| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import copy |
| 4 | import dataclasses |
| 5 | from collections.abc import Callable |
| 6 | from math import prod |
| 7 | |
| 8 | import pytest |
| 9 | import torch |
| 10 | |
| 11 | import vllm.model_executor.layers.fused_moe.modular_kernel as mk |
| 12 | from tests.kernels.moe.utils import make_dummy_moe_config |
| 13 | from vllm import _custom_ops as ops |
| 14 | from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config |
| 15 | from vllm.model_executor.layers.fused_moe import ( |
| 16 | ApplyMoEActivationConfig, |
| 17 | fused_experts, |
| 18 | fused_topk, |
| 19 | ) |
| 20 | from vllm.model_executor.layers.fused_moe.activation import ( |
| 21 | MoEActivation, |
| 22 | apply_moe_activation_supported, |
| 23 | ) |
| 24 | from vllm.model_executor.layers.fused_moe.all2all_utils import ( |
| 25 | maybe_make_prepare_finalize, |
| 26 | ) |
| 27 | from vllm.model_executor.layers.fused_moe.config import ( |
| 28 | FUSED_MOE_UNQUANTIZED_CONFIG, |
| 29 | FusedMoEQuantConfig, |
| 30 | fp8_w8a8_moe_quant_config, |
| 31 | ) |
| 32 | from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import ( |
| 33 | CutlassExpertsFp4, |
| 34 | CutlassExpertsFp8, |
| 35 | CutlassExpertsMxfp4, |
| 36 | CutlassExpertsW4A8Fp8, |
| 37 | run_cutlass_moe_fp4, |
| 38 | run_cutlass_moe_fp8, |
| 39 | run_cutlass_moe_mxfp4, |
| 40 | ) |
| 41 | from vllm.model_executor.layers.fused_moe.oracle import nvfp4 as nvfp4_oracle |
| 42 | from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input |
| 43 | from vllm.platforms import current_platform |
| 44 | from vllm.utils.torch_utils import set_random_seed |
| 45 | |
| 46 | NUM_EXPERTS = [40, 64] |
| 47 | TOP_KS = [6, 8] |
| 48 | |
| 49 | MNK_FACTORS = [ |
| 50 | (2, 1024, 1024), |
| 51 | (2, 3072, 1024), |
| 52 | (2, 3072, 1536), |
| 53 | (7, 3072, 1536), |
| 54 | (64, 1024, 1024), |
| 55 | (64, 1024, 1536), |
| 56 | (64, 3072, 1024), |
| 57 | (224, 1024, 1024), |
| 58 | (224, 3072, 1024), |
| 59 | (224, 3072, 1536), |
| 60 | (32768, 1024, 1024), |
| 61 | # These sizes trigger wrong answers. |
| 62 | # (7232, 2048, 5120), |
| 63 | # (40000, 2048, 5120), |
| 64 | ] |
| 65 | |
| 66 | vllm_config = VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) |
| 67 | |
| 68 | |
| 69 | @pytest.mark.parametrize( |
| 70 | "experts_cls", |
| 71 | [ |
| 72 | CutlassExpertsFp8, |
| 73 | CutlassExpertsFp4, |
| 74 | CutlassExpertsMxfp4, |
| 75 | CutlassExpertsW4A8Fp8, |
| 76 | ], |
| 77 | ) |
| 78 | @pytest.mark.parametrize("activation", list(MoEActivation)) |
| 79 | def test_cutlass_moe_activation_metadata_tracks_shared_apply( |
| 80 | experts_cls: type[mk.FusedMoEExperts], activation: MoEActivation |
| 81 | ): |
| 82 | supports_shape = experts_cls._supports_no_act_and_mul() or activation.is_gated |
| 83 | expected = supports_shape and apply_moe_activation_supported(activation) |
| 84 | |
| 85 | assert experts_cls._supports_activation(activation) == expected |
| 86 | |
| 87 | |
| 88 | def test_cutlass_moe_forwards_shared_activation_parameters(): |
| 89 | moe_config = make_dummy_moe_config() |
| 90 | moe_config.swiglu_limit = 7.0 |
| 91 | moe_config.swiglu_alpha = 1.5 |
| 92 | moe_config.swiglu_beta = 0.25 |
| 93 | moe_config.activation_situ_beta = 2.0 |
| 94 | moe_config.activation_situ_linear_beta = 3.0 |
| 95 | quant_config = FusedMoEQuantConfig.make(gemm1_alpha=1.75) |
| 96 | |
| 97 | assert ApplyMoEActivationConfig.from_configs( |
| 98 | moe_config, quant_config |
| 99 | ) == ApplyMoEActivationConfig( |
| 100 | clamp_limit=7.0, |
| 101 | alpha=1.75, |
| 102 | beta=0.25, |
| 103 | activation_situ_beta=2.0, |
| 104 | activation_situ_linear_beta=3.0, |
| 105 | ) |
| 106 | |
| 107 | |
| 108 | @pytest.mark.parametrize( |
| 109 | ("backend", "expected"), |
| 110 | [ |
| 111 | ("cutlass", nvfp4_oracle.NvFp4MoeBackend.VLLM_CUTLASS), |
| 112 | ("humming", nvfp4_oracle.NvFp4MoeBackend.HUMMING), |
| 113 | ], |
| 114 | ) |
| 115 | def test_nvfp4_clamp_allows_shared_activation_backends( |
| 116 | monkeypatch, backend: str, expected: nvfp4_oracle.NvFp4MoeBackend |
| 117 | ): |
| 118 | class SupportedExperts: |
| 119 | @staticmethod |
| 120 | def is_supported_config(*args, **kwargs): |
| 121 | return True, None |
| 122 | |
| 123 | monkeypatch.setattr( |
| 124 | nvfp4_oracle, "backend_to_kernel_cls", lambda backend: [SupportedExperts] |
| 125 | ) |
| 126 | moe_config = make_dummy_moe_config() |
| 127 | moe_config.moe_backend = backend |
| 128 | moe_config.swiglu_limit = 7.0 |
| 129 | |
| 130 | selected, _ = nvfp4_oracle.select_nvfp4_moe_backend( |
| 131 | moe_config, weight_key=None, activation_key=None |
| 132 | ) |
| 133 | |
| 134 | assert selected == expected |
| 135 | |
| 136 | |
| 137 | @pytest.mark.skipif( |
| 138 | (lambda x: x is None or not ops.cutlass_group_gemm_supported(x.to_int()))( |
| 139 | current_platform.get_device_capability() |
| 140 | ), |
| 141 | reason="Grouped gemm is not supported on this GPU type.", |
| 142 | ) |
| 143 | def test_cutlass_moe_permutation_maps_padding_to_zero(): |
| 144 | """Invalid top-k routes must not leave unsafe permutation-map entries.""" |
| 145 | num_experts = 2 |
| 146 | topk_ids = torch.tensor( |
| 147 | [[0, -1], [1, num_experts]], device="cuda", dtype=torch.int32 |
| 148 | ) |
| 149 | num_routes = topk_ids.numel() |
| 150 | |
| 151 | expert_offsets = torch.empty(num_experts + 1, device="cuda", dtype=torch.int32) |
| 152 | blockscale_offsets = torch.empty_like(expert_offsets) |
| 153 | problem_sizes1 = torch.empty(num_experts, 3, device="cuda", dtype=torch.int32) |
| 154 | problem_sizes2 = torch.empty_like(problem_sizes1) |
| 155 | input_permutation = torch.empty(num_routes, device="cuda", dtype=torch.int32) |
| 156 | output_permutation = torch.empty_like(input_permutation) |
| 157 | |
| 158 | ops.get_cutlass_moe_mm_data( |
| 159 | topk_ids, |
| 160 | expert_offsets, |
| 161 | problem_sizes1, |
| 162 | problem_sizes2, |
| 163 | input_permutation, |
| 164 | output_permutation, |
| 165 | num_experts, |
| 166 | 128, |
| 167 | 128, |
| 168 | blockscale_offsets, |
| 169 | ) |
| 170 | |
| 171 | input_tensor = torch.randn(2, 128, device="cuda", dtype=torch.bfloat16) |
| 172 | permuted = ops.shuffle_rows(input_tensor, input_permutation) |
| 173 | restored = ops.shuffle_rows(permuted, output_permutation) |
| 174 | |
| 175 | torch.testing.assert_close(permuted[:2], input_tensor) |
| 176 | torch.testing.assert_close(permuted[2:], torch.zeros_like(permuted[2:])) |
| 177 | torch.testing.assert_close(restored[0], input_tensor[0]) |
| 178 | torch.testing.assert_close(restored[1], torch.zeros_like(restored[1])) |
| 179 | torch.testing.assert_close(restored[2], input_tensor[1]) |
| 180 | torch.testing.assert_close(restored[3], torch.zeros_like(restored[3])) |
| 181 | |
| 182 | |
| 183 | @pytest.mark.parametrize("quantization", ["nvfp4", "mxfp4"]) |
| 184 | @torch.inference_mode() |
| 185 | def test_cutlass_fp4_moe_padded_routes_do_not_change_valid_output(quantization: str): |
| 186 | """Padded routes must not participate in either activation quantization.""" |
| 187 | experts_cls = CutlassExpertsFp4 if quantization == "nvfp4" else CutlassExpertsMxfp4 |
| 188 | if not experts_cls._supports_current_device(): |
| 189 | pytest.skip(f"CUTLASS {quantization} MoE is not supported on this GPU") |
| 190 | |
| 191 | set_random_seed(7) |
| 192 | num_experts, n, k = 4, 128, 128 |
| 193 | dtype = torch.bfloat16 |
| 194 | device = torch.device("cuda") |
| 195 | w1 = torch.full( |
| 196 | (num_experts, 2 * n, k // 2), 0x11, device=device, dtype=torch.uint8 |
| 197 | ) |
| 198 | w2 = torch.full((num_experts, k, n // 2), 0x11, device=device, dtype=torch.uint8) |
| 199 | |
| 200 | runner: Callable[..., None] |
| 201 | quant_args: tuple[torch.Tensor, ...] |
| 202 | if quantization == "nvfp4": |
| 203 | scale_kwargs = { |
| 204 | "device": device, |
| 205 | "dtype": torch.float8_e4m3fn, |
| 206 | } |
| 207 | w1_scale = torch.full((num_experts, 2 * n, k // 16), 1 / 16, **scale_kwargs) |
| 208 | w2_scale = torch.full((num_experts, k, n // 16), 1 / 16, **scale_kwargs) |
| 209 | gscale = torch.ones(num_experts, device=device, dtype=torch.float32) |
| 210 | runner = run_cutlass_moe_fp4 |
| 211 | quant_args = ( |
| 212 | gscale, |
| 213 | w1, |
| 214 | w1_scale, |
| 215 | gscale, |
| 216 | gscale, |
| 217 | w2, |
| 218 | w2_scale, |
| 219 | gscale, |
| 220 | ) |
| 221 | else: |
| 222 | scale_exponent = 127 - 4 |
| 223 | w1_scale = torch.full( |
| 224 | (num_experts, 2 * n, k // 32), |
| 225 | scale_exponent, |
| 226 | device=device, |
| 227 | dtype=torch.uint8, |
| 228 | ) |
| 229 | w2_scale = torch.full( |
| 230 | (num_experts, k, n // 32), |
| 231 | scale_exponent, |
| 232 | device=device, |
| 233 | dtype=torch.uint8, |
| 234 | ) |
| 235 | runner = run_cutlass_moe_mxfp4 |
| 236 | quant_args = (w1, w1_scale, w2, w2_scale) |
| 237 | |
| 238 | def run_moe( |
| 239 | hidden_states: torch.Tensor, |
| 240 | topk_ids: torch.Tensor, |
| 241 | topk_weights: torch.Tensor, |
| 242 | workspace_value: float, |
| 243 | ) -> torch.Tensor: |
| 244 | m, topk = topk_ids.shape |
| 245 | output = torch.empty((m, k), device=device, dtype=dtype) |
| 246 | workspace13 = torch.full( |
| 247 | (m * topk, max(2 * n, k)), |
| 248 | workspace_value, |
| 249 | device=device, |
| 250 | dtype=dtype, |
| 251 | ) |
| 252 | workspace2 = torch.full( |
| 253 | (m * topk, n), workspace_value, device=device, dtype=dtype |
| 254 | ) |
| 255 | |
| 256 | runner( |
| 257 | output, |
| 258 | hidden_states, |
| 259 | *quant_args, |
| 260 | topk_weights, |
| 261 | topk_ids, |
| 262 | MoEActivation.SILU, |
| 263 | workspace13, |
| 264 | workspace2, |
| 265 | m, |
| 266 | n, |
| 267 | k, |
| 268 | num_experts, |
| 269 | device, |
| 270 | ) |
| 271 | return output |
| 272 | |
| 273 | valid_input = torch.randn((1, k), device=device, dtype=dtype) |
| 274 | expected = run_moe( |
| 275 | valid_input, |
| 276 | torch.zeros((1, 1), device=device, dtype=torch.int32), |
| 277 | torch.ones((1, 1), device=device, dtype=torch.float32), |
| 278 | workspace_value=0, |
| 279 | ) |
| 280 | assert expected.abs().max() > 0 |
| 281 | |
| 282 | # Force the quantizers' grid-stride loops to revisit padded tail rows. |
| 283 | num_tokens = 4096 |
| 284 | padded_input = torch.zeros((num_tokens, k), device=device, dtype=dtype) |
| 285 | padded_input[0] = valid_input[0] |
| 286 | padded_ids = torch.full((num_tokens, 2), -1, device=device, dtype=torch.int32) |
| 287 | padded_ids[0, 0] = 0 |
| 288 | padded_weights = torch.zeros((num_tokens, 2), device=device, dtype=torch.float32) |
| 289 | padded_weights[0, 0] = 1 |
| 290 | |
| 291 | for workspace_value in (0, 64): |
| 292 | actual = run_moe(padded_input, padded_ids, padded_weights, workspace_value) |
| 293 | torch.testing.assert_close(actual[0], expected[0], atol=1e-2, rtol=1e-2) |
| 294 | torch.testing.assert_close(actual[1:], torch.zeros_like(actual[1:])) |
| 295 | |
| 296 | |
| 297 | @dataclasses.dataclass |
| 298 | class MOETensors: |
| 299 | a: torch.Tensor |
| 300 | w1: torch.Tensor |
| 301 | w2: torch.Tensor |
| 302 | ab_strides1: torch.Tensor |
| 303 | c_strides1: torch.Tensor |
| 304 | ab_strides2: torch.Tensor |
| 305 | c_strides2: torch.Tensor |
| 306 | |
| 307 | @staticmethod |
| 308 | def make_moe_tensors( |
| 309 | m: int, k: int, n: int, e: int, dtype: torch.dtype |
| 310 | ) -> "MOETensors": |
| 311 | a = torch.randn((m, k), device="cuda", dtype=dtype) / 10 |
| 312 | w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10 |
| 313 | w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10 |
| 314 | ab_strides1 = torch.full((e,), k, device="cuda", dtype=torch.int64) |
| 315 | c_strides1 = torch.full((e,), 2 * n, device="cuda", dtype=torch.int64) |
| 316 | ab_strides2 = torch.full((e,), n, device="cuda", dtype=torch.int64) |
| 317 | c_strides2 = torch.full((e,), k, device="cuda", dtype=torch.int64) |
| 318 | return MOETensors( |
| 319 | a=a, |
| 320 | w1=w1, |
| 321 | w2=w2, |
| 322 | ab_strides1=ab_strides1, |
| 323 | c_strides1=c_strides1, |
| 324 | ab_strides2=ab_strides2, |
| 325 | c_strides2=c_strides2, |
| 326 | ) |
| 327 | |
| 328 | |
| 329 | @dataclasses.dataclass |
| 330 | class MOETensors8Bit(MOETensors): |
| 331 | # quantized |
| 332 | a_q: torch.Tensor | None = None # a -> a_q |
| 333 | w1_q: torch.Tensor | None = None # w1 -> w1_q |
| 334 | w2_q: torch.Tensor | None = None # w2 -> w2_q |
| 335 | a_scale: torch.Tensor | None = None |
| 336 | w1_scale: torch.Tensor | None = None |
| 337 | w2_scale: torch.Tensor | None = None |
| 338 | # dequantized |
| 339 | a_d: torch.Tensor | None = None # a -> a_q -> a_d |
| 340 | w1_d: torch.Tensor | None = None # w1 -> w1_q -> w1_d |
| 341 | w2_d: torch.Tensor | None = None # w2 -> w2_q -> w2_d |
| 342 | |
| 343 | @staticmethod |
| 344 | def make_moe_tensors_8bit( |
| 345 | m: int, k: int, n: int, e: int, per_act_token: bool, per_out_channel: bool |
| 346 | ) -> "MOETensors8Bit": |
| 347 | dtype = torch.half |
| 348 | q_dtype = torch.float8_e4m3fn |
| 349 | |
| 350 | moe_tensors_fp16 = MOETensors.make_moe_tensors(m, k, n, e, dtype) |
| 351 | |
| 352 | # a -> a_q, w1 -> w1_q, w2 -> w2_q |
| 353 | n_b_scales = 2 * n if per_out_channel else 1 |
| 354 | k_b_scales = k if per_out_channel else 1 |
| 355 | # Get the right scale for tests. |
| 356 | a_q, a_scale = ops.scaled_fp8_quant( |
| 357 | moe_tensors_fp16.a, None, use_per_token_if_dynamic=per_act_token |
| 358 | ) |
| 359 | |
| 360 | w1_q = torch.empty((e, 2 * n, k), device="cuda", dtype=q_dtype) |
| 361 | w2_q = torch.empty((e, k, n), device="cuda", dtype=q_dtype) |
| 362 | |
| 363 | w1_scale = torch.empty((e, n_b_scales, 1), device="cuda", dtype=torch.float32) |
| 364 | w2_scale = torch.empty((e, k_b_scales, 1), device="cuda", dtype=torch.float32) |
| 365 | for expert in range(e): |
| 366 | w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant( |
| 367 | moe_tensors_fp16.w1[expert], use_per_token_if_dynamic=per_out_channel |
| 368 | ) |
| 369 | w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant( |
| 370 | moe_tensors_fp16.w2[expert], use_per_token_if_dynamic=per_out_channel |
| 371 | ) |
| 372 | |
| 373 | # a_q -> a_d, w1_q -> w1_d, w2_q -> w2_d |
| 374 | a_d = a_q.float().mul(a_scale).to(dtype) |
| 375 | w1_d = torch.empty_like(moe_tensors_fp16.w1) |
| 376 | w2_d = torch.empty_like(moe_tensors_fp16.w2) |
| 377 | for expert in range(e): |
| 378 | w1_d[expert] = (w1_q[expert].float() * w1_scale[expert]).half() |
| 379 | w2_d[expert] = (w2_q[expert].float() * w2_scale[expert]).half() |
| 380 | |
| 381 | return MOETensors8Bit( |
| 382 | a=moe_tensors_fp16.a, |
| 383 | w1=moe_tensors_fp16.w1, |
| 384 | w2=moe_tensors_fp16.w2, |
| 385 | ab_strides1=moe_tensors_fp16.ab_strides1, |
| 386 | c_strides1=moe_tensors_fp16.c_strides1, |
| 387 | ab_strides2=moe_tensors_fp16.ab_strides2, |
| 388 | c_strides2=moe_tensors_fp16.c_strides2, |
| 389 | a_q=a_q, |
| 390 | w1_q=w1_q, |
| 391 | w2_q=w2_q, |
| 392 | a_scale=a_scale, |
| 393 | w1_scale=w1_scale, |
| 394 | w2_scale=w2_scale, |
| 395 | a_d=a_d, |
| 396 | w1_d=w1_d, |
| 397 | w2_d=w2_d, |
| 398 | ) |
| 399 | |
| 400 | |
| 401 | def run_with_expert_maps( |
| 402 | num_experts: int, |
| 403 | num_local_experts: int, |
| 404 | quant_config: FusedMoEQuantConfig, |
| 405 | **cutlass_moe_kwargs, |
| 406 | ): |
| 407 | def slice_experts(): |
| 408 | slice_params = [ |
| 409 | "w1", |
| 410 | "w2", |
| 411 | ] |
| 412 | full_tensors = { |
| 413 | k: v |
| 414 | for k, v in cutlass_moe_kwargs.items() |
| 415 | if k in slice_params and k in cutlass_moe_kwargs |
| 416 | } |
| 417 | |
| 418 | for i in range(0, num_experts, num_local_experts): |
| 419 | s, e = i, i + num_local_experts |
| 420 | |
| 421 | # make expert map |
| 422 | expert_map = [-1] * num_experts |
| 423 | expert_map[s:e] = list(range(num_local_experts)) |
| 424 | expert_map = torch.tensor(expert_map, dtype=torch.int32, device="cuda") |
| 425 | |
| 426 | # update cutlass moe arg with expert_map |
| 427 | cutlass_moe_kwargs["expert_map"] = expert_map |
| 428 | # update cutlass moe arg tensors |
| 429 | for k, t in full_tensors.items(): |
| 430 | cutlass_moe_kwargs[k] = t[s:e] |
| 431 | |
| 432 | new_quant_config = copy.deepcopy(quant_config) |
| 433 | new_quant_config._w1.scale = quant_config.w1_scale[s:e] |
| 434 | new_quant_config._w2.scale = quant_config.w2_scale[s:e] |
| 435 | |
| 436 | yield cutlass_moe_kwargs, new_quant_config |
| 437 | |
| 438 | out_tensor = torch.zeros_like(cutlass_moe_kwargs["hidden_states"]) |
| 439 | for kwargs, new_quant_config in slice_experts(): |
| 440 | w2 = kwargs["w2"] |
| 441 | a = kwargs["hidden_states"] |
| 442 | moe_config = make_dummy_moe_config( |
| 443 | max_num_tokens=kwargs.get("hidden_states").shape[0], |
| 444 | experts_per_token=kwargs.get("topk_ids").shape[1], |
| 445 | num_experts=num_experts, |
| 446 | num_local_experts=num_local_experts, |
| 447 | hidden_dim=w2.shape[1], |
| 448 | intermediate_size=w2.shape[2], |
| 449 | in_dtype=a.dtype, |
| 450 | ) |
| 451 | kernel = mk.FusedMoEKernel( |
| 452 | maybe_make_prepare_finalize( |
| 453 | moe=moe_config, |
| 454 | quant_config=new_quant_config, |
| 455 | allow_new_interface=True, |
| 456 | use_monolithic=False, |
| 457 | ), |
| 458 | CutlassExpertsFp8( |
| 459 | moe_config=moe_config, |
| 460 | quant_config=new_quant_config, |
| 461 | ), |
| 462 | ) |
| 463 | out_tensor = out_tensor + kernel.apply(**kwargs) |
| 464 | |
| 465 | return out_tensor |
| 466 | |
| 467 | |
| 468 | def run_8_bit( |
| 469 | moe_tensors: MOETensors8Bit, |
| 470 | topk_weights: torch.Tensor, |
| 471 | topk_ids: torch.Tensor, |
| 472 | per_act_token: bool, |
| 473 | per_out_ch: bool, |
| 474 | num_local_experts: int | None = None, |
| 475 | ) -> torch.Tensor: |
| 476 | assert not any( |
| 477 | [ |
| 478 | t is None |
| 479 | for t in [ |
| 480 | moe_tensors.w1_q, |
| 481 | moe_tensors.w2_q, |
| 482 | moe_tensors.w1_scale, |
| 483 | moe_tensors.w2_scale, |
| 484 | moe_tensors.a_scale, |
| 485 | ] |
| 486 | ] |
| 487 | ) |
| 488 | |
| 489 | quant_config = fp8_w8a8_moe_quant_config( |
| 490 | w1_scale=moe_tensors.w1_scale, |
| 491 | w2_scale=moe_tensors.w2_scale, |
| 492 | per_act_token_quant=per_act_token, |
| 493 | per_out_ch_quant=per_out_ch, |
| 494 | # Set to moe_tensors.a_scale iff static scales + per tensor. |
| 495 | # This is not currently being tested. |
| 496 | a1_scale=None, |
| 497 | ) |
| 498 | |
| 499 | num_experts = moe_tensors.w1.size(0) # type: ignore[attr-defined] |
| 500 | with_ep = num_local_experts is not None or num_local_experts == num_experts |
| 501 | |
| 502 | kwargs = { |
| 503 | "hidden_states": moe_tensors.a, |
| 504 | "w1": moe_tensors.w1_q, # type: ignore[union-attr] |
| 505 | "w2": moe_tensors.w2_q, # type: ignore[union-attr] |
| 506 | "topk_weights": topk_weights, |
| 507 | "topk_ids": topk_ids, |
| 508 | "global_num_experts": num_experts, |
| 509 | "activation": MoEActivation.SILU, |
| 510 | "expert_map": None, |
| 511 | "apply_router_weight_on_input": False, |
| 512 | } |
| 513 | |
| 514 | if not with_ep: |
| 515 | moe_config = make_dummy_moe_config( |
| 516 | max_num_tokens=moe_tensors.a.shape[0], |
| 517 | experts_per_token=topk_ids.shape[1], |
| 518 | num_experts=num_experts, |
| 519 | num_local_experts=num_local_experts, |
| 520 | hidden_dim=moe_tensors.w2_q.shape[1], # type: ignore[union-attr] |
| 521 | intermediate_size=moe_tensors.w2_q.shape[2], # type: ignore[union-attr] |
| 522 | in_dtype=moe_tensors.a.dtype, |
| 523 | ) |
| 524 | kernel = mk.FusedMoEKernel( |
| 525 | maybe_make_prepare_finalize( |
| 526 | moe=moe_config, |
| 527 | quant_config=quant_config, |
| 528 | allow_new_interface=True, |
| 529 | use_monolithic=False, |
| 530 | ), |
| 531 | CutlassExpertsFp8( |
| 532 | moe_config=moe_config, |
| 533 | quant_config=quant_config, |
| 534 | ), |
| 535 | ) |
| 536 | return kernel.apply(**kwargs) |
| 537 | |
| 538 | assert num_local_experts is not None |
| 539 | return run_with_expert_maps( |
| 540 | num_experts, |
| 541 | num_local_experts, |
| 542 | quant_config, |
| 543 | **kwargs, |
| 544 | ) |
| 545 | |
| 546 | |
| 547 | @pytest.mark.parametrize("m,n,k", MNK_FACTORS) |
| 548 | @pytest.mark.parametrize("e", NUM_EXPERTS) |
| 549 | @pytest.mark.parametrize("topk", TOP_KS) |
| 550 | @pytest.mark.parametrize("per_act_token", [True, False]) |
| 551 | @pytest.mark.parametrize("per_out_ch", [True, False]) |
| 552 | @pytest.mark.skipif( |
| 553 | (lambda x: x is None or not ops.cutlass_group_gemm_supported(x.to_int()))( |
| 554 | current_platform.get_device_capability() |
| 555 | ), |
| 556 | reason="Grouped gemm is not supported on this GPU type.", |
| 557 | ) |
| 558 | def test_cutlass_moe_8_bit_no_graph( |
| 559 | m: int, |
| 560 | n: int, |
| 561 | k: int, |
| 562 | e: int, |
| 563 | topk: int, |
| 564 | per_act_token: bool, |
| 565 | per_out_ch: bool, |
| 566 | monkeypatch, |
| 567 | workspace_init, |
| 568 | ep_size: int | None = None, |
| 569 | ): |
| 570 | set_random_seed(7) |
| 571 | with set_current_vllm_config(vllm_config): |
| 572 | mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token, per_out_ch) |
| 573 | |
| 574 | score = torch.randn((m, e), device="cuda", dtype=torch.half) |
| 575 | topk_weights, topk_ids, _ = fused_topk(mt.a, score, topk, renormalize=False) |
| 576 | |
| 577 | # Note that we are using the dequantized versions of the tensors. |
| 578 | # Using a, w1 and w2 directly results in minor output differences. |
| 579 | |
| 580 | quant_config = FUSED_MOE_UNQUANTIZED_CONFIG |
| 581 | triton_output = fused_experts( |
| 582 | mt.a_d, mt.w1_d, mt.w2_d, topk_weights, topk_ids, quant_config=quant_config |
| 583 | ) |
| 584 | |
| 585 | if ep_size is not None: |
| 586 | assert e % ep_size == 0, "Cannot distribute experts evenly" |
| 587 | number_local_experts = e // ep_size |
| 588 | else: |
| 589 | number_local_experts = None |
| 590 | |
| 591 | cutlass_output = run_8_bit( |
| 592 | mt, topk_weights, topk_ids, per_act_token, per_out_ch, number_local_experts |
| 593 | ) |
| 594 | |
| 595 | # Note 5.5 only needed for larger problem sizes, 5 works ok for |
| 596 | # the rest. |
| 597 | torch.testing.assert_close( |
| 598 | triton_output, cutlass_output, atol=5.5e-2, rtol=1e-2 |
| 599 | ) |
| 600 | |
| 601 | |
| 602 | @pytest.mark.parametrize("m,n,k", MNK_FACTORS) |
| 603 | @pytest.mark.parametrize("e", NUM_EXPERTS) |
| 604 | @pytest.mark.parametrize("topk", TOP_KS) |
| 605 | @pytest.mark.parametrize("per_act_token", [True, False]) |
| 606 | @pytest.mark.parametrize("per_out_ch", [True, False]) |
| 607 | @pytest.mark.skipif( |
| 608 | (lambda x: x is None or not ops.cutlass_group_gemm_supported(x.to_int()))( |
| 609 | current_platform.get_device_capability() |
| 610 | ), |
| 611 | reason="Grouped gemm is not supported on this GPU type.", |
| 612 | ) |
| 613 | def test_cutlass_moe_8_bit_cuda_graph( |
| 614 | m: int, |
| 615 | n: int, |
| 616 | k: int, |
| 617 | e: int, |
| 618 | topk: int, |
| 619 | per_act_token: bool, |
| 620 | per_out_ch: bool, |
| 621 | monkeypatch, |
| 622 | workspace_init, |
| 623 | ): |
| 624 | set_random_seed(7) |
| 625 | with set_current_vllm_config(vllm_config): |
| 626 | dtype = torch.half |
| 627 | |
| 628 | mt = MOETensors8Bit.make_moe_tensors_8bit(m, k, n, e, per_act_token, per_out_ch) |
| 629 | |
| 630 | score = torch.randn((m, e), device="cuda", dtype=dtype) |
| 631 | topk_weights, topk_ids, _ = fused_topk(mt.a, score, topk, renormalize=False) |
| 632 | |
| 633 | # Note that we are using the dequantized versions of the tensors. |
| 634 | # Using a, w1 and w2 directly results in minor output differences. |
| 635 | quant_config = FUSED_MOE_UNQUANTIZED_CONFIG |
| 636 | triton_output = fused_experts( |
| 637 | mt.a_d, mt.w1_d, mt.w2_d, topk_weights, topk_ids, quant_config=quant_config |
| 638 | ) |
| 639 | |
| 640 | stream = torch.cuda.Stream() |
| 641 | graph = torch.cuda.CUDAGraph() |
| 642 | with torch.cuda.graph(graph, stream=stream): |
| 643 | cutlass_output = run_8_bit( |
| 644 | mt, topk_weights, topk_ids, per_act_token, per_out_ch |
| 645 | ) |
| 646 | |
| 647 | torch.accelerator.synchronize() |
| 648 | graph.replay() |
| 649 | torch.accelerator.synchronize() |
| 650 | |
| 651 | torch.testing.assert_close(triton_output, cutlass_output, atol=9e-2, rtol=1e-2) |
| 652 | |
| 653 | |
| 654 | @pytest.mark.parametrize("m", [64]) |
| 655 | @pytest.mark.parametrize("n", [1024]) |
| 656 | @pytest.mark.parametrize("k", [4096]) |
| 657 | @pytest.mark.parametrize("e", [16]) |
| 658 | @pytest.mark.parametrize("topk", [1, 8]) |
| 659 | @pytest.mark.parametrize("per_act_token", [True]) |
| 660 | @pytest.mark.parametrize("per_out_channel", [True]) |
| 661 | @pytest.mark.parametrize("ep_size", [1, 2, 4, 8, 16]) |
| 662 | @pytest.mark.skipif( |
| 663 | (lambda x: x is None or not ops.cutlass_group_gemm_supported(x.to_int()))( |
| 664 | current_platform.get_device_capability() |
| 665 | ), |
| 666 | reason="Grouped gemm is not supported on this GPU type.", |
| 667 | ) |
| 668 | def test_cutlass_moe_8_bit_EP( |
| 669 | m: int, |
| 670 | n: int, |
| 671 | k: int, |
| 672 | e: int, |
| 673 | topk: int, |
| 674 | per_act_token: bool, |
| 675 | per_out_channel: bool, |
| 676 | ep_size: int, |
| 677 | monkeypatch, |
| 678 | workspace_init, |
| 679 | ): |
| 680 | test_cutlass_moe_8_bit_no_graph( |
| 681 | m, |
| 682 | n, |
| 683 | k, |
| 684 | e, |
| 685 | topk, |
| 686 | per_act_token, |
| 687 | per_out_channel, |
| 688 | monkeypatch, |
| 689 | workspace_init, |
| 690 | ep_size, |
| 691 | ) |
| 692 | |
| 693 | |
| 694 | LARGE_MNK_FACTORS = [ |
| 695 | (1, 8192, 5120, 31), |
| 696 | (32768, 1024, 1024, 16), |
| 697 | (65536, 512, 1024, 16), |
| 698 | ] |
| 699 | |
| 700 | |
| 701 | @pytest.mark.parametrize("m,n,k,topk", LARGE_MNK_FACTORS) |
| 702 | @pytest.mark.parametrize("e", [128]) |
| 703 | @pytest.mark.parametrize("per_act_token", [False]) |
| 704 | @pytest.mark.parametrize("per_out_channel", [True]) |
| 705 | @pytest.mark.parametrize("ep_size", [8]) |
| 706 | @pytest.mark.skipif( |
| 707 | (lambda x: x is None or not ops.cutlass_group_gemm_supported(x.to_int()))( |
| 708 | current_platform.get_device_capability() |
| 709 | ), |
| 710 | reason="Grouped gemm is not supported on this GPU type.", |
| 711 | ) |
| 712 | def test_cutlass_moe_8_bit_EP_large( |
| 713 | m: int, |
| 714 | n: int, |
| 715 | k: int, |
| 716 | e: int, |
| 717 | topk: int, |
| 718 | per_act_token: bool, |
| 719 | per_out_channel: bool, |
| 720 | ep_size: int, |
| 721 | monkeypatch, |
| 722 | workspace_init, |
| 723 | ): |
| 724 | test_cutlass_moe_8_bit_no_graph( |
| 725 | m, |
| 726 | n, |
| 727 | k, |
| 728 | e, |
| 729 | topk, |
| 730 | per_act_token, |
| 731 | per_out_channel, |
| 732 | monkeypatch, |
| 733 | workspace_init, |
| 734 | ep_size, |
| 735 | ) |
| 736 | |
| 737 | |
| 738 | @pytest.mark.parametrize("m,n,k,topk", [(1, 8192, 5120, 31)]) |
| 739 | @pytest.mark.parametrize("e", [128]) |
| 740 | @pytest.mark.parametrize("per_act_token", [False]) |
| 741 | @pytest.mark.parametrize("per_out_channel", [True]) |
| 742 | @pytest.mark.parametrize("ep_size", [8]) |
| 743 | @pytest.mark.skipif( |
| 744 | (lambda x: x is None or not ops.cutlass_group_gemm_supported(x.to_int()))( |
| 745 | current_platform.get_device_capability() |
| 746 | ), |
| 747 | reason="Grouped gemm is not supported on this GPU type.", |
| 748 | ) |
| 749 | def test_run_cutlass_moe_fp8( |
| 750 | m: int, |
| 751 | n: int, |
| 752 | k: int, |
| 753 | e: int, |
| 754 | topk: int, |
| 755 | per_act_token: bool, |
| 756 | per_out_channel: bool, |
| 757 | ep_size: int, |
| 758 | workspace_init, |
| 759 | ): |
| 760 | set_random_seed(7) |
| 761 | with set_current_vllm_config(vllm_config): |
| 762 | mt = MOETensors8Bit.make_moe_tensors_8bit( |
| 763 | m, k, n, e, per_act_token, per_out_channel |
| 764 | ) |
| 765 | |
| 766 | score = torch.randn((m, e), device="cuda", dtype=torch.half) |
| 767 | topk_weights, topk_ids, _ = fused_topk(mt.a, score, topk, renormalize=False) |
| 768 | # we want to make sure there is at least one token that's generated in |
| 769 | # this expert shard and at least one token that's NOT generated in this |
| 770 | # expert shard |
| 771 | topk_ids[0][0] = -1 |
| 772 | topk_ids[0][1] = 1 |
| 773 | |
| 774 | workspace13_shape = (m * topk, max(2 * n, k)) |
| 775 | workspace2_shape = (m * topk, max(n, k)) |
| 776 | output_shape = (m, k) |
| 777 | |
| 778 | workspace13 = torch.empty( |
| 779 | prod(workspace13_shape), device="cuda", dtype=mt.a.dtype |
| 780 | ) |
| 781 | workspace2 = torch.empty( |
| 782 | prod(workspace2_shape), device="cuda", dtype=mt.a.dtype |
| 783 | ) |
| 784 | |
| 785 | num_local_experts = e // ep_size |
| 786 | start, end = 0, num_local_experts |
| 787 | expert_map = [-1] * e |
| 788 | expert_map[start:end] = list(range(num_local_experts)) |
| 789 | expert_map = torch.tensor(expert_map, dtype=torch.int32, device="cuda") |
| 790 | |
| 791 | ab_strides1 = torch.full((e,), k, device="cuda", dtype=torch.int64) |
| 792 | ab_strides2 = torch.full((e,), n, device="cuda", dtype=torch.int64) |
| 793 | c_strides1 = torch.full((e,), 2 * n, device="cuda", dtype=torch.int64) |
| 794 | c_strides2 = torch.full((e,), k, device="cuda", dtype=torch.int64) |
| 795 | |
| 796 | activation = MoEActivation.SILU |
| 797 | a1q, a1q_scale = moe_kernel_quantize_input( |
| 798 | mt.a, mt.a_scale, torch.float8_e4m3fn, per_act_token |
| 799 | ) |
| 800 | global_num_experts = -1 if mt.w1_q is None else mt.w1_q.size(0) |
| 801 | func = lambda output: run_cutlass_moe_fp8( |
| 802 | output, |
| 803 | a1q, |
| 804 | mt.w1_q, |
| 805 | mt.w2_q, |
| 806 | topk_ids, |
| 807 | activation, |
| 808 | global_num_experts, |
| 809 | expert_map, |
| 810 | mt.w1_scale, |
| 811 | mt.w2_scale, |
| 812 | a1q_scale, |
| 813 | None, |
| 814 | ab_strides1, |
| 815 | ab_strides2, |
| 816 | c_strides1, |
| 817 | c_strides2, |
| 818 | workspace13, |
| 819 | workspace2, |
| 820 | None, |
| 821 | mt.a.dtype, |
| 822 | per_act_token, |
| 823 | per_out_channel, |
| 824 | False, |
| 825 | topk_weights, |
| 826 | None, |
| 827 | ) |
| 828 | |
| 829 | workspace13.random_() |
| 830 | output_random_workspace = torch.empty( |
| 831 | output_shape, device="cuda", dtype=mt.a.dtype |
| 832 | ) |
| 833 | func(output_random_workspace) |
| 834 | |
| 835 | workspace13.fill_(0) |
| 836 | output_zero_workspace = torch.zeros( |
| 837 | output_shape, device="cuda", dtype=mt.a.dtype |
| 838 | ) |
| 839 | func(output_zero_workspace) |
| 840 | |
| 841 | torch.testing.assert_close( |
| 842 | output_random_workspace, output_zero_workspace, atol=5e-3, rtol=1e-3 |
| 843 | ) |