| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for the b12x tensor-parallel MoE integration.""" |
| 4 | |
| 5 | from dataclasses import dataclass, replace |
| 6 | from types import SimpleNamespace |
| 7 | |
| 8 | import pytest |
| 9 | import torch |
| 10 | |
| 11 | import vllm.model_executor.layers.fused_moe.b12x as b12x |
| 12 | import vllm.model_executor.layers.fused_moe.modular_kernel as mk |
| 13 | import vllm.model_executor.layers.fused_moe.oracle.mxfp4 as mxfp4_oracle |
| 14 | import vllm.model_executor.layers.fused_moe.oracle.nvfp4 as nvfp4_oracle |
| 15 | from tests.kernels.moe.utils import make_dummy_moe_config |
| 16 | from tests.kernels.quantization.nvfp4_utils import ( |
| 17 | dequantize_nvfp4_to_dtype, |
| 18 | quant_nvfp4_tensor, |
| 19 | ) |
| 20 | from tests.kernels.utils import torch_moe |
| 21 | from tests.quantization.reference_mxfp4 import dq_mxfp4_torch |
| 22 | from vllm import _custom_ops as ops |
| 23 | from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config |
| 24 | from vllm.model_executor.layers.fused_moe import fused_topk |
| 25 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 26 | from vllm.model_executor.layers.fused_moe.all2all_utils import ( |
| 27 | maybe_make_prepare_finalize, |
| 28 | ) |
| 29 | from vllm.model_executor.layers.fused_moe.b12x import B12xExperts |
| 30 | from vllm.model_executor.layers.fused_moe.config import ( |
| 31 | FusedMoEParallelConfig, |
| 32 | FusedMoEQuantConfig, |
| 33 | mxfp4_w4a16_moe_quant_config, |
| 34 | nvfp4_w4a16_moe_quant_config, |
| 35 | ) |
| 36 | from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import ( |
| 37 | Mxfp4MoeBackend, |
| 38 | select_deepseek_v4_mxfp4_moe_backend, |
| 39 | select_mxfp4_moe_backend, |
| 40 | ) |
| 41 | from vllm.model_executor.layers.fused_moe.oracle.nvfp4 import ( |
| 42 | NvFp4MoeBackend, |
| 43 | select_nvfp4_moe_backend, |
| 44 | ) |
| 45 | from vllm.model_executor.layers.quantization.utils.b12x_moe import ( |
| 46 | prepare_nvfp4_moe_layer_for_b12x, |
| 47 | ) |
| 48 | from vllm.model_executor.layers.quantization.utils.mxfp4_utils import mxfp4_quantize |
| 49 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 50 | kMxfp4Static, |
| 51 | kMxfp8Dynamic, |
| 52 | kNvfp4Dynamic, |
| 53 | kNvfp4Static, |
| 54 | ) |
| 55 | from vllm.platforms import current_platform |
| 56 | from vllm.utils.torch_utils import set_random_seed |
| 57 | |
| 58 | |
| 59 | def _quantize_nvfp4_linear( |
| 60 | weight: torch.Tensor, |
| 61 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: |
| 62 | weights_q = [] |
| 63 | scales = [] |
| 64 | global_scales = [] |
| 65 | for expert_weight in weight: |
| 66 | weight_q, scale, global_scale = quant_nvfp4_tensor( |
| 67 | expert_weight, |
| 68 | is_sf_swizzled_layout=False, |
| 69 | ) |
| 70 | weights_q.append(weight_q) |
| 71 | scales.append(scale) |
| 72 | global_scales.append(global_scale) |
| 73 | return torch.stack(weights_q), torch.stack(scales), torch.stack(global_scales) |
| 74 | |
| 75 | |
| 76 | def _dequantize_nvfp4_linear( |
| 77 | tensor_fp4: torch.Tensor, |
| 78 | tensor_sf: torch.Tensor, |
| 79 | global_scale: torch.Tensor, |
| 80 | dtype: torch.dtype, |
| 81 | ) -> torch.Tensor: |
| 82 | return dequantize_nvfp4_to_dtype( |
| 83 | tensor_fp4, |
| 84 | tensor_sf, |
| 85 | global_scale, |
| 86 | dtype=dtype, |
| 87 | device=tensor_fp4.device, |
| 88 | is_sf_linear_layout=True, |
| 89 | ) |
| 90 | |
| 91 | |
| 92 | def _nvfp4_activation_reference( |
| 93 | hidden_states: torch.Tensor, |
| 94 | w1: torch.Tensor, |
| 95 | w2: torch.Tensor, |
| 96 | topk_weights: torch.Tensor, |
| 97 | topk_ids: torch.Tensor, |
| 98 | a1_scale: torch.Tensor, |
| 99 | a2_scale: torch.Tensor, |
| 100 | ) -> torch.Tensor: |
| 101 | tokens, hidden_size = hidden_states.shape |
| 102 | topk = topk_ids.shape[1] |
| 103 | routed_input = ( |
| 104 | hidden_states[:, None, :] |
| 105 | .expand(-1, topk, -1) |
| 106 | .reshape(tokens * topk, hidden_size) |
| 107 | ) |
| 108 | routed_output = torch.zeros( |
| 109 | tokens * topk, |
| 110 | hidden_size, |
| 111 | dtype=torch.float32, |
| 112 | device=hidden_states.device, |
| 113 | ) |
| 114 | flat_ids = topk_ids.reshape(-1) |
| 115 | |
| 116 | for expert in range(w1.shape[0]): |
| 117 | mask = flat_ids == expert |
| 118 | if not mask.any(): |
| 119 | continue |
| 120 | a1_q, a1_block_scale = ops.scaled_fp4_quant( |
| 121 | routed_input[mask], |
| 122 | a1_scale[expert], |
| 123 | is_sf_swizzled_layout=False, |
| 124 | ) |
| 125 | a1 = _dequantize_nvfp4_linear( |
| 126 | a1_q, |
| 127 | a1_block_scale, |
| 128 | a1_scale[expert], |
| 129 | torch.float32, |
| 130 | ) |
| 131 | fc1 = a1 @ w1[expert].float().t() |
| 132 | gate, up = fc1.chunk(2, dim=-1) |
| 133 | intermediate = (torch.nn.functional.silu(gate) * up).to(torch.bfloat16) |
| 134 | a2_q, a2_block_scale = ops.scaled_fp4_quant( |
| 135 | intermediate, |
| 136 | a2_scale[expert], |
| 137 | is_sf_swizzled_layout=False, |
| 138 | ) |
| 139 | a2 = _dequantize_nvfp4_linear( |
| 140 | a2_q, |
| 141 | a2_block_scale, |
| 142 | a2_scale[expert], |
| 143 | torch.float32, |
| 144 | ) |
| 145 | routed_output[mask] = a2 @ w2[expert].float().t() |
| 146 | |
| 147 | return ( |
| 148 | routed_output.view(tokens, topk, hidden_size) |
| 149 | .mul(topk_weights[..., None]) |
| 150 | .sum(dim=1) |
| 151 | .to(hidden_states.dtype) |
| 152 | ) |
| 153 | |
| 154 | |
| 155 | def _has_b12x_moe() -> bool: |
| 156 | return ( |
| 157 | torch.cuda.is_available() |
| 158 | and current_platform.is_device_capability_family(120) |
| 159 | and B12xExperts._supports_current_device() |
| 160 | ) |
| 161 | |
| 162 | |
| 163 | def _count_fp4_negative_zeros(packed: torch.Tensor) -> int: |
| 164 | low = (packed & 0x0F) == 0x08 |
| 165 | high = (packed & 0xF0) == 0x80 |
| 166 | return int(low.sum().item() + high.sum().item()) |
| 167 | |
| 168 | |
| 169 | def _make_b12x_moe_kernel( |
| 170 | hidden_states: torch.Tensor, |
| 171 | w1: torch.Tensor, |
| 172 | w2: torch.Tensor, |
| 173 | topk: int, |
| 174 | activation: MoEActivation, |
| 175 | quant_config: FusedMoEQuantConfig, |
| 176 | ) -> mk.FusedMoEKernel: |
| 177 | num_experts = w1.shape[0] |
| 178 | moe_config = make_dummy_moe_config( |
| 179 | num_experts=num_experts, |
| 180 | experts_per_token=topk, |
| 181 | hidden_dim=hidden_states.shape[1], |
| 182 | intermediate_size=w2.shape[2] * 2, |
| 183 | in_dtype=hidden_states.dtype, |
| 184 | activation=activation, |
| 185 | ) |
| 186 | experts = B12xExperts(moe_config, quant_config) |
| 187 | layer = SimpleNamespace( |
| 188 | activation=activation, |
| 189 | apply_router_weight_on_input=False, |
| 190 | w13_weight=w1, |
| 191 | w2_weight=w2, |
| 192 | w13_weight_scale=quant_config.w1_scale, |
| 193 | w2_weight_scale=quant_config.w2_scale, |
| 194 | ) |
| 195 | if quant_config.weight_quant_dtype == "nvfp4": |
| 196 | layer.w13_weight_scale_2 = quant_config.g1_alphas |
| 197 | layer.w2_weight_scale_2 = quant_config.g2_alphas |
| 198 | if quant_config.quant_dtype is not None: |
| 199 | assert quant_config.a1_gscale is not None |
| 200 | assert quant_config.a2_gscale is not None |
| 201 | layer.w13_input_scale = 1.0 / quant_config.a1_gscale |
| 202 | layer.w2_input_scale = 1.0 / quant_config.a2_gscale |
| 203 | experts.process_weights_after_loading(layer) |
| 204 | return mk.FusedMoEKernel( |
| 205 | maybe_make_prepare_finalize( |
| 206 | moe=moe_config, |
| 207 | quant_config=quant_config, |
| 208 | allow_new_interface=True, |
| 209 | use_monolithic=False, |
| 210 | ), |
| 211 | experts, |
| 212 | ) |
| 213 | |
| 214 | |
| 215 | def _run_b12x_moe( |
| 216 | hidden_states: torch.Tensor, |
| 217 | w1: torch.Tensor, |
| 218 | w2: torch.Tensor, |
| 219 | score: torch.Tensor, |
| 220 | topk: int, |
| 221 | activation: MoEActivation, |
| 222 | quant_config: FusedMoEQuantConfig, |
| 223 | ) -> torch.Tensor: |
| 224 | num_experts = w1.shape[0] |
| 225 | kernel = _make_b12x_moe_kernel( |
| 226 | hidden_states, |
| 227 | w1, |
| 228 | w2, |
| 229 | topk, |
| 230 | activation, |
| 231 | quant_config, |
| 232 | ) |
| 233 | topk_weights, topk_ids, _ = fused_topk( |
| 234 | hidden_states, score, topk, renormalize=False |
| 235 | ) |
| 236 | return kernel.apply( |
| 237 | hidden_states=hidden_states, |
| 238 | w1=w1, |
| 239 | w2=w2, |
| 240 | topk_weights=topk_weights, |
| 241 | topk_ids=topk_ids, |
| 242 | activation=activation, |
| 243 | global_num_experts=num_experts, |
| 244 | expert_map=None, |
| 245 | apply_router_weight_on_input=False, |
| 246 | ) |
| 247 | |
| 248 | |
| 249 | def _quant_config(weight_dtype: str, activation_dtype: str | None): |
| 250 | scale = torch.ones(1, dtype=torch.float32) |
| 251 | return FusedMoEQuantConfig.make( |
| 252 | quant_dtype=activation_dtype, |
| 253 | weight_dtype=weight_dtype, |
| 254 | w1_scale=scale, |
| 255 | w2_scale=scale, |
| 256 | g1_alphas=scale, |
| 257 | g2_alphas=scale, |
| 258 | a1_gscale=scale, |
| 259 | a2_gscale=scale, |
| 260 | ) |
| 261 | |
| 262 | |
| 263 | def test_b12x_moe_supports_only_tensor_parallel() -> None: |
| 264 | parallel = FusedMoEParallelConfig.make_no_parallel() |
| 265 | |
| 266 | assert B12xExperts._supports_parallel_config(parallel) |
| 267 | assert not B12xExperts._supports_parallel_config( |
| 268 | replace(parallel, use_ep=True, ep_size=2) |
| 269 | ) |
| 270 | all2all = replace(parallel, use_ep=True, dp_size=2) |
| 271 | assert all2all.use_all2all_kernels |
| 272 | assert not B12xExperts._supports_parallel_config(all2all) |
| 273 | assert not B12xExperts._supports_parallel_config( |
| 274 | replace(parallel, enable_eplb=True) |
| 275 | ) |
| 276 | |
| 277 | |
| 278 | _SITU_REASON = "kernel supports only SiTU beta=4 and linear_beta=25" |
| 279 | _UNINTERLEAVED_W4A8_REASON = "kernel does not support swigluoai_uninterleave with W4A8" |
| 280 | |
| 281 | |
| 282 | @pytest.mark.parametrize( |
| 283 | "config_kwargs,overrides,weight_key,activation_key,expected_reason", |
| 284 | [ |
| 285 | pytest.param( |
| 286 | {"hidden_dim": 128, "activation": MoEActivation.SWIGLUOAI}, |
| 287 | {}, |
| 288 | kMxfp4Static, |
| 289 | None, |
| 290 | "kernel does not support MoEActivation.SWIGLUOAI activation", |
| 291 | id="interleaved-swigluoai", |
| 292 | ), |
| 293 | pytest.param( |
| 294 | { |
| 295 | "hidden_dim": 128, |
| 296 | "activation": MoEActivation.SWIGLUOAI_UNINTERLEAVE, |
| 297 | }, |
| 298 | {}, |
| 299 | kMxfp4Static, |
| 300 | kMxfp8Dynamic, |
| 301 | _UNINTERLEAVED_W4A8_REASON, |
| 302 | id="mxfp4-w4a8-uninterleaved-swigluoai", |
| 303 | ), |
| 304 | pytest.param( |
| 305 | { |
| 306 | "hidden_dim": 128, |
| 307 | "activation": MoEActivation.SWIGLUOAI_UNINTERLEAVE, |
| 308 | }, |
| 309 | {}, |
| 310 | kMxfp4Static, |
| 311 | None, |
| 312 | None, |
| 313 | id="w4a16-uninterleaved-swigluoai", |
| 314 | ), |
| 315 | pytest.param( |
| 316 | {"activation": MoEActivation.RELU2_NO_MUL}, |
| 317 | {}, |
| 318 | kMxfp4Static, |
| 319 | kMxfp8Dynamic, |
| 320 | "MXFP4 W4A8 supports only SiLU and SiTU", |
| 321 | id="mxfp4-w4a8-relu2", |
| 322 | ), |
| 323 | pytest.param( |
| 324 | {"hidden_dim": 128}, |
| 325 | {}, |
| 326 | kMxfp4Static, |
| 327 | kMxfp8Dynamic, |
| 328 | ( |
| 329 | "MXFP4 W4A8 requires hidden size divisible by 256 and per-rank " |
| 330 | "intermediate size divisible by 32" |
| 331 | ), |
| 332 | id="mxfp4-w4a8-alignment", |
| 333 | ), |
| 334 | pytest.param( |
| 335 | {"hidden_dim": 128, "intermediate_size": 48}, |
| 336 | {"intermediate_size_per_partition": 64}, |
| 337 | kMxfp4Static, |
| 338 | None, |
| 339 | "MXFP4 requires the per-rank intermediate size to be divisible by 32", |
| 340 | id="mxfp4-tp-scale-groups", |
| 341 | ), |
| 342 | pytest.param( |
| 343 | {"activation": MoEActivation.SITU}, |
| 344 | {"activation_situ_beta": 3.0, "activation_situ_linear_beta": 25.0}, |
| 345 | kMxfp4Static, |
| 346 | None, |
| 347 | _SITU_REASON, |
| 348 | id="situ-beta", |
| 349 | ), |
| 350 | pytest.param( |
| 351 | {"activation": MoEActivation.SITU}, |
| 352 | {"activation_situ_beta": 4.0, "activation_situ_linear_beta": 25.0}, |
| 353 | kMxfp4Static, |
| 354 | None, |
| 355 | None, |
| 356 | id="situ-standard-parameters", |
| 357 | ), |
| 358 | ], |
| 359 | ) |
| 360 | def test_b12x_moe_config_support( |
| 361 | monkeypatch: pytest.MonkeyPatch, |
| 362 | config_kwargs, |
| 363 | overrides, |
| 364 | weight_key, |
| 365 | activation_key, |
| 366 | expected_reason: str | None, |
| 367 | ) -> None: |
| 368 | monkeypatch.setattr(B12xExperts, "_supports_current_device", lambda: True) |
| 369 | config = make_dummy_moe_config( |
| 370 | **{"hidden_dim": 256, "intermediate_size": 64, **config_kwargs} |
| 371 | ) |
| 372 | for name, value in overrides.items(): |
| 373 | setattr(config, name, value) |
| 374 | |
| 375 | supported, reason = B12xExperts.is_supported_config( |
| 376 | B12xExperts, |
| 377 | config, |
| 378 | weight_key, |
| 379 | activation_key, |
| 380 | mk.FusedMoEActivationFormat.Standard, |
| 381 | ) |
| 382 | |
| 383 | assert (supported, reason) == (expected_reason is None, expected_reason) |
| 384 | |
| 385 | |
| 386 | @pytest.mark.parametrize( |
| 387 | "activation_key,force_a16,expected_backend", |
| 388 | [ |
| 389 | (kMxfp8Dynamic, False, Mxfp4MoeBackend.B12X_MXFP4_MXFP8), |
| 390 | (None, False, Mxfp4MoeBackend.B12X_MXFP4_MXFP8), |
| 391 | ], |
| 392 | ) |
| 393 | def test_explicit_b12x_mxfp4_selection( |
| 394 | monkeypatch: pytest.MonkeyPatch, |
| 395 | activation_key, |
| 396 | force_a16: bool, |
| 397 | expected_backend: Mxfp4MoeBackend, |
| 398 | ) -> None: |
| 399 | monkeypatch.setattr(B12xExperts, "_supports_current_device", lambda: True) |
| 400 | monkeypatch.setattr(mxfp4_oracle, "_user_moe_activation_override", lambda: None) |
| 401 | monkeypatch.setattr( |
| 402 | mxfp4_oracle.envs, |
| 403 | "VLLM_B12X_MOE_FP4_FORCE_A16", |
| 404 | force_a16, |
| 405 | ) |
| 406 | config = make_dummy_moe_config(hidden_dim=256, intermediate_size=64) |
| 407 | config.moe_backend = "b12x" |
| 408 | |
| 409 | backend, experts_cls = select_mxfp4_moe_backend( |
| 410 | config, |
| 411 | activation_key=activation_key, |
| 412 | ) |
| 413 | |
| 414 | assert backend == expected_backend |
| 415 | assert experts_cls is B12xExperts |
| 416 | |
| 417 | |
| 418 | def test_explicit_b12x_mxfp4_force_a16_uses_a16_contract( |
| 419 | monkeypatch: pytest.MonkeyPatch, |
| 420 | ) -> None: |
| 421 | monkeypatch.setattr(B12xExperts, "_supports_current_device", lambda: True) |
| 422 | monkeypatch.setattr(mxfp4_oracle, "_user_moe_activation_override", lambda: None) |
| 423 | monkeypatch.setattr( |
| 424 | mxfp4_oracle.envs, |
| 425 | "VLLM_B12X_MOE_FP4_FORCE_A16", |
| 426 | True, |
| 427 | ) |
| 428 | config = make_dummy_moe_config(hidden_dim=128, intermediate_size=64) |
| 429 | config.moe_backend = "b12x" |
| 430 | |
| 431 | backend, experts_cls = select_mxfp4_moe_backend( |
| 432 | config, |
| 433 | activation_key=kMxfp8Dynamic, |
| 434 | ) |
| 435 | |
| 436 | assert backend == Mxfp4MoeBackend.B12X_MXFP4_BF16 |
| 437 | assert experts_cls is B12xExperts |
| 438 | |
| 439 | |
| 440 | @pytest.mark.parametrize( |
| 441 | "force_a16,expected_backend", |
| 442 | [ |
| 443 | (False, Mxfp4MoeBackend.B12X_MXFP4_MXFP8), |
| 444 | (True, Mxfp4MoeBackend.B12X_MXFP4_BF16), |
| 445 | ], |
| 446 | ) |
| 447 | def test_deepseek_v4_b12x_activation_selection( |
| 448 | monkeypatch: pytest.MonkeyPatch, |
| 449 | force_a16: bool, |
| 450 | expected_backend: Mxfp4MoeBackend, |
| 451 | ) -> None: |
| 452 | monkeypatch.setattr(B12xExperts, "_supports_current_device", lambda: True) |
| 453 | monkeypatch.setattr( |
| 454 | mxfp4_oracle.envs, |
| 455 | "VLLM_B12X_MOE_FP4_FORCE_A16", |
| 456 | force_a16, |
| 457 | ) |
| 458 | config = make_dummy_moe_config(hidden_dim=256, intermediate_size=64) |
| 459 | config.moe_backend = "b12x" |
| 460 | |
| 461 | backend, experts_cls = select_deepseek_v4_mxfp4_moe_backend(config) |
| 462 | |
| 463 | assert backend == expected_backend |
| 464 | assert experts_cls is B12xExperts |
| 465 | |
| 466 | |
| 467 | def test_deepseek_v4_flashinfer_cutlass_falls_through_to_w4a8( |
| 468 | monkeypatch: pytest.MonkeyPatch, |
| 469 | ) -> None: |
| 470 | """Explicit flashinfer_cutlass must try the W4A8 variant when the BF16 |
| 471 | variant is unsupported (the BF16 variant is gated to SM90).""" |
| 472 | from vllm.model_executor.layers.fused_moe.experts.flashinfer_cutlass_moe import ( |
| 473 | FlashInferExperts, |
| 474 | ) |
| 475 | |
| 476 | monkeypatch.setattr(FlashInferExperts, "_supports_current_device", lambda: True) |
| 477 | |
| 478 | def sm120_quant_gate(weight_key, activation_key): |
| 479 | return (weight_key, activation_key) == ( |
| 480 | mxfp4_oracle.kMxfp4Static, |
| 481 | mxfp4_oracle.kMxfp8Dynamic, |
| 482 | ) |
| 483 | |
| 484 | monkeypatch.setattr( |
| 485 | FlashInferExperts, "_supports_quant_scheme", staticmethod(sm120_quant_gate) |
| 486 | ) |
| 487 | config = make_dummy_moe_config(hidden_dim=256, intermediate_size=64) |
| 488 | config.moe_backend = "flashinfer_cutlass" |
| 489 | |
| 490 | backend, experts_cls = select_deepseek_v4_mxfp4_moe_backend(config) |
| 491 | |
| 492 | assert backend == Mxfp4MoeBackend.FLASHINFER_CUTLASS_MXFP4_MXFP8 |
| 493 | assert experts_cls is FlashInferExperts |
| 494 | |
| 495 | |
| 496 | def test_compressed_tensors_mxfp4_preserves_checkpoint_packing( |
| 497 | monkeypatch: pytest.MonkeyPatch, |
| 498 | ) -> None: |
| 499 | from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import ( # noqa: E501 |
| 500 | compressed_tensors_moe_w4a4_mxfp4 as ct_mxfp4, |
| 501 | ) |
| 502 | |
| 503 | monkeypatch.setattr( |
| 504 | ct_mxfp4.CutlassExpertsMxfp4, |
| 505 | "_supports_current_device", |
| 506 | lambda: False, |
| 507 | ) |
| 508 | monkeypatch.setattr( |
| 509 | ct_mxfp4, |
| 510 | "select_mxfp4_moe_backend", |
| 511 | lambda moe: (Mxfp4MoeBackend.B12X_MXFP4_MXFP8, B12xExperts), |
| 512 | ) |
| 513 | monkeypatch.setattr( |
| 514 | ct_mxfp4, |
| 515 | "prepare_moe_fp4_layer_for_marlin", |
| 516 | lambda layer: pytest.fail("b12x must not use Marlin packing"), |
| 517 | ) |
| 518 | moe_config = SimpleNamespace(w13_num_shards=2, moe_backend="b12x") |
| 519 | method = ct_mxfp4.CompressedTensorsW4A4Mxfp4MoEMethod(moe_config) |
| 520 | processed_layers: list[torch.nn.Module] = [] |
| 521 | fake_experts = SimpleNamespace( |
| 522 | process_weights_after_loading=processed_layers.append |
| 523 | ) |
| 524 | kernel = SimpleNamespace(fused_experts=fake_experts) |
| 525 | monkeypatch.setattr(method, "get_fused_moe_quant_config", lambda _: object()) |
| 526 | monkeypatch.setattr(ct_mxfp4, "make_mxfp4_moe_kernel", lambda **_: kernel) |
| 527 | layer = torch.nn.Module() |
| 528 | layer._expert_routing_tables = lambda: () |
| 529 | method.create_weights( |
| 530 | layer, |
| 531 | num_experts=2, |
| 532 | hidden_size=64, |
| 533 | intermediate_size_per_partition=32, |
| 534 | params_dtype=torch.bfloat16, |
| 535 | ) |
| 536 | w13_packed_data = layer.w13_weight_packed.data |
| 537 | w2_packed_data = layer.w2_weight_packed.data |
| 538 | |
| 539 | method.process_weights_after_loading(layer) |
| 540 | |
| 541 | assert layer.w13_weight.data.data_ptr() == w13_packed_data.data_ptr() |
| 542 | assert layer.w2_weight.data.data_ptr() == w2_packed_data.data_ptr() |
| 543 | assert processed_layers == [layer] |
| 544 | |
| 545 | |
| 546 | def test_b12x_mxfp4_falls_back_to_a16( |
| 547 | monkeypatch: pytest.MonkeyPatch, |
| 548 | ) -> None: |
| 549 | monkeypatch.setattr(B12xExperts, "_supports_current_device", lambda: True) |
| 550 | monkeypatch.setattr(mxfp4_oracle, "_user_moe_activation_override", lambda: None) |
| 551 | monkeypatch.setattr( |
| 552 | mxfp4_oracle.envs, |
| 553 | "VLLM_B12X_MOE_FP4_FORCE_A16", |
| 554 | False, |
| 555 | ) |
| 556 | config = make_dummy_moe_config(hidden_dim=128, intermediate_size=64) |
| 557 | config.moe_backend = "b12x" |
| 558 | |
| 559 | backend, experts_cls = select_mxfp4_moe_backend(config) |
| 560 | |
| 561 | assert backend == Mxfp4MoeBackend.B12X_MXFP4_BF16 |
| 562 | assert experts_cls is B12xExperts |
| 563 | |
| 564 | |
| 565 | @pytest.mark.parametrize( |
| 566 | "activation_key,force_a16,expected_activation_key", |
| 567 | [ |
| 568 | (kNvfp4Dynamic, False, kNvfp4Dynamic), |
| 569 | (kMxfp8Dynamic, False, kMxfp8Dynamic), |
| 570 | (kNvfp4Dynamic, True, None), |
| 571 | ], |
| 572 | ) |
| 573 | def test_explicit_b12x_nvfp4_selection( |
| 574 | monkeypatch: pytest.MonkeyPatch, |
| 575 | activation_key, |
| 576 | force_a16: bool, |
| 577 | expected_activation_key, |
| 578 | ) -> None: |
| 579 | selected_activation_keys = [] |
| 580 | |
| 581 | def is_supported_config(cls, config, weight_key, activation_key, activation_format): |
| 582 | selected_activation_keys.append(activation_key) |
| 583 | return True, None |
| 584 | |
| 585 | monkeypatch.setattr(B12xExperts, "is_supported_config", is_supported_config) |
| 586 | monkeypatch.setattr( |
| 587 | nvfp4_oracle.envs, |
| 588 | "VLLM_B12X_MOE_FP4_FORCE_A16", |
| 589 | force_a16, |
| 590 | ) |
| 591 | config = make_dummy_moe_config(hidden_dim=128, intermediate_size=64) |
| 592 | config.moe_backend = "b12x" |
| 593 | |
| 594 | backend, experts_cls = select_nvfp4_moe_backend( |
| 595 | config, |
| 596 | weight_key=kNvfp4Static, |
| 597 | activation_key=activation_key, |
| 598 | ) |
| 599 | |
| 600 | assert backend == NvFp4MoeBackend.B12X |
| 601 | assert experts_cls is B12xExperts |
| 602 | assert selected_activation_keys == [expected_activation_key] |
| 603 | |
| 604 | |
| 605 | @pytest.mark.parametrize( |
| 606 | "force_a16,expected_quant_dtype", [(False, "nvfp4"), (True, None)] |
| 607 | ) |
| 608 | def test_b12x_nvfp4_force_a16_updates_quant_config( |
| 609 | monkeypatch: pytest.MonkeyPatch, |
| 610 | force_a16: bool, |
| 611 | expected_quant_dtype, |
| 612 | ) -> None: |
| 613 | monkeypatch.setattr( |
| 614 | nvfp4_oracle.envs, |
| 615 | "VLLM_B12X_MOE_FP4_FORCE_A16", |
| 616 | force_a16, |
| 617 | ) |
| 618 | scale = torch.ones(1) |
| 619 | |
| 620 | quant_config = nvfp4_oracle.make_nvfp4_moe_quant_config( |
| 621 | backend=NvFp4MoeBackend.B12X, |
| 622 | w13_scale=scale, |
| 623 | w2_scale=scale, |
| 624 | w13_scale_2=scale, |
| 625 | w2_scale_2=scale, |
| 626 | a13_scale=scale, |
| 627 | a2_scale=scale, |
| 628 | ) |
| 629 | |
| 630 | assert quant_config.quant_dtype == expected_quant_dtype |
| 631 | |
| 632 | |
| 633 | def test_b12x_nvfp4_force_a16_updates_weight_preparation( |
| 634 | monkeypatch: pytest.MonkeyPatch, |
| 635 | ) -> None: |
| 636 | monkeypatch.setattr( |
| 637 | nvfp4_oracle.envs, |
| 638 | "VLLM_B12X_MOE_FP4_FORCE_A16", |
| 639 | True, |
| 640 | ) |
| 641 | reorder_w13 = None |
| 642 | |
| 643 | def prepare_for_b12x(**kwargs): |
| 644 | nonlocal reorder_w13 |
| 645 | reorder_w13 = kwargs["reorder_w13"] |
| 646 | return ( |
| 647 | kwargs["w13"], |
| 648 | kwargs["w13_scale"], |
| 649 | kwargs["w13_scale_2"], |
| 650 | kwargs["a13_scale"], |
| 651 | kwargs["w2"], |
| 652 | kwargs["w2_scale"], |
| 653 | kwargs["w2_scale_2"], |
| 654 | kwargs["a2_scale"], |
| 655 | ) |
| 656 | |
| 657 | monkeypatch.setattr( |
| 658 | nvfp4_oracle, |
| 659 | "prepare_nvfp4_moe_layer_for_b12x", |
| 660 | prepare_for_b12x, |
| 661 | ) |
| 662 | tensor = torch.ones(1) |
| 663 | |
| 664 | nvfp4_oracle.convert_to_nvfp4_moe_kernel_format( |
| 665 | nvfp4_backend=NvFp4MoeBackend.B12X, |
| 666 | layer=SimpleNamespace(), |
| 667 | w13=tensor, |
| 668 | w13_scale=tensor, |
| 669 | w13_scale_2=tensor, |
| 670 | a13_scale=tensor, |
| 671 | w2=tensor, |
| 672 | w2_scale=tensor, |
| 673 | w2_scale_2=tensor, |
| 674 | a2_scale=tensor, |
| 675 | is_act_and_mul=True, |
| 676 | ) |
| 677 | |
| 678 | assert reorder_w13 is True |
| 679 | |
| 680 | |
| 681 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 682 | def test_b12x_nvfp4_preparation_pads_each_gated_half() -> None: |
| 683 | device = torch.device("cuda") |
| 684 | num_experts, hidden_size, intermediate_size = 2, 64, 48 |
| 685 | w13 = torch.ones( |
| 686 | num_experts, |
| 687 | 2 * intermediate_size, |
| 688 | hidden_size // 2, |
| 689 | dtype=torch.uint8, |
| 690 | device=device, |
| 691 | ) |
| 692 | w13_scale = torch.ones( |
| 693 | num_experts, |
| 694 | 2 * intermediate_size, |
| 695 | hidden_size // 16, |
| 696 | dtype=torch.float8_e4m3fn, |
| 697 | device=device, |
| 698 | ) |
| 699 | w2 = torch.ones( |
| 700 | num_experts, |
| 701 | hidden_size, |
| 702 | intermediate_size // 2, |
| 703 | dtype=torch.uint8, |
| 704 | device=device, |
| 705 | ) |
| 706 | w2_scale = torch.ones( |
| 707 | num_experts, |
| 708 | hidden_size, |
| 709 | intermediate_size // 16, |
| 710 | dtype=torch.float8_e4m3fn, |
| 711 | device=device, |
| 712 | ) |
| 713 | global_scale = torch.ones(num_experts, device=device) |
| 714 | input_scale = torch.tensor([[1.0, 2.0], [3.0, 1.0]], device=device) |
| 715 | |
| 716 | prepared = prepare_nvfp4_moe_layer_for_b12x( |
| 717 | w13, |
| 718 | w13_scale, |
| 719 | global_scale, |
| 720 | input_scale, |
| 721 | w2, |
| 722 | w2_scale, |
| 723 | global_scale, |
| 724 | input_scale, |
| 725 | is_act_and_mul=True, |
| 726 | ) |
| 727 | |
| 728 | prepared_w13, prepared_w13_scale, _, prepared_a13 = prepared[:4] |
| 729 | prepared_w2, prepared_w2_scale, _, prepared_a2 = prepared[4:] |
| 730 | assert prepared_w13.shape == (num_experts, 128, hidden_size // 2) |
| 731 | assert prepared_w13_scale.shape == (num_experts, 128, hidden_size // 16) |
| 732 | assert prepared_w2.shape == (num_experts, hidden_size, 32) |
| 733 | assert prepared_w2_scale.shape == (num_experts, 128, 4) |
| 734 | torch.testing.assert_close(prepared_a13, torch.tensor([2.0, 3.0], device=device)) |
| 735 | torch.testing.assert_close(prepared_a2, torch.tensor([2.0, 3.0], device=device)) |
| 736 | |
| 737 | |
| 738 | def test_b12x_moe_uses_minimax_swiglu_parameters() -> None: |
| 739 | config = make_dummy_moe_config( |
| 740 | hidden_dim=128, |
| 741 | intermediate_size=64, |
| 742 | activation=MoEActivation.SWIGLUOAI_UNINTERLEAVE, |
| 743 | ) |
| 744 | config.swiglu_limit = 7.0 |
| 745 | config.swiglu_alpha = 1.702 |
| 746 | config.swiglu_beta = 1.0 |
| 747 | experts = B12xExperts(config, _quant_config("mxfp4", None)) |
| 748 | |
| 749 | assert experts._swiglu_params(config.activation) == (7.0, 1.702, 1.0) |
| 750 | |
| 751 | |
| 752 | def test_b12x_moe_warmup_runs_each_planner_regime_once( |
| 753 | monkeypatch: pytest.MonkeyPatch, |
| 754 | ) -> None: |
| 755 | experts = B12xExperts( |
| 756 | make_dummy_moe_config( |
| 757 | num_experts=4, |
| 758 | experts_per_token=2, |
| 759 | hidden_dim=128, |
| 760 | intermediate_size=64, |
| 761 | ), |
| 762 | _quant_config("mxfp4", None), |
| 763 | ) |
| 764 | prepared = SimpleNamespace( |
| 765 | num_experts=4, |
| 766 | hidden_size=128, |
| 767 | intermediate_size=64, |
| 768 | w1_fp4=torch.empty(0), |
| 769 | ) |
| 770 | layer = SimpleNamespace( |
| 771 | activation=MoEActivation.SILU, |
| 772 | apply_router_weight_on_input=False, |
| 773 | ) |
| 774 | planned_tokens = [] |
| 775 | launched_tokens = [] |
| 776 | |
| 777 | with pytest.raises(RuntimeError, match="process_weights_after_loading"): |
| 778 | experts.warmup_launches(layer, token_counts=(1,)) |
| 779 | experts._prepared_experts = prepared |
| 780 | |
| 781 | def fake_execution_plan(**kwargs): |
| 782 | tokens = kwargs["tokens"] |
| 783 | if tokens <= 2: |
| 784 | signature = ("micro", "decode") |
| 785 | elif tokens <= 4: |
| 786 | signature = ("dynamic", "small") |
| 787 | else: |
| 788 | signature = ("dynamic", "large") |
| 789 | return SimpleNamespace( |
| 790 | implementation=signature[0], |
| 791 | execution=signature[1], |
| 792 | ) |
| 793 | |
| 794 | def fake_plan(**kwargs): |
| 795 | planned_tokens.append(kwargs["tokens"]) |
| 796 | return SimpleNamespace( |
| 797 | scratch_specs=lambda: [SimpleNamespace(dtype=torch.uint8, shape=(64,))] |
| 798 | ) |
| 799 | |
| 800 | def fake_run(**kwargs): |
| 801 | launched_tokens.append(kwargs["hidden_states"].shape[0]) |
| 802 | |
| 803 | monkeypatch.setattr(b12x, "_b12x_moe_execution_plan", fake_execution_plan) |
| 804 | monkeypatch.setattr(b12x, "_run_b12x_moe_plan", fake_run) |
| 805 | monkeypatch.setattr(experts, "_plan", fake_plan) |
| 806 | |
| 807 | warmed = experts.warmup_launches(layer, token_counts=(1, 2, 3, 4, 8)) |
| 808 | |
| 809 | assert warmed == 3 |
| 810 | assert planned_tokens == [1, 3, 8] |
| 811 | assert launched_tokens == planned_tokens |
| 812 | |
| 813 | |
| 814 | def test_b12x_moe_reload_reprepares_current_parameters( |
| 815 | monkeypatch: pytest.MonkeyPatch, |
| 816 | ) -> None: |
| 817 | plan = SimpleNamespace( |
| 818 | discards_source_parameters=False, |
| 819 | quant_modes=("w4a16",), |
| 820 | io_dtype="bfloat16", |
| 821 | activation="silu", |
| 822 | ) |
| 823 | prepared_inputs: list[ |
| 824 | tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor] |
| 825 | ] = [] |
| 826 | |
| 827 | def prepare_weights(**kwargs): |
| 828 | prepared_inputs.append( |
| 829 | ( |
| 830 | kwargs["w1_fp4"], |
| 831 | kwargs["w2_fp4"], |
| 832 | kwargs["w1_blockscale"], |
| 833 | kwargs["w2_blockscale"], |
| 834 | ) |
| 835 | ) |
| 836 | return SimpleNamespace(plan=plan) |
| 837 | |
| 838 | extension = SimpleNamespace( |
| 839 | plan_weights=lambda **_: plan, prepare_weights=prepare_weights |
| 840 | ) |
| 841 | monkeypatch.setattr(b12x, "_require_b12x_fused_moe", lambda: extension) |
| 842 | experts = B12xExperts( |
| 843 | make_dummy_moe_config(num_experts=2, hidden_dim=4, intermediate_size=8), |
| 844 | _quant_config("mxfp4", None), |
| 845 | ) |
| 846 | layer = SimpleNamespace( |
| 847 | activation=MoEActivation.SILU, |
| 848 | apply_router_weight_on_input=False, |
| 849 | w13_weight=torch.full((2, 8, 2), 1, dtype=torch.uint8), |
| 850 | w2_weight=torch.full((2, 4, 4), 1, dtype=torch.uint8), |
| 851 | w13_weight_scale=torch.full((2, 8, 1), 1, dtype=torch.uint8), |
| 852 | w2_weight_scale=torch.full((2, 4, 1), 1, dtype=torch.uint8), |
| 853 | ) |
| 854 | |
| 855 | experts.process_weights_after_loading(layer) |
| 856 | layer.w13_weight = torch.full_like(layer.w13_weight, 2) |
| 857 | layer.w2_weight = torch.full_like(layer.w2_weight, 2) |
| 858 | layer.w13_weight_scale = torch.full_like(layer.w13_weight_scale, 3) |
| 859 | layer.w2_weight_scale = torch.full_like(layer.w2_weight_scale, 3) |
| 860 | experts.process_weights_after_loading(layer) |
| 861 | |
| 862 | assert len(prepared_inputs) == 2 |
| 863 | assert prepared_inputs[-1][0] is layer.w13_weight |
| 864 | assert prepared_inputs[-1][1] is layer.w2_weight |
| 865 | assert prepared_inputs[-1][2] is layer.w13_weight_scale |
| 866 | assert prepared_inputs[-1][3] is layer.w2_weight_scale |
| 867 | |
| 868 | |
| 869 | def test_b12x_source_release_preserves_prepared_storage_owner() -> None: |
| 870 | layer = torch.nn.Module() |
| 871 | for name, shape in ( |
| 872 | ("w13_weight", (4, 32, 16)), |
| 873 | ("w2_weight", (4, 64, 8)), |
| 874 | ("w13_weight_scale", (4, 32, 2)), |
| 875 | ("w2_weight_scale", (4, 64, 1)), |
| 876 | ): |
| 877 | layer.register_parameter( |
| 878 | name, |
| 879 | torch.nn.Parameter( |
| 880 | torch.empty(shape, dtype=torch.uint8), |
| 881 | requires_grad=False, |
| 882 | ), |
| 883 | ) |
| 884 | experts = B12xExperts( |
| 885 | make_dummy_moe_config(hidden_dim=128, intermediate_size=64), |
| 886 | _quant_config("mxfp4", None), |
| 887 | ) |
| 888 | owner = SimpleNamespace( |
| 889 | w1_fp4=layer.w13_weight, |
| 890 | w2_fp4=layer.w2_weight, |
| 891 | w1_blockscale=layer.w13_weight_scale, |
| 892 | w2_blockscale=layer.w2_weight_scale, |
| 893 | ) |
| 894 | experts._prepared_experts = owner |
| 895 | owner_tensors = ( |
| 896 | owner.w1_fp4, |
| 897 | owner.w2_fp4, |
| 898 | owner.w1_blockscale, |
| 899 | owner.w2_blockscale, |
| 900 | ) |
| 901 | owner_ptrs = tuple(tensor.untyped_storage().data_ptr() for tensor in owner_tensors) |
| 902 | |
| 903 | experts._release_source_parameters(layer) |
| 904 | experts._release_source_parameters(layer) |
| 905 | |
| 906 | assert layer.w13_weight.numel() == 0 |
| 907 | assert layer.w2_weight.numel() == 0 |
| 908 | assert layer.w13_weight_scale.numel() == 0 |
| 909 | assert layer.w2_weight_scale.numel() == 0 |
| 910 | assert ( |
| 911 | tuple(tensor.untyped_storage().data_ptr() for tensor in owner_tensors) |
| 912 | == owner_ptrs |
| 913 | ) |
| 914 | |
| 915 | |
| 916 | def test_b12x_moe_rejects_router_weight_on_input_for_w4a8() -> None: |
| 917 | experts = B12xExperts( |
| 918 | make_dummy_moe_config(hidden_dim=256, intermediate_size=64), |
| 919 | _quant_config("mxfp4", "mxfp8"), |
| 920 | ) |
| 921 | layer = SimpleNamespace( |
| 922 | activation=MoEActivation.SILU, |
| 923 | apply_router_weight_on_input=True, |
| 924 | ) |
| 925 | |
| 926 | with pytest.raises( |
| 927 | ValueError, |
| 928 | match="apply_router_weight_on_input only with W4A16", |
| 929 | ): |
| 930 | experts.process_weights_after_loading(layer) |
| 931 | |
| 932 | |
| 933 | def test_b12x_moe_workspace_uses_prepared_router_weight_contract( |
| 934 | monkeypatch: pytest.MonkeyPatch, |
| 935 | ) -> None: |
| 936 | experts = B12xExperts( |
| 937 | make_dummy_moe_config(hidden_dim=128, intermediate_size=64), |
| 938 | _quant_config("mxfp4", None), |
| 939 | ) |
| 940 | prepared = SimpleNamespace( |
| 941 | plan=SimpleNamespace(discards_source_parameters=False), |
| 942 | ) |
| 943 | layer = SimpleNamespace( |
| 944 | activation=MoEActivation.SILU, |
| 945 | apply_router_weight_on_input=True, |
| 946 | w13_weight=torch.empty(0), |
| 947 | w2_weight=torch.empty(0), |
| 948 | w13_weight_scale=torch.empty(0), |
| 949 | w2_weight_scale=torch.empty(0), |
| 950 | ) |
| 951 | monkeypatch.setattr(experts, "_prepare_experts", lambda **kwargs: prepared) |
| 952 | planned = [] |
| 953 | |
| 954 | def fake_plan(**kwargs): |
| 955 | planned.append(kwargs) |
| 956 | return SimpleNamespace( |
| 957 | scratch_specs=lambda: [SimpleNamespace(dtype=torch.uint8, shape=(64,))] |
| 958 | ) |
| 959 | |
| 960 | monkeypatch.setattr(experts, "_plan", fake_plan) |
| 961 | |
| 962 | experts.process_weights_after_loading(layer) |
| 963 | assert layer.b12x_warmup_provider is experts |
| 964 | experts.workspace_shapes( |
| 965 | 8, |
| 966 | 128, |
| 967 | 128, |
| 968 | 2, |
| 969 | 4, |
| 970 | 4, |
| 971 | None, |
| 972 | MoEActivation.SILU, |
| 973 | ) |
| 974 | |
| 975 | assert planned == [ |
| 976 | { |
| 977 | "tokens": 8, |
| 978 | "topk": 2, |
| 979 | "activation": MoEActivation.SILU, |
| 980 | "apply_router_weight_on_input": True, |
| 981 | } |
| 982 | ] |
| 983 | |
| 984 | |
| 985 | @dataclass |
| 986 | class _B12xMoeCase: |
| 987 | hidden_states: torch.Tensor |
| 988 | score: torch.Tensor |
| 989 | w1: torch.Tensor |
| 990 | w2: torch.Tensor |
| 991 | w1_ref: torch.Tensor |
| 992 | w2_ref: torch.Tensor |
| 993 | quant_config: FusedMoEQuantConfig |
| 994 | activation: MoEActivation |
| 995 | activation_dtype: str | None |
| 996 | topk: int = 2 |
| 997 | |
| 998 | |
| 999 | def _make_b12x_moe_case( |
| 1000 | weight_dtype: str, |
| 1001 | activation_dtype: str | None, |
| 1002 | *, |
| 1003 | activation: MoEActivation = MoEActivation.SILU, |
| 1004 | tokens: int = 16, |
| 1005 | seed: int = 19, |
| 1006 | ) -> _B12xMoeCase: |
| 1007 | set_random_seed(seed) |
| 1008 | num_experts, hidden_size, intermediate_size = 4, 512, 128 |
| 1009 | dtype = torch.bfloat16 |
| 1010 | hidden_states = torch.randn((tokens, hidden_size), device="cuda", dtype=dtype) / 10 |
| 1011 | w1_rows = 2 * intermediate_size if activation.is_gated else intermediate_size |
| 1012 | w1 = ( |
| 1013 | torch.randn( |
| 1014 | (num_experts, w1_rows, hidden_size), |
| 1015 | device="cuda", |
| 1016 | dtype=dtype, |
| 1017 | ) |
| 1018 | / 15 |
| 1019 | ) |
| 1020 | w2 = ( |
| 1021 | torch.randn( |
| 1022 | (num_experts, hidden_size, intermediate_size), |
| 1023 | device="cuda", |
| 1024 | dtype=dtype, |
| 1025 | ) |
| 1026 | / 15 |
| 1027 | ) |
| 1028 | |
| 1029 | if weight_dtype == "mxfp4": |
| 1030 | w1_q, w1_scale = mxfp4_quantize(w1) |
| 1031 | w2_q, w2_scale = mxfp4_quantize(w2) |
| 1032 | w1_ref = torch.stack( |
| 1033 | [dq_mxfp4_torch(w1_q[e], w1_scale[e], dtype) for e in range(num_experts)] |
| 1034 | ) |
| 1035 | w2_ref = torch.stack( |
| 1036 | [dq_mxfp4_torch(w2_q[e], w2_scale[e], dtype) for e in range(num_experts)] |
| 1037 | ) |
| 1038 | if activation_dtype is None: |
| 1039 | quant_config = mxfp4_w4a16_moe_quant_config( |
| 1040 | w1_scale=w1_scale, |
| 1041 | w2_scale=w2_scale, |
| 1042 | ) |
| 1043 | else: |
| 1044 | quant_config = FusedMoEQuantConfig.make( |
| 1045 | quant_dtype=activation_dtype, |
| 1046 | weight_dtype=weight_dtype, |
| 1047 | w1_scale=w1_scale, |
| 1048 | w2_scale=w2_scale, |
| 1049 | ) |
| 1050 | elif weight_dtype == "nvfp4": |
| 1051 | w1_q, w1_scale, w1_global_scale = _quantize_nvfp4_linear(w1) |
| 1052 | w2_q, w2_scale, w2_global_scale = _quantize_nvfp4_linear(w2) |
| 1053 | w1_ref = torch.stack( |
| 1054 | [ |
| 1055 | _dequantize_nvfp4_linear( |
| 1056 | w1_q[e], w1_scale[e], w1_global_scale[e], dtype |
| 1057 | ) |
| 1058 | for e in range(num_experts) |
| 1059 | ] |
| 1060 | ) |
| 1061 | w2_ref = torch.stack( |
| 1062 | [ |
| 1063 | _dequantize_nvfp4_linear( |
| 1064 | w2_q[e], w2_scale[e], w2_global_scale[e], dtype |
| 1065 | ) |
| 1066 | for e in range(num_experts) |
| 1067 | ] |
| 1068 | ) |
| 1069 | input_scale = torch.full( |
| 1070 | (num_experts,), |
| 1071 | 1.0 if activation_dtype is None else 1.0 / 1024.0, |
| 1072 | device="cuda", |
| 1073 | dtype=torch.float32, |
| 1074 | ) |
| 1075 | prepared = prepare_nvfp4_moe_layer_for_b12x( |
| 1076 | w1_q, |
| 1077 | w1_scale, |
| 1078 | 1.0 / w1_global_scale, |
| 1079 | input_scale, |
| 1080 | w2_q, |
| 1081 | w2_scale, |
| 1082 | 1.0 / w2_global_scale, |
| 1083 | input_scale, |
| 1084 | is_act_and_mul=activation.is_gated, |
| 1085 | reorder_w13=activation_dtype is None and activation.is_gated, |
| 1086 | ) |
| 1087 | w1_q, w1_scale, w1_alpha, a1_scale = prepared[:4] |
| 1088 | w2_q, w2_scale, w2_alpha, a2_scale = prepared[4:] |
| 1089 | if activation_dtype is None: |
| 1090 | quant_config = nvfp4_w4a16_moe_quant_config( |
| 1091 | g1_alphas=w1_alpha, |
| 1092 | g2_alphas=w2_alpha, |
| 1093 | w1_scale=w1_scale, |
| 1094 | w2_scale=w2_scale, |
| 1095 | ) |
| 1096 | else: |
| 1097 | quant_config = FusedMoEQuantConfig.make( |
| 1098 | quant_dtype=activation_dtype, |
| 1099 | weight_dtype=weight_dtype, |
| 1100 | w1_scale=w1_scale, |
| 1101 | w2_scale=w2_scale, |
| 1102 | g1_alphas=w1_alpha, |
| 1103 | g2_alphas=w2_alpha, |
| 1104 | a1_gscale=1.0 / a1_scale, |
| 1105 | a2_gscale=1.0 / a2_scale, |
| 1106 | ) |
| 1107 | else: |
| 1108 | raise ValueError(f"unsupported test weight dtype {weight_dtype}") |
| 1109 | |
| 1110 | return _B12xMoeCase( |
| 1111 | hidden_states=hidden_states, |
| 1112 | score=torch.randn((tokens, num_experts), device="cuda", dtype=dtype), |
| 1113 | w1=w1_q, |
| 1114 | w2=w2_q, |
| 1115 | w1_ref=w1_ref, |
| 1116 | w2_ref=w2_ref, |
| 1117 | quant_config=quant_config, |
| 1118 | activation=activation, |
| 1119 | activation_dtype=activation_dtype, |
| 1120 | ) |
| 1121 | |
| 1122 | |
| 1123 | @pytest.mark.skipif(not _has_b12x_moe(), reason="requires b12x MoE on SM120") |
| 1124 | @pytest.mark.parametrize( |
| 1125 | "weight_dtype,activation_dtype,activation", |
| 1126 | [ |
| 1127 | pytest.param("mxfp4", "mxfp8", MoEActivation.SILU, id="mxfp4-mxfp8"), |
| 1128 | pytest.param("mxfp4", None, MoEActivation.SILU, id="mxfp4-bf16"), |
| 1129 | pytest.param("nvfp4", "nvfp4", MoEActivation.SILU, id="nvfp4-nvfp4"), |
| 1130 | pytest.param("nvfp4", "mxfp8", MoEActivation.SILU, id="nvfp4-mxfp8"), |
| 1131 | pytest.param("nvfp4", None, MoEActivation.SILU, id="nvfp4-bf16-silu"), |
| 1132 | pytest.param( |
| 1133 | "nvfp4", |
| 1134 | None, |
| 1135 | MoEActivation.RELU2_NO_MUL, |
| 1136 | id="nvfp4-bf16-relu2", |
| 1137 | ), |
| 1138 | ], |
| 1139 | ) |
| 1140 | @torch.inference_mode() |
| 1141 | def test_b12x_moe_matches_torch( |
| 1142 | weight_dtype: str, |
| 1143 | activation_dtype: str | None, |
| 1144 | activation: MoEActivation, |
| 1145 | workspace_init, |
| 1146 | ) -> None: |
| 1147 | with set_current_vllm_config( |
| 1148 | VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) |
| 1149 | ): |
| 1150 | case = _make_b12x_moe_case( |
| 1151 | weight_dtype, |
| 1152 | activation_dtype, |
| 1153 | activation=activation, |
| 1154 | ) |
| 1155 | reference = torch_moe( |
| 1156 | case.hidden_states, |
| 1157 | case.w1_ref, |
| 1158 | case.w2_ref, |
| 1159 | case.score, |
| 1160 | case.topk, |
| 1161 | activation=case.activation, |
| 1162 | ) |
| 1163 | if activation_dtype == "nvfp4": |
| 1164 | topk_weights, topk_ids, _ = fused_topk( |
| 1165 | case.hidden_states, |
| 1166 | case.score, |
| 1167 | case.topk, |
| 1168 | renormalize=False, |
| 1169 | ) |
| 1170 | reference = _nvfp4_activation_reference( |
| 1171 | case.hidden_states, |
| 1172 | case.w1_ref, |
| 1173 | case.w2_ref, |
| 1174 | topk_weights, |
| 1175 | topk_ids, |
| 1176 | case.quant_config.a1_gscale, |
| 1177 | case.quant_config.a2_gscale, |
| 1178 | ) |
| 1179 | |
| 1180 | checks_zero_canonicalization = ( |
| 1181 | weight_dtype == "nvfp4" and activation_dtype is None |
| 1182 | ) |
| 1183 | if checks_zero_canonicalization: |
| 1184 | assert _count_fp4_negative_zeros(case.w1) > 0 |
| 1185 | assert _count_fp4_negative_zeros(case.w2) > 0 |
| 1186 | output = _run_b12x_moe( |
| 1187 | case.hidden_states, |
| 1188 | case.w1, |
| 1189 | case.w2, |
| 1190 | case.score, |
| 1191 | case.topk, |
| 1192 | case.activation, |
| 1193 | case.quant_config, |
| 1194 | ) |
| 1195 | if checks_zero_canonicalization: |
| 1196 | assert _count_fp4_negative_zeros(case.w1) == 0 |
| 1197 | assert _count_fp4_negative_zeros(case.w2) == 0 |
| 1198 | |
| 1199 | torch.testing.assert_close(output, reference, atol=2e-1, rtol=2e-1) |
| 1200 | cosine = torch.nn.functional.cosine_similarity( |
| 1201 | output.flatten().float(), |
| 1202 | reference.flatten().float(), |
| 1203 | dim=0, |
| 1204 | ) |
| 1205 | assert cosine > 0.99 |
| 1206 | |
| 1207 | |
| 1208 | @pytest.mark.skipif(not _has_b12x_moe(), reason="requires b12x MoE on SM120") |
| 1209 | @pytest.mark.parametrize( |
| 1210 | "weight_dtype,activation_dtype", |
| 1211 | [ |
| 1212 | pytest.param("mxfp4", "mxfp8", id="w4a8"), |
| 1213 | pytest.param("nvfp4", None, id="w4a16"), |
| 1214 | ], |
| 1215 | ) |
| 1216 | @torch.inference_mode() |
| 1217 | def test_b12x_moe_cuda_graph_replay( |
| 1218 | weight_dtype: str, |
| 1219 | activation_dtype: str | None, |
| 1220 | workspace_init, |
| 1221 | ) -> None: |
| 1222 | from vllm.v1.worker.workspace import lock_workspace |
| 1223 | |
| 1224 | with set_current_vllm_config( |
| 1225 | VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1)) |
| 1226 | ): |
| 1227 | case = _make_b12x_moe_case( |
| 1228 | weight_dtype, |
| 1229 | activation_dtype, |
| 1230 | tokens=4, |
| 1231 | seed=23, |
| 1232 | ) |
| 1233 | kernel = _make_b12x_moe_kernel( |
| 1234 | case.hidden_states, |
| 1235 | case.w1, |
| 1236 | case.w2, |
| 1237 | case.topk, |
| 1238 | case.activation, |
| 1239 | case.quant_config, |
| 1240 | ) |
| 1241 | topk_weights, topk_ids, _ = fused_topk( |
| 1242 | case.hidden_states, |
| 1243 | case.score, |
| 1244 | case.topk, |
| 1245 | renormalize=False, |
| 1246 | ) |
| 1247 | assert topk_weights.dtype == torch.float32 and topk_weights.is_contiguous() |
| 1248 | assert topk_ids.dtype == torch.int32 and topk_ids.is_contiguous() |
| 1249 | |
| 1250 | def apply() -> torch.Tensor: |
| 1251 | return kernel.apply( |
| 1252 | hidden_states=case.hidden_states, |
| 1253 | w1=case.w1, |
| 1254 | w2=case.w2, |
| 1255 | topk_weights=topk_weights, |
| 1256 | topk_ids=topk_ids, |
| 1257 | activation=case.activation, |
| 1258 | global_num_experts=case.w1.shape[0], |
| 1259 | expert_map=None, |
| 1260 | apply_router_weight_on_input=False, |
| 1261 | ) |
| 1262 | |
| 1263 | expected = apply().clone() |
| 1264 | lock_workspace() |
| 1265 | graph = torch.cuda.CUDAGraph() |
| 1266 | stream = torch.cuda.Stream() |
| 1267 | with torch.cuda.graph(graph, stream=stream): |
| 1268 | actual = apply() |
| 1269 | graph.replay() |
| 1270 | torch.accelerator.synchronize() |
| 1271 | |
| 1272 | assert torch.isfinite(expected).all() |
| 1273 | assert torch.isfinite(actual).all() |
| 1274 | torch.testing.assert_close(actual, expected, atol=2e-2, rtol=2e-2) |