 | kernel_cls = next( |
| 57 | cls |
| 58 | for cls in backend_to_kernel_cls(UnquantizedMoeBackend.CPU) |
| 59 | if cls.is_supported_config( |
| 60 | cls, |
| 61 | moe_config, |
| 62 | None, |
| 63 | None, |
| 64 | CPUUnquantizedExperts.activation_format(), |
| 65 | )[0] |
| 66 | ) |
| 67 | |
| 68 | expected_kernel_cls = ( |
| 69 | X86CPUUnquantizedExperts if expect_amx_kernel else CPUUnquantizedExperts |
| 70 | ) |
| 71 | assert kernel_cls is expected_kernel_cls |
| 72 | |
| 73 | |
| 74 | @pytest.mark.parametrize( |
| 75 | ("platform", "in_dtype", "expect_arm_kernel"), |
| 76 | [ |
| 77 | ("linux", torch.bfloat16, True), |
| 78 | ("linux", torch.float16, False), |
| 79 | ("darwin", torch.bfloat16, False), |
| 80 | ], |
| 81 | ) |
| 82 | def test_arm_cpu_unquantized_kernel_selection( |
| 83 | platform: str, |
| 84 | in_dtype: torch.dtype, |
| 85 | expect_arm_kernel: bool, |
| 86 | ): |
| 87 | from vllm.model_executor.layers.fused_moe.experts.cpu_moe import ( |
| 88 | ArmCPUUnquantizedExperts, |
| 89 | CPUUnquantizedExperts, |
| 90 | ) |
| 91 | |
| 92 | with ( |
| 93 | patch.object(current_platform, "is_cpu", return_value=True), |
| 94 | patch.object( |
| 95 | current_platform, |
| 96 | "get_cpu_architecture", |
| 97 | return_value=CpuArchEnum.ARM, |
| 98 | ), |
| 99 | patch( |
| 100 | "vllm.model_executor.layers.fused_moe.experts.cpu_moe.sys.platform", |
| 101 | platform, |
| 102 | ), |
| 103 | ): |
| 104 | moe_config = make_dummy_moe_config( |
| 105 | hidden_dim=128, |
| 106 | intermediate_size=128, |
| 107 | in_dtype=in_dtype, |
| 108 | ) |
| 109 | kernel_cls = next( |
| 110 | cls |
| 111 | for cls in backend_to_kernel_cls(UnquantizedMoeBackend.CPU) |
| 112 | if cls.is_supported_config( |
| 113 | cls, |
| 114 | moe_config, |
| 115 | None, |
| 116 | None, |
| 117 | CPUUnquantizedExperts.activation_format(), |
| 118 | )[0] |
| 119 | ) |
| 120 | |
| 121 | expected_kernel_cls = ( |
| 122 | ArmCPUUnquantizedExperts if expect_arm_kernel else CPUUnquantizedExperts |
| 123 | ) |
| 124 | assert kernel_cls is expected_kernel_cls |
| 125 | |
| 126 | |
| 127 | @pytest.mark.parametrize( |
| 128 | "platform_method,expected_backend", |
| 129 | [ |
| 130 | ("is_cuda", UnquantizedMoeBackend.TRITON), # Default CUDA without FlashInfer |
| 131 | ("is_rocm", UnquantizedMoeBackend.TRITON), # ROCm without AITER |
| 132 | ("is_cpu", UnquantizedMoeBackend.CPU), |
| 133 | ("is_xpu", UnquantizedMoeBackend.XPU), |
| 134 | ("is_tpu", UnquantizedMoeBackend.TPU), |
| 135 | ("is_out_of_tree", UnquantizedMoeBackend.OOT), |
| 136 | ], |
| 137 | ) |
| 138 | @patch( |
| 139 | "vllm.utils.flashinfer.has_flashinfer", |
| 140 | return_value=False, |
| 141 | ) |
| 142 | @patch( |
| 143 | "vllm.model_executor.layers.fused_moe.oracle.unquantized.rocm_aiter_ops.is_fused_moe_enabled", |
| 144 | return_value=False, |
| 145 | ) |
| 146 | def test_select_default_backend_by_platform( |
| 147 | mock_aiter_enabled, |
| 148 | mock_has_flashinfer, |
| 149 | monkeypatch, |
| 150 | platform_method, |
| 151 | expected_backend, |
| 152 | ): |
| 153 | """Test default backend selection per platform with all optional |
| 154 | accelerators (FlashInfer, AITER) disabled.""" |
| 155 | with patch( |
| 156 | "vllm.model_executor.layers.fused_moe.oracle.unquantized.current_platform" |
| 157 | ) as mock_platform: |
| 158 | # Set all platform checks to False |
| 159 | mock_platform.is_cuda.return_value = False |
| 160 | mock_platform.is_rocm.return_value = False |
| 161 | mock_platform.is_cpu.return_value = False |
| 162 | mock_platform.is_xpu.return_value = False |
| 163 | mock_platform.is_tpu.return_value = False |
| 164 | mock_platform.is_out_of_tree.return_value = False |
| 165 | |
| 166 | # Set only the specified platform to True |
| 167 | getattr(mock_platform, platform_method).return_value = True |
| 168 | |
| 169 | with ( |
| 170 | patch.object(current_platform, "is_cuda", return_value=False), |
| 171 | patch.object(current_platform, "is_rocm", return_value=False), |
| 172 | patch.object(current_platform, "is_cpu", return_value=False), |
| 173 | patch.object(current_platform, "is_xpu", return_value=False), |
| 174 | patch.object(current_platform, "is_tpu", return_value=False), |
| 175 | patch.object(current_platform, "is_out_of_tree", return_value=False), |
| 176 | patch.object(current_platform, platform_method, return_value=True), |
| 177 | ): |
| 178 | # CPU's grouped-gemm kernels require hidden/intermediate sizes |
| 179 | # aligned to 32; the size-1 defaults only work for backends that |
| 180 | # don't check shapes at selection time. |
| 181 | moe_config = ( |
| 182 | make_dummy_moe_config(hidden_dim=128, intermediate_size=128) |
| 183 | if expected_backend == UnquantizedMoeBackend.CPU |
| 184 | else make_dummy_moe_config() |
| 185 | ) |
| 186 | selected_backend, expert_cls = select_unquantized_moe_backend( |
| 187 | moe_config=moe_config |
| 188 | ) |
| 189 | |
| 190 | assert selected_backend == expected_backend |
| 191 | if expected_backend in [ |
| 192 | UnquantizedMoeBackend.OOT, |
| 193 | UnquantizedMoeBackend.TPU, |
| 194 | ]: |
| 195 | assert expert_cls is None |
| 196 | else: |
| 197 | assert expert_cls is not None |
| 198 | |
| 199 | |
| 200 | @patch( |
| 201 | "vllm.utils.flashinfer.has_flashinfer", |
| 202 | return_value=False, |
| 203 | ) |
| 204 | @patch( |
| 205 | "vllm.model_executor.layers.fused_moe.oracle.unquantized.rocm_aiter_ops.is_fused_moe_enabled", |
| 206 | return_value=True, |
| 207 | ) |
| 208 | @pytest.mark.skipif( |
| 209 | not current_platform.is_rocm(), reason="ROCm-specific backend selection test" |
| 210 | ) |
| 211 | def test_select_rocm_aiter_backend(mock_aiter_enabled, mock_has_flashinfer): |
| 212 | """Test ROCm backend selection when AITER is available.""" |
| 213 | with patch( |
| 214 | "vllm.model_executor.layers.fused_moe.oracle.unquantized.current_platform" |
| 215 | ) as mock_platform: |
| 216 | mock_platform.is_cuda.return_value = False |
| 217 | mock_platform.is_rocm.return_value = True |
| 218 | mock_platform.is_cpu.return_value = False |
| 219 | mock_platform.is_xpu.return_value = False |
| 220 | mock_platform.is_tpu.return_value = False |
| 221 | mock_platform.is_out_of_tree.return_value = False |
| 222 | |
| 223 | moe_config = make_dummy_moe_config() |
| 224 | selected_backend, expert_cls = select_unquantized_moe_backend( |
| 225 | moe_config=moe_config, |
| 226 | ) |
| 227 | |
| 228 | assert selected_backend == UnquantizedMoeBackend.AITER |
| 229 | assert expert_cls is not None |
| 230 | |
| 231 | |
| 232 | @patch( |
| 233 | "vllm.utils.flashinfer.has_flashinfer", |
| 234 | return_value=False, |
| 235 | ) |
| 236 | @patch( |
| 237 | "vllm.model_executor.layers.fused_moe.oracle.unquantized.rocm_aiter_ops." |
| 238 | "is_fused_moe_enabled", |
| 239 | return_value=True, |
| 240 | ) |
| 241 | @patch( |
| 242 | "vllm.model_executor.layers.fused_moe.oracle.unquantized.rocm_aiter_ops." |
| 243 | "is_rdna_aiter_enabled", |
| 244 | return_value=False, |
| 245 | ) |
| 246 | @pytest.mark.skipif( |
| 247 | not current_platform.is_rocm(), reason="ROCm-specific backend selection test" |
| 248 | ) |
| 249 | def test_select_rocm_aiter_backend_non_gated_activation_falls_back( |
| 250 | mock_rdna_disabled, mock_aiter_enabled, mock_has_flashinfer, monkeypatch |
| 251 | ): |
| 252 | """Test ROCm backend selection falls back (not raises) for non-gated MoE.""" |
| 253 | monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1") |
| 254 | monkeypatch.setenv("VLLM_ROCM_USE_AITER_MOE", "1") |
| 255 | |
| 256 | with patch( |
| 257 | "vllm.model_executor.layers.fused_moe.oracle.unquantized.current_platform" |
| 258 | ) as mock_platform: |
| 259 | mock_platform.is_cuda.return_value = False |
| 260 | mock_platform.is_rocm.return_value = True |
| 261 | mock_platform.is_cpu.return_value = False |
| 262 | mock_platform.is_xpu.return_value = False |
| 263 | mock_platform.is_tpu.return_value = False |
| 264 | mock_platform.is_out_of_tree.return_value = False |
| 265 | |
| 266 | moe_config = make_dummy_moe_config(activation=MoEActivation.SILU_NO_MUL) |
| 267 | assert moe_config.is_act_and_mul is False |
| 268 | |
| 269 | selected_backend, expert_cls = select_unquantized_moe_backend( |
| 270 | moe_config=moe_config, |
| 271 | ) |
| 272 | |
| 273 | assert selected_backend != UnquantizedMoeBackend.AITER |
| 274 | assert expert_cls is not None |
| 275 | |
| 276 | |
| 277 | @pytest.mark.skipif( |
| 278 | not current_platform.is_rocm(), reason="ROCm-specific backend selection test" |
| 279 | ) |
| 280 | def test_explicit_aiter_backend_non_gated_activation_still_raises(): |
| 281 | """Explicit `--moe-backend aiter` still raises for non-gated MoE; |
| 282 | only the env-var opt-in path falls back.""" |
| 283 | moe_config = make_dummy_moe_config(activation=MoEActivation.SILU_NO_MUL) |
| 284 | moe_config.moe_backend = "aiter" # explicit pin, not "auto" |
| 285 | assert moe_config.is_act_and_mul is False |
| 286 | |
| 287 | with pytest.raises(ValueError): |
| 288 | select_unquantized_moe_backend(moe_config=moe_config) |
| 289 | |
| 290 | |
| 291 | @patch( |
| 292 | "vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsMonolithic.is_supported_config", |
| 293 | return_value=(True, None), |
| 294 | ) |
| 295 | @pytest.mark.skipif( |
| 296 | not current_platform.is_cuda(), reason="Only supported on NVIDIA platforms." |
| 297 | ) |
| 298 | def test_select_cuda_flashinfer_trtllm_backend( |
| 299 | mock_is_supported_trtllm_monolithic, |
| 300 | ): |
| 301 | """Test CUDA backend selection when FlashInfer TRTLLM is available and enabled.""" |
| 302 | with ( |
| 303 | patch.object(current_platform, "is_cuda", return_value=True), |
| 304 | patch.object(current_platform, "is_rocm", return_value=False), |
| 305 | patch.object(current_platform, "is_cpu", return_value=False), |
| 306 | patch.object(current_platform, "is_xpu", return_value=False), |
| 307 | patch.object(current_platform, "is_tpu", return_value=False), |
| 308 | patch.object(current_platform, "is_out_of_tree", return_value=False), |
| 309 | patch.object(current_platform, "has_device_capability", return_value=True), |
| 310 | ): |
| 311 | moe_config = make_dummy_moe_config() |
| 312 | moe_config.moe_backend = "flashinfer_trtllm" |
| 313 | # TRTLLM requires EP and does not support DP |
| 314 | moe_config.moe_parallel_config.use_ep = True |
| 315 | moe_config.moe_parallel_config.use_dp = False |
| 316 | |
| 317 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 318 | moe_config=moe_config |
| 319 | ) |
| 320 | |
| 321 | assert selected_backend == UnquantizedMoeBackend.FLASHINFER_TRTLLM |
| 322 | assert experts_cls is not None |
| 323 | assert experts_cls.__name__ == "TrtLlmBf16ExpertsMonolithic" |
| 324 | |
| 325 | |
| 326 | @patch( |
| 327 | "vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsMonolithic.is_supported_config", |
| 328 | return_value=(False, "monolithic unsupported"), |
| 329 | ) |
| 330 | @patch( |
| 331 | "vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsModular.is_supported_config", |
| 332 | return_value=(True, None), |
| 333 | ) |
| 334 | @pytest.mark.skipif( |
| 335 | not current_platform.is_cuda(), reason="Only supported on NVIDIA platforms." |
| 336 | ) |
| 337 | def test_select_cuda_flashinfer_trtllm_modular_backend( |
| 338 | mock_is_supported_trtllm_modular, |
| 339 | mock_is_supported_trtllm_monolithic, |
| 340 | ): |
| 341 | """Test CUDA backend selection falls back to FlashInfer TRTLLM modular.""" |
| 342 | with ( |
| 343 | patch.object(current_platform, "is_cuda", return_value=True), |
| 344 | patch.object(current_platform, "is_rocm", return_value=False), |
| 345 | patch.object(current_platform, "is_cpu", return_value=False), |
| 346 | patch.object(current_platform, "is_xpu", return_value=False), |
| 347 | patch.object(current_platform, "is_tpu", return_value=False), |
| 348 | patch.object(current_platform, "is_out_of_tree", return_value=False), |
| 349 | patch.object(current_platform, "has_device_capability", return_value=True), |
| 350 | ): |
| 351 | moe_config = make_dummy_moe_config() |
| 352 | moe_config.moe_backend = "flashinfer_trtllm" |
| 353 | moe_config.moe_parallel_config.use_ep = True |
| 354 | moe_config.moe_parallel_config.use_dp = False |
| 355 | |
| 356 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 357 | moe_config=moe_config |
| 358 | ) |
| 359 | |
| 360 | assert selected_backend == UnquantizedMoeBackend.FLASHINFER_TRTLLM |
| 361 | assert experts_cls is not None |
| 362 | assert experts_cls.__name__ == "TrtLlmBf16ExpertsModular" |
| 363 | |
| 364 | |
| 365 | @patch( |
| 366 | "vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsBase._supports_current_device", |
| 367 | return_value=True, |
| 368 | ) |
| 369 | @pytest.mark.parametrize( |
| 370 | "all2all_backend", |
| 371 | [ |
| 372 | "mori_high_throughput", |
| 373 | "mori_low_latency", |
| 374 | "flashinfer_nvlink_two_sided", |
| 375 | "flashinfer_nvlink_one_sided", |
| 376 | ], |
| 377 | ) |
| 378 | def test_select_cuda_flashinfer_trtllm_modular_for_standard_all2all( |
| 379 | mock_supports_current_device, |
| 380 | all2all_backend, |
| 381 | ): |
| 382 | """Test non-AG/RS standard-format all2all backends select modular BF16.""" |
| 383 | with ( |
| 384 | patch.object(current_platform, "is_cuda", return_value=True), |
| 385 | patch.object(current_platform, "is_rocm", return_value=False), |
| 386 | patch.object(current_platform, "is_cpu", return_value=False), |
| 387 | patch.object(current_platform, "is_xpu", return_value=False), |
| 388 | patch.object(current_platform, "is_tpu", return_value=False), |
| 389 | patch.object(current_platform, "is_out_of_tree", return_value=False), |
| 390 | patch.object( |
| 391 | current_platform, "is_device_capability_family", return_value=False |
| 392 | ), |
| 393 | ): |
| 394 | moe_config = make_dummy_moe_config(num_experts=4, num_local_experts=2) |
| 395 | moe_config.moe_backend = "flashinfer_trtllm" |
| 396 | moe_config.moe_parallel_config.use_ep = True |
| 397 | moe_config.moe_parallel_config.dp_size = 2 |
| 398 | moe_config.moe_parallel_config.ep_size = 2 |
| 399 | moe_config.moe_parallel_config.all2all_backend = all2all_backend |
| 400 | |
| 401 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 402 | moe_config=moe_config |
| 403 | ) |
| 404 | |
| 405 | assert selected_backend == UnquantizedMoeBackend.FLASHINFER_TRTLLM |
| 406 | assert experts_cls is not None |
| 407 | assert experts_cls.__name__ == "TrtLlmBf16ExpertsModular" |
| 408 | |
| 409 | |
| 410 | @patch( |
| 411 | "vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsBase._supports_current_device", |
| 412 | return_value=True, |
| 413 | ) |
| 414 | def test_select_cuda_deepep_ht_falls_back_from_trtllm( |
| 415 | mock_supports_current_device, |
| 416 | ): |
| 417 | """Test DeepEP HT avoids the unsupported BF16 TRTLLM modular path.""" |
| 418 | with ( |
| 419 | patch.object(current_platform, "is_cuda", return_value=True), |
| 420 | patch.object(current_platform, "is_rocm", return_value=False), |
| 421 | patch.object(current_platform, "is_cpu", return_value=False), |
| 422 | patch.object(current_platform, "is_xpu", return_value=False), |
| 423 | patch.object(current_platform, "is_tpu", return_value=False), |
| 424 | patch.object(current_platform, "is_out_of_tree", return_value=False), |
| 425 | patch.object( |
| 426 | current_platform, "is_device_capability_family", return_value=False |
| 427 | ), |
| 428 | ): |
| 429 | moe_config = make_dummy_moe_config(num_experts=4, num_local_experts=2) |
| 430 | moe_config.moe_backend = "auto" |
| 431 | moe_config.moe_parallel_config.use_ep = True |
| 432 | moe_config.moe_parallel_config.dp_size = 2 |
| 433 | moe_config.moe_parallel_config.ep_size = 2 |
| 434 | moe_config.moe_parallel_config.all2all_backend = "deepep_high_throughput" |
| 435 | |
| 436 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 437 | moe_config=moe_config |
| 438 | ) |
| 439 | |
| 440 | assert selected_backend == UnquantizedMoeBackend.TRITON |
| 441 | assert experts_cls is not None |
| 442 | assert experts_cls.__name__ == "TritonExperts" |
| 443 | |
| 444 | |
| 445 | @patch( |
| 446 | "vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsBase._supports_current_device", |
| 447 | return_value=True, |
| 448 | ) |
| 449 | def test_select_cuda_flashinfer_trtllm_ag_rs_uses_monolithic( |
| 450 | mock_supports_current_device, |
| 451 | ): |
| 452 | """Test AG/RS stays on BF16 TRTLLM monolithic when TRTLLM is supported.""" |
| 453 | with ( |
| 454 | patch.object(current_platform, "is_cuda", return_value=True), |
| 455 | patch.object(current_platform, "is_rocm", return_value=False), |
| 456 | patch.object(current_platform, "is_cpu", return_value=False), |
| 457 | patch.object(current_platform, "is_xpu", return_value=False), |
| 458 | patch.object(current_platform, "is_tpu", return_value=False), |
| 459 | patch.object(current_platform, "is_out_of_tree", return_value=False), |
| 460 | patch.object( |
| 461 | current_platform, "is_device_capability_family", return_value=False |
| 462 | ), |
| 463 | ): |
| 464 | moe_config = make_dummy_moe_config(num_experts=4, num_local_experts=2) |
| 465 | moe_config.moe_backend = "flashinfer_trtllm" |
| 466 | moe_config.routing_method = RoutingMethodType.Renormalize |
| 467 | moe_config.moe_parallel_config.use_ep = True |
| 468 | moe_config.moe_parallel_config.dp_size = 2 |
| 469 | moe_config.moe_parallel_config.ep_size = 2 |
| 470 | moe_config.moe_parallel_config.all2all_backend = "allgather_reducescatter" |
| 471 | |
| 472 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 473 | moe_config=moe_config |
| 474 | ) |
| 475 | |
| 476 | assert selected_backend == UnquantizedMoeBackend.FLASHINFER_TRTLLM |
| 477 | assert experts_cls is not None |
| 478 | assert experts_cls.__name__ == "TrtLlmBf16ExpertsMonolithic" |
| 479 | |
| 480 | |
| 481 | @patch( |
| 482 | "vllm.utils.flashinfer.has_flashinfer", |
| 483 | return_value=True, |
| 484 | ) |
| 485 | @patch( |
| 486 | "vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsMonolithic.is_supported_config", |
| 487 | return_value=(False, None), |
| 488 | ) |
| 489 | @patch( |
| 490 | "vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe.TrtLlmBf16ExpertsModular.is_supported_config", |
| 491 | return_value=(False, None), |
| 492 | ) |
| 493 | @patch( |
| 494 | "vllm.model_executor.layers.fused_moe.experts.flashinfer_cutlass_moe.FlashInferExperts.is_supported_config", |
| 495 | return_value=(True, None), |
| 496 | ) |
| 497 | @pytest.mark.skipif( |
| 498 | not current_platform.is_cuda(), reason="Only supported on NVIDIA platforms." |
| 499 | ) |
| 500 | def test_select_cuda_flashinfer_cutlass_backend( |
| 501 | mock_is_supported_cutlass, |
| 502 | mock_is_supported_trtllm_modular, |
| 503 | mock_is_supported_trtllm_monolithic, |
| 504 | mock_has_flashinfer, |
| 505 | ): |
| 506 | """Test CUDA backend selection when FlashInfer TRTLLM is not available |
| 507 | and FlashInfer CUTLASS is available.""" |
| 508 | with ( |
| 509 | patch.object(current_platform, "is_cuda", return_value=True), |
| 510 | patch.object(current_platform, "is_rocm", return_value=False), |
| 511 | patch.object(current_platform, "is_cpu", return_value=False), |
| 512 | patch.object(current_platform, "is_xpu", return_value=False), |
| 513 | patch.object(current_platform, "is_tpu", return_value=False), |
| 514 | patch.object(current_platform, "is_out_of_tree", return_value=False), |
| 515 | patch.object(current_platform, "has_device_capability", return_value=True), |
| 516 | ): |
| 517 | moe_config = make_dummy_moe_config() |
| 518 | # Select FlashInfer CUTLASS explicitly |
| 519 | moe_config.moe_backend = "flashinfer_cutlass" |
| 520 | # CUTLASS requires EP and does not support DP |
| 521 | moe_config.moe_parallel_config.use_ep = True |
| 522 | moe_config.moe_parallel_config.use_dp = False |
| 523 | |
| 524 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 525 | moe_config=moe_config |
| 526 | ) |
| 527 | |
| 528 | assert selected_backend == UnquantizedMoeBackend.FLASHINFER_CUTLASS |
| 529 | assert experts_cls is not None |
| 530 | |
| 531 | |
| 532 | @skipif_not_cuda_rocm |
| 533 | def test_select_lora_backend_prefers_triton(): |
| 534 | """LoRA-enabled unquantized MoE should select Triton backend.""" |
| 535 | moe_config = make_dummy_moe_config() |
| 536 | moe_config.is_lora_enabled = True |
| 537 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 538 | moe_config=moe_config |
| 539 | ) |
| 540 | |
| 541 | assert selected_backend == UnquantizedMoeBackend.TRITON |
| 542 | assert experts_cls is not None |
| 543 | |
| 544 | |
| 545 | @skipif_not_cuda_rocm |
| 546 | def test_select_lora_explicit_non_triton_backend(): |
| 547 | """LoRA should override explicit non-Triton backend to Triton.""" |
| 548 | moe_config = make_dummy_moe_config() |
| 549 | moe_config.is_lora_enabled = True |
| 550 | |
| 551 | # Use string from mapping in function map_unquantized_backend() |
| 552 | moe_config.moe_backend = "flashinfer_cutlass" |
| 553 | |
| 554 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 555 | moe_config=moe_config |
| 556 | ) |
| 557 | |
| 558 | assert selected_backend == UnquantizedMoeBackend.TRITON |
| 559 | assert experts_cls is not None |
| 560 | |
| 561 | |
| 562 | @skipif_not_cuda_rocm |
| 563 | @pytest.mark.parametrize("is_lora_enabled", [False, True]) |
| 564 | def test_select_explicit_triton_backend(is_lora_enabled): |
| 565 | """Explicit triton backend selection should return Triton.""" |
| 566 | moe_config = make_dummy_moe_config() |
| 567 | moe_config.is_lora_enabled = is_lora_enabled |
| 568 | moe_config.moe_backend = "triton" |
| 569 | |
| 570 | selected_backend, experts_cls = select_unquantized_moe_backend( |
| 571 | moe_config=moe_config |
| 572 | ) |
| 573 | |
| 574 | assert selected_backend == UnquantizedMoeBackend.TRITON |
| 575 | assert experts_cls is not None |