| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Callable, Iterable |
| 4 | from contextlib import nullcontext |
| 5 | from typing import TYPE_CHECKING, cast |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn.functional as F |
| 9 | |
| 10 | from vllm.config import VllmConfig, get_current_vllm_config |
| 11 | from vllm.config.parallel import ExpertPlacementStrategy |
| 12 | from vllm.distributed import ( |
| 13 | get_ep_group, |
| 14 | get_pcp_group, |
| 15 | tensor_model_parallel_all_reduce, |
| 16 | ) |
| 17 | from vllm.distributed.eplb.eplb_state import EplbLayerState |
| 18 | from vllm.forward_context import ( |
| 19 | ForwardContext, |
| 20 | get_forward_context, |
| 21 | is_forward_context_available, |
| 22 | ) |
| 23 | from vllm.logger import init_logger |
| 24 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 25 | from vllm.model_executor.layers.fused_moe.config import ( |
| 26 | FusedMoEConfig, |
| 27 | ) |
| 28 | from vllm.model_executor.layers.fused_moe.fused_moe_method_base import ( |
| 29 | FusedMoEMethodBase, |
| 30 | ) |
| 31 | from vllm.model_executor.layers.fused_moe.moe_output import UnfinalizedMoEOutput |
| 32 | from vllm.model_executor.layers.fused_moe.routed_experts import ( |
| 33 | RoutedExperts, |
| 34 | ) |
| 35 | from vllm.model_executor.layers.fused_moe.router.fused_moe_router import ( |
| 36 | FusedMoERouter, |
| 37 | ) |
| 38 | from vllm.model_executor.layers.fused_moe.router.zero_expert_router import ( |
| 39 | ZeroExpertRouter, |
| 40 | ) |
| 41 | from vllm.model_executor.layers.fused_moe.runner.moe_runner_interface import ( |
| 42 | MoERunnerInterface, |
| 43 | ) |
| 44 | from vllm.model_executor.layers.fused_moe.runner.shared_experts import ( |
| 45 | SharedExperts, |
| 46 | SharedExpertsOrder, |
| 47 | ) |
| 48 | from vllm.model_executor.layers.utils import dispatch_unquantized_gemm |
| 49 | from vllm.platforms import current_platform |
| 50 | from vllm.utils.torch_utils import ( |
| 51 | _USE_LAYERNAME, |
| 52 | LayerName, |
| 53 | direct_register_custom_op, |
| 54 | ) |
| 55 | |
| 56 | logger = init_logger(__name__) |
| 57 | |
| 58 | |
| 59 | def register_layer_for_moe_forward_op( |
| 60 | vllm_config: VllmConfig, |
| 61 | layer: "MoERunner", |
| 62 | ): |
| 63 | # For smuggling this layer into the fused moe custom op |
| 64 | prefix = layer.layer_name |
| 65 | compilation_config = vllm_config.compilation_config |
| 66 | if prefix in compilation_config.static_forward_context: |
| 67 | raise ValueError("Duplicate layer name: {}".format(prefix)) |
| 68 | compilation_config.static_forward_context[prefix] = layer |
| 69 | compilation_config.static_all_moe_layers.append(prefix) |
| 70 | |
| 71 | |
| 72 | def get_layer_from_name(layer_name: str) -> MoERunnerInterface: |
| 73 | forward_context: ForwardContext = get_forward_context() |
| 74 | if not _USE_LAYERNAME and layer_name == "from_forward_context": |
| 75 | all_moe_layers = forward_context.all_moe_layers |
| 76 | assert all_moe_layers is not None |
| 77 | moe_layer_index = forward_context.moe_layer_index |
| 78 | if moe_layer_index >= len(all_moe_layers): |
| 79 | raise AssertionError( |
| 80 | "We expected the number of MOE layers in `all_moe_layers` " |
| 81 | "to be equal to the number of " |
| 82 | "{vllm.moe_forward, vllm.moe_forward_shared} calls." |
| 83 | ) |
| 84 | layer_name = all_moe_layers[moe_layer_index] |
| 85 | forward_context.moe_layer_index += 1 |
| 86 | layer = forward_context.no_compile_layers[layer_name] |
| 87 | assert isinstance(layer, MoERunnerInterface) |
| 88 | return layer |
| 89 | |
| 90 | |
| 91 | # On torch >= 2.11, layer_name is a hoisted LayerName opaque object; |
| 92 | # on older versions it remains a plain str. |
| 93 | if TYPE_CHECKING: |
| 94 | from typing import TypeAlias |
| 95 | |
| 96 | _layer_name_type: TypeAlias = str | LayerName |
| 97 | else: |
| 98 | _layer_name_type = LayerName if _USE_LAYERNAME else str |
| 99 | |
| 100 | |
| 101 | @torch.compiler.assume_constant_result |
| 102 | def _resolve_layer_name(layer_name: str | LayerName) -> str: |
| 103 | from torch._library.fake_class_registry import FakeScriptObject |
| 104 | |
| 105 | if isinstance(layer_name, LayerName): |
| 106 | return layer_name.value |
| 107 | elif isinstance(layer_name, FakeScriptObject): |
| 108 | return layer_name.real_obj.value |
| 109 | return layer_name |
| 110 | |
| 111 | |
| 112 | # Note: _moe_forward and _moe_forward_shared should not contain any |
| 113 | # implementation details, They should merely pass along control to |
| 114 | # the runner's '_forward_impl' method. |
| 115 | # These functions should never be called directly since they do not |
| 116 | # include all the functionality of the MoE layer. |
| 117 | def _moe_forward( |
| 118 | hidden_states: torch.Tensor, |
| 119 | router_logits: torch.Tensor, |
| 120 | shared_experts_input: torch.Tensor | None, |
| 121 | input_ids: torch.Tensor | None, |
| 122 | layer_name: _layer_name_type, |
| 123 | hidden_dim_unpadded: int, |
| 124 | ) -> torch.Tensor: |
| 125 | layer = get_layer_from_name(_resolve_layer_name(layer_name)) |
| 126 | return cast( |
| 127 | torch.Tensor, |
| 128 | layer._forward_impl( |
| 129 | hidden_states, |
| 130 | router_logits, |
| 131 | shared_experts_input, |
| 132 | input_ids, |
| 133 | ), |
| 134 | ) |
| 135 | |
| 136 | |
| 137 | def _moe_forward_fake( |
| 138 | hidden_states: torch.Tensor, |
| 139 | router_logits: torch.Tensor, |
| 140 | shared_experts_input: torch.Tensor | None, |
| 141 | input_ids: torch.Tensor | None, |
| 142 | layer_name: _layer_name_type, |
| 143 | hidden_dim_unpadded: int, |
| 144 | ) -> torch.Tensor: |
| 145 | # `hidden_dim_unpadded > 0` only on the TRT-LLM MXFP4 path, where the |
| 146 | # real kernel writes narrower than `hidden_states.shape[-1]`. Plumbed |
| 147 | # as an op arg (not peeked from the layer registry) to keep the fake |
| 148 | # a pure shape function of its inputs and preserve subgraph dedup. |
| 149 | if hidden_dim_unpadded > 0: |
| 150 | return hidden_states.new_empty((*hidden_states.shape[:-1], hidden_dim_unpadded)) |
| 151 | return torch.empty_like(hidden_states) |
| 152 | |
| 153 | |
| 154 | def _moe_forward_shared( |
| 155 | hidden_states: torch.Tensor, |
| 156 | router_logits: torch.Tensor, |
| 157 | shared_experts_input: torch.Tensor | None, |
| 158 | input_ids: torch.Tensor | None, |
| 159 | layer_name: _layer_name_type, |
| 160 | hidden_dim_unpadded: int, |
| 161 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 162 | layer = get_layer_from_name(_resolve_layer_name(layer_name)) |
| 163 | return cast( |
| 164 | tuple[torch.Tensor, torch.Tensor], |
| 165 | layer._forward_impl( |
| 166 | hidden_states, |
| 167 | router_logits, |
| 168 | shared_experts_input, |
| 169 | input_ids, |
| 170 | ), |
| 171 | ) |
| 172 | |
| 173 | |
| 174 | def _moe_forward_shared_fake( |
| 175 | hidden_states: torch.Tensor, |
| 176 | router_logits: torch.Tensor, |
| 177 | shared_experts_input: torch.Tensor | None, |
| 178 | input_ids: torch.Tensor | None, |
| 179 | layer_name: _layer_name_type, |
| 180 | hidden_dim_unpadded: int, |
| 181 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 182 | # `fused_out`: see `_moe_forward_fake` for hidden_dim_unpadded semantics. |
| 183 | # `shared_out`: matches `shared_experts_input` if provided (latent MoE), |
| 184 | # else `hidden_states`. |
| 185 | if hidden_dim_unpadded > 0: |
| 186 | fused_out = hidden_states.new_empty( |
| 187 | (*hidden_states.shape[:-1], hidden_dim_unpadded) |
| 188 | ) |
| 189 | else: |
| 190 | fused_out = torch.empty_like(hidden_states) |
| 191 | if shared_experts_input is not None: |
| 192 | shared_out = torch.empty_like(shared_experts_input) |
| 193 | else: |
| 194 | shared_out = torch.empty_like(hidden_states) |
| 195 | return shared_out, fused_out |
| 196 | |
| 197 | |
| 198 | # NOTE: `moe_forward` and `moe_forward_shared` being opaque custom ops is a |
| 199 | # load-bearing assumption for the MoE-LoRA dual-stream path. |
| 200 | direct_register_custom_op( |
| 201 | op_name="moe_forward", |
| 202 | op_func=_moe_forward, |
| 203 | mutates_args=["hidden_states"], |
| 204 | fake_impl=_moe_forward_fake, |
| 205 | tags=(torch.Tag.needs_fixed_stride_order,), |
| 206 | ) |
| 207 | |
| 208 | |
| 209 | direct_register_custom_op( |
| 210 | op_name="moe_forward_shared", |
| 211 | op_func=_moe_forward_shared, |
| 212 | fake_impl=_moe_forward_shared_fake, |
| 213 | tags=(torch.Tag.needs_fixed_stride_order,), |
| 214 | ) |
| 215 | |
| 216 | |
| 217 | def _unpack( |
| 218 | result: torch.Tensor |
| 219 | | UnfinalizedMoEOutput |
| 220 | | tuple[torch.Tensor, torch.Tensor | UnfinalizedMoEOutput], |
| 221 | ) -> tuple[torch.Tensor | None, torch.Tensor | UnfinalizedMoEOutput]: |
| 222 | if isinstance(result, tuple): |
| 223 | return result |
| 224 | else: |
| 225 | return (None, result) |
| 226 | |
| 227 | |
| 228 | class MoERunner(MoERunnerInterface): |
| 229 | """Standard MoE runner implementation for executing Mixture of Experts layers. |
| 230 | |
| 231 | This is the primary concrete implementation of MoE execution logic, providing |
| 232 | comprehensive support for standard MoE operations. It handles: |
| 233 | - Expert routing and token dispatching using various routing strategies |
| 234 | - Shared experts computation with optional parallel execution using CUDA streams |
| 235 | - Tensor model parallel and expert parallel operations |
| 236 | - Multiple quantization methods and optimized kernel selection |
| 237 | - Both monolithic and decomposed expert execution paths |
| 238 | - Integration with various parallel execution modes (TP, EP, DP) |
| 239 | |
| 240 | The runner orchestrates the complete MoE forward pass including routing tokens |
| 241 | to experts, executing expert computations in parallel, and combining results. |
| 242 | It supports advanced features like overlapped execution of shared experts, |
| 243 | optimized kernels for different parallel configurations, and seamless |
| 244 | integration with vLLM's distributed execution framework. |
| 245 | |
| 246 | Eventually, this class may be split into more specialized implementations |
| 247 | for different configurations (e.g., with/without shared experts, gates, etc.). |
| 248 | """ |
| 249 | |
| 250 | def __init__( |
| 251 | self, |
| 252 | layer_name: str, |
| 253 | moe_config: FusedMoEConfig, |
| 254 | router: FusedMoERouter, |
| 255 | routed_experts: RoutedExperts, |
| 256 | enable_dbo: bool = False, |
| 257 | gate: torch.nn.Module | None = None, |
| 258 | shared_experts: torch.nn.Module | None = None, |
| 259 | shared_expert_gate: torch.nn.Module | None = None, |
| 260 | routed_input_transform: torch.nn.Module | None = None, |
| 261 | routed_output_transform: torch.nn.Module | None = None, |
| 262 | routed_scaling_factor: float = 1.0, |
| 263 | ): |
| 264 | super().__init__() |
| 265 | self.moe_config = moe_config |
| 266 | self.router = router |
| 267 | self.routed_input_transform = routed_input_transform |
| 268 | self.routed_output_transform = routed_output_transform |
| 269 | self.routed_scaling_factor = routed_scaling_factor |
| 270 | self.gate = gate |
| 271 | self.shared_expert_gate = shared_expert_gate |
| 272 | self.routed_experts = routed_experts |
| 273 | self.enable_dbo = enable_dbo |
| 274 | |
| 275 | # When both gates are present and FSE is enabled, fuse their |
| 276 | # weight matrices into [num_experts + num_shared, hidden] so one |
| 277 | # GEMM produces combined logits. The topk kernel can then |
| 278 | # apply routing softmax and shared expert activation (sigmoid) |
| 279 | # in a single launch. |
| 280 | self._fse_fuse_gate = gate is not None and shared_expert_gate is not None |
| 281 | self._combined_gate_weight: torch.Tensor | None = None |
| 282 | |
| 283 | self._shared_experts: SharedExperts | None = None |
| 284 | if shared_experts is not None: |
| 285 | can_overlap = lambda: self._quant_method.mk_can_overlap_shared_experts |
| 286 | self._shared_experts = SharedExperts( |
| 287 | shared_experts, |
| 288 | moe_config=moe_config, |
| 289 | enable_dbo=enable_dbo, |
| 290 | mk_can_overlap_shared_experts=can_overlap, |
| 291 | ) |
| 292 | |
| 293 | # Needed for string -> MoERunner layer lookup in custom ops. |
| 294 | self.layer_name = layer_name |
| 295 | |
| 296 | self._forward_entry = self._select_forward() |
| 297 | |
| 298 | # For smuggling this layer into the fused moe custom op |
| 299 | register_layer_for_moe_forward_op(get_current_vllm_config(), self) |
| 300 | |
| 301 | def load_weights( |
| 302 | self, weights: Iterable[tuple[str, torch.Tensor]] |
| 303 | ) -> Iterable[str]: |
| 304 | return self.routed_experts.load_weights(weights) |
| 305 | |
| 306 | def _select_forward(self) -> Callable: |
| 307 | if current_platform.is_tpu(): |
| 308 | # TODO: Once the OOM issue for the TPU backend is resolved, we |
| 309 | # will switch to using the moe_forward custom op. |
| 310 | return _moe_forward if self._shared_experts is None else _moe_forward_shared |
| 311 | |
| 312 | if current_platform.is_cpu(): |
| 313 | # CPU never touches the workspace manager (Monolithic experts |
| 314 | # skip it entirely; Modular experts' _allocate_buffers bypasses |
| 315 | # it too, see modular_kernel.py) -- the ContextVar-based lane |
| 316 | # lookup was the only part of this call graph Dynamo can't |
| 317 | # trace, so CPU can always call the fused-MoE op directly. |
| 318 | return _moe_forward if self._shared_experts is None else _moe_forward_shared |
| 319 | |
| 320 | return ( |
| 321 | torch.ops.vllm.moe_forward |
| 322 | if self._shared_experts is None |
| 323 | else torch.ops.vllm.moe_forward_shared |
| 324 | ) |
| 325 | |
| 326 | @property |
| 327 | def shared_experts(self) -> SharedExperts | None: |
| 328 | return self._shared_experts |
| 329 | |
| 330 | # TODO(bnell): Temporary hack. Get rid of this. |
| 331 | def _replace_quant_method(self, quant_method: FusedMoEMethodBase): |
| 332 | self.routed_experts._replace_quant_method(quant_method) |
| 333 | |
| 334 | # TODO(bnell): Hack for elastic_ep. Get rid of this |
| 335 | def _set_moe_config(self, new_moe_config: FusedMoEConfig): |
| 336 | self.moe_config = new_moe_config |
| 337 | self.routed_experts._set_moe_config(new_moe_config) |
| 338 | if self._shared_experts is not None: |
| 339 | self._shared_experts._set_moe_config(new_moe_config) |
| 340 | |
| 341 | def _maybe_fuse_gate_weights(self): |
| 342 | """Fuse router and shared expert gate weights on first call. |
| 343 | |
| 344 | Cannot be done at __init__ because gate weights are loaded after |
| 345 | module construction (via weight_loader). Called once from |
| 346 | _forward_impl before the first forward pass. |
| 347 | """ |
| 348 | if self._combined_gate_weight is None: |
| 349 | assert self.gate is not None and self.shared_expert_gate is not None |
| 350 | self._combined_gate_weight = torch.cat( |
| 351 | [self.gate.weight, self.shared_expert_gate.weight], |
| 352 | dim=0, |
| 353 | ) |
| 354 | |
| 355 | @property |
| 356 | def _quant_method(self) -> FusedMoEMethodBase: |
| 357 | return self.routed_experts.quant_method |
| 358 | |
| 359 | def apply_routed_input_transform( |
| 360 | self, |
| 361 | hidden_states: torch.Tensor, |
| 362 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 363 | """Apply transform for routed experts (e.g., latent projection). |
| 364 | |
| 365 | This is called by MoERunner.forward_native. The original hidden_states |
| 366 | is saved separately so shared experts get [S, hidden_size] while |
| 367 | routed experts get the transformed [S, moe_latent_size]. |
| 368 | |
| 369 | Returns (possibly transformed) hidden states and the input for shared |
| 370 | experts (or None if there are no shared experts). |
| 371 | """ |
| 372 | if self.routed_input_transform is not None: |
| 373 | result = self.routed_input_transform(hidden_states) |
| 374 | # ReplicatedLinear returns (output, extra_bias) tuple. |
| 375 | # We only need the output tensor; extra_bias is not used here. |
| 376 | if isinstance(result, tuple): |
| 377 | return result[0], hidden_states |
| 378 | return result, hidden_states |
| 379 | |
| 380 | return ( |
| 381 | hidden_states, |
| 382 | hidden_states if self._shared_experts is not None else None, |
| 383 | ) |
| 384 | |
| 385 | def apply_routed_output_transform( |
| 386 | self, |
| 387 | fused_output: torch.Tensor, |
| 388 | ) -> torch.Tensor: |
| 389 | """Apply transform to routed expert output (e.g., latent to full dim). |
| 390 | |
| 391 | Used by latent MoE models (e.g., NemotronH) where routed experts |
| 392 | operate in a compressed latent space and need projection back to |
| 393 | the full hidden dimension before combining with shared expert output. |
| 394 | """ |
| 395 | if self.routed_output_transform is not None: |
| 396 | r = self.routed_output_transform(fused_output) |
| 397 | fused_output = r[0] if isinstance(r, tuple) else r |
| 398 | return fused_output |
| 399 | |
| 400 | def _maybe_apply_routed_scale_to_output( |
| 401 | self, |
| 402 | shared_output: torch.Tensor | None, |
| 403 | fused_output: torch.Tensor, |
| 404 | ) -> tuple[torch.Tensor | None, torch.Tensor]: |
| 405 | """Apply routed_scaling_factor to the output with FP16 overflow |
| 406 | protection. |
| 407 | |
| 408 | Scale the fused expert output by routed_scaling_factor. For FP16, |
| 409 | avoid overflow by dividing shared_output by the scale instead |
| 410 | (the decoder layer compensates with matching divisions). |
| 411 | """ |
| 412 | if self.routed_scaling_factor != 1.0: |
| 413 | if fused_output.dtype != torch.float16 or shared_output is None: |
| 414 | fused_output *= self.routed_scaling_factor |
| 415 | elif shared_output is not None: |
| 416 | shared_output *= 1.0 / self.routed_scaling_factor |
| 417 | return shared_output, fused_output |
| 418 | |
| 419 | @property |
| 420 | def _fused_output_is_reduced(self) -> bool: |
| 421 | return ( |
| 422 | self._quant_method.moe_kernel is not None |
| 423 | and self._quant_method.moe_kernel.output_is_reduced() |
| 424 | ) |
| 425 | |
| 426 | def _maybe_reduce_shared_expert_output( |
| 427 | self, |
| 428 | shared_output: torch.Tensor | None, |
| 429 | fused_output_is_reduced: bool | None = None, |
| 430 | ) -> torch.Tensor | None: |
| 431 | """All-reduce shared expert output when the combine kernel already |
| 432 | reduced fused output. |
| 433 | |
| 434 | * If the combine kernel does the reduction for fused_output, reduce |
| 435 | shared_output separately. O.w, reduce fused_output+shared_output later. |
| 436 | * If we have SP (TP=N, DP=M, EP), there is a separate AG step handled |
| 437 | in the model. |
| 438 | """ |
| 439 | if fused_output_is_reduced is None: |
| 440 | fused_output_is_reduced = self._fused_output_is_reduced |
| 441 | |
| 442 | if ( |
| 443 | shared_output is not None |
| 444 | and not self.moe_config.is_sequence_parallel |
| 445 | and fused_output_is_reduced |
| 446 | ): |
| 447 | shared_output = tensor_model_parallel_all_reduce(shared_output) |
| 448 | return shared_output |
| 449 | |
| 450 | def _maybe_reduce_routed_output_before_transform( |
| 451 | self, |
| 452 | fused_output: torch.Tensor, |
| 453 | fused_output_is_reduced: bool, |
| 454 | ) -> tuple[torch.Tensor, bool]: |
| 455 | """All-reduce latent routed output before its output transform. |
| 456 | |
| 457 | Latent MoE output transforms may contain non-linear ops, e.g. RMSNorm. |
| 458 | TP partial routed outputs must be summed in latent space before such |
| 459 | transforms are applied. |
| 460 | |
| 461 | A transform that commutes with the TP sum is exempt: if |
| 462 | ``sum_r T(x_r) == T(sum_r x_r)``, applying the transform to the local |
| 463 | partial output and letting the existing late all-reduce sum the |
| 464 | combined result is equivalent, and costs one collective instead of two. |
| 465 | Such a transform opts out by setting ``reduce_commutative = True``. |
| 466 | The default is False, so transforms that do not declare themselves |
| 467 | keep being reduced early. |
| 468 | """ |
| 469 | if ( |
| 470 | self.routed_output_transform is not None |
| 471 | and not getattr(self.routed_output_transform, "reduce_commutative", False) |
| 472 | and not self.moe_config.is_sequence_parallel |
| 473 | and (self.moe_config.tp_size > 1 or self.moe_config.ep_size > 1) |
| 474 | and not fused_output_is_reduced |
| 475 | ): |
| 476 | fused_output = tensor_model_parallel_all_reduce(fused_output) |
| 477 | fused_output_is_reduced = True |
| 478 | return fused_output, fused_output_is_reduced |
| 479 | |
| 480 | def _maybe_reduce_final_output( |
| 481 | self, |
| 482 | states: torch.Tensor, |
| 483 | trunc_size: int | None, |
| 484 | output_is_reduced: bool | None = None, |
| 485 | ) -> torch.Tensor: |
| 486 | """All-reduce the combined output if needed. |
| 487 | |
| 488 | This is the "late" all-reduce path. When neither fused nor shared |
| 489 | output was individually reduced, the combined sum is all-reduced |
| 490 | here. Skipped when sequence-parallel is active (SP handles its |
| 491 | own reduction) or when the early path already reduced both outputs. |
| 492 | """ |
| 493 | # skip_final_all_reduce must not coexist with a pre-reduced fused |
| 494 | # output. This should be enforced by MoE config initialization. |
| 495 | if self.moe_config.skip_final_all_reduce: |
| 496 | assert not self._fused_output_is_reduced, ( |
| 497 | "skip_final_all_reduce requires an un-reduced fused output" |
| 498 | ) |
| 499 | |
| 500 | # We don't need to reduce the final output if: |
| 501 | # - We are not running with TP or DP |
| 502 | # - The MK already reduced the fused output itself. |
| 503 | if output_is_reduced is None: |
| 504 | output_is_reduced = self._fused_output_is_reduced |
| 505 | |
| 506 | if ( |
| 507 | not self.moe_config.is_sequence_parallel |
| 508 | and not self.moe_config.skip_final_all_reduce |
| 509 | and (self.moe_config.tp_size > 1 or self.moe_config.ep_size > 1) |
| 510 | and not output_is_reduced |
| 511 | ): |
| 512 | states = tensor_model_parallel_all_reduce(states) |
| 513 | |
| 514 | return states[..., :trunc_size] if trunc_size is not None else states |
| 515 | |
| 516 | def _encode_layer_name(self) -> str | LayerName: |
| 517 | if _USE_LAYERNAME: |
| 518 | return LayerName(self.layer_name) |
| 519 | # Can be unavailable or None in unittests |
| 520 | if ( |
| 521 | is_forward_context_available() |
| 522 | and get_forward_context().all_moe_layers is not None |
| 523 | ): |
| 524 | return "from_forward_context" |
| 525 | return self.layer_name |
| 526 | |
| 527 | def _maybe_pad_hidden_states( |
| 528 | self, |
| 529 | shared_experts_input: torch.Tensor | None, |
| 530 | hidden_states: torch.Tensor, |
| 531 | ) -> tuple[torch.Tensor, int | None, int | None]: |
| 532 | """Pad hidden_states to moe_config.hidden_dim and compute the |
| 533 | original dimension for later truncation. |
| 534 | |
| 535 | For latent MoE, the routed hidden_states may be smaller than |
| 536 | hidden_dim. Padding ensures uniform tensor sizes through the |
| 537 | fused MoE kernel. The returned trunc_size is used by |
| 538 | _maybe_reduce_final_output to strip the padding from the result. |
| 539 | """ |
| 540 | shared_experts_hidden_dim = ( |
| 541 | shared_experts_input.shape[-1] if shared_experts_input is not None else 0 |
| 542 | ) |
| 543 | transformed_hidden_dim: int | None = hidden_states.shape[-1] |
| 544 | if ( |
| 545 | not self._quant_method.skip_forward_padding |
| 546 | and self.moe_config.hidden_dim != transformed_hidden_dim |
| 547 | ): |
| 548 | assert transformed_hidden_dim is not None |
| 549 | hidden_states = F.pad( |
| 550 | hidden_states, |
| 551 | (0, self.moe_config.hidden_dim - transformed_hidden_dim), |
| 552 | mode="constant", |
| 553 | value=0.0, |
| 554 | ) |
| 555 | |
| 556 | # Truncation sizes for stripping kernel padding from the output. |
| 557 | # None means no truncation needed (no padding was applied). |
| 558 | # |
| 559 | # Two truncation points exist in forward(): |
| 560 | # pre_xform: applied to fused_output BEFORE routed_output_transform |
| 561 | # post_xform: applied to the final result AFTER all-reduce |
| 562 | # |
| 563 | # MoE with routed output transform or shared experts: |
| 564 | # - pre_xform applies if the transform needs unpadded routed output |
| 565 | # or shared+routed add needs matching hidden dims. For Nemotron-3 |
| 566 | # Nano, TRTLLM NVFP4 pads routed MoE hidden dim 2688->2816, while |
| 567 | # shared output stays 2688. |
| 568 | # - post_xform uses shared_experts_hidden_dim when transform and shared |
| 569 | # experts make the final output full hidden dim. |
| 570 | # |
| 571 | # Standard MoE / MoE without transforms (GPT-OSS, Mixtral): |
| 572 | # - pre_xform is None (no early truncation) |
| 573 | # - post_xform strips padding after all-reduce (or None if unpadded) |
| 574 | if transformed_hidden_dim == hidden_states.shape[-1]: |
| 575 | transformed_hidden_dim = None |
| 576 | |
| 577 | pre_xform_trunc_size = None |
| 578 | if self.routed_output_transform is not None or shared_experts_hidden_dim > 0: |
| 579 | pre_xform_trunc_size = transformed_hidden_dim |
| 580 | post_xform_trunc_size = transformed_hidden_dim |
| 581 | if self.routed_output_transform is not None and shared_experts_hidden_dim > 0: |
| 582 | post_xform_trunc_size = shared_experts_hidden_dim |
| 583 | |
| 584 | return hidden_states, pre_xform_trunc_size, post_xform_trunc_size |
| 585 | |
| 586 | def _maybe_apply_shared_experts( |
| 587 | self, |
| 588 | shared_experts_input: torch.Tensor | None, |
| 589 | order: SharedExpertsOrder, |
| 590 | ): |
| 591 | if self._shared_experts is not None: |
| 592 | assert shared_experts_input is not None |
| 593 | self._shared_experts(shared_experts_input, order) |
| 594 | |
| 595 | def _apply_quant_method( |
| 596 | self, |
| 597 | hidden_states: torch.Tensor, |
| 598 | router_logits: torch.Tensor, |
| 599 | shared_experts_input: torch.Tensor | None, |
| 600 | input_ids: torch.Tensor | None = None, |
| 601 | shared_experts_overlapping: bool = False, |
| 602 | ) -> tuple[torch.Tensor | None, torch.Tensor | UnfinalizedMoEOutput]: |
| 603 | """Run expert routing and the fused MoE kernel via the quant method. |
| 604 | |
| 605 | Orchestrates shared expert execution (before/after), expert selection |
| 606 | via the router, and the actual fused MoE computation. Returns |
| 607 | (shared_expert_output, fused_expert_output). |
| 608 | |
| 609 | `shared_experts_overlapping` should be True only if using multi-stream |
| 610 | overlap. Then the shared expert was already launched in a separate |
| 611 | stream, so the results only have to be awaited here. |
| 612 | """ |
| 613 | self._maybe_apply_shared_experts( |
| 614 | shared_experts_input, SharedExpertsOrder.NO_OVERLAP |
| 615 | ) |
| 616 | |
| 617 | if self.routed_experts.quant_method.is_monolithic: |
| 618 | # Monolithic kernels: pass router_logits to routed_experts |
| 619 | fused_out = self.routed_experts.forward_monolithic( |
| 620 | x=hidden_states, |
| 621 | router_logits=router_logits, |
| 622 | input_ids=input_ids, |
| 623 | ) |
| 624 | else: |
| 625 | # Modular kernels: select experts first, then call routed_experts |
| 626 | topk_weights, topk_ids = self.router.select_experts( |
| 627 | hidden_states=hidden_states, |
| 628 | router_logits=router_logits, |
| 629 | topk_indices_dtype=self._quant_method.topk_indices_dtype, |
| 630 | input_ids=input_ids, |
| 631 | ) |
| 632 | |
| 633 | fused_out = self.routed_experts.forward_modular( |
| 634 | x=hidden_states, |
| 635 | topk_weights=topk_weights, |
| 636 | topk_ids=topk_ids, |
| 637 | shared_experts=self._shared_experts, |
| 638 | shared_experts_input=shared_experts_input, |
| 639 | ) |
| 640 | |
| 641 | if shared_experts_overlapping: |
| 642 | assert self._shared_experts is not None |
| 643 | self._shared_experts.wait() |
| 644 | |
| 645 | return ( |
| 646 | self._shared_experts.output if self._shared_experts is not None else None, |
| 647 | fused_out, |
| 648 | ) |
| 649 | |
| 650 | def _sequence_parallel_context(self): |
| 651 | """Return a context manager for sequence-parallel token |
| 652 | redistribution. |
| 653 | |
| 654 | When sequence parallelism is active, returns a context that handles |
| 655 | local size tracking for proper token scatter/gather. Otherwise |
| 656 | returns a no-op context. |
| 657 | """ |
| 658 | ctx = get_forward_context() |
| 659 | return ( |
| 660 | ctx.dp_metadata.sp_local_sizes( |
| 661 | self.moe_config.sp_size, |
| 662 | pcp_size=self.moe_config.pcp_size, |
| 663 | use_ep=self.moe_config.use_ep, |
| 664 | ) |
| 665 | if ctx.dp_metadata |
| 666 | else nullcontext() |
| 667 | ) |
| 668 | |
| 669 | def _maybe_add_zero_expert_output( |
| 670 | self, |
| 671 | result: torch.Tensor, |
| 672 | ) -> torch.Tensor: |
| 673 | """Add the zero expert's contribution to the final result. |
| 674 | |
| 675 | When a ZeroExpertRouter is used, it computes a bias-like output |
| 676 | from the "zero expert" that is added to the combined routed+shared |
| 677 | expert output. |
| 678 | """ |
| 679 | if isinstance(self.router, ZeroExpertRouter): |
| 680 | zero_expert_output = self.router.zero_expert_output |
| 681 | assert zero_expert_output is not None |
| 682 | result = result + zero_expert_output |
| 683 | return result |
| 684 | |
| 685 | def forward( |
| 686 | self, |
| 687 | hidden_states: torch.Tensor, |
| 688 | router_logits: torch.Tensor, |
| 689 | input_ids: torch.Tensor | None = None, |
| 690 | shared_experts_input: torch.Tensor | None = None, |
| 691 | ) -> torch.Tensor: |
| 692 | """Invoke the fused moe layer. |
| 693 | |
| 694 | Input: |
| 695 | - hidden_states |
| 696 | - router_logits |
| 697 | |
| 698 | Output: |
| 699 | - The new hidden_states. |
| 700 | |
| 701 | Calling sequence |
| 702 | - forward |
| 703 | - self._forward_entry (_moe_forward or _moe_forward_shared custom op) |
| 704 | - _forward_impl |
| 705 | |
| 706 | Note: The existence of _moe_forward and _moe_forward_shared custom ops are due |
| 707 | to the following reason: |
| 708 | 1. pytorch cannot handle union types in custom op signatures so |
| 709 | _moe_forward and _moe_forward_shared must be split. |
| 710 | """ |
| 711 | # Apply transform for routed experts (e.g., latent projection for |
| 712 | # latent MoE). When the caller pre-applies the routed input transform |
| 713 | # outside the runner (e.g. to overlap it on a separate stream), it |
| 714 | # passes the already-transformed routed input as ``hidden_states`` and |
| 715 | # the original hidden states as ``shared_experts_input``; skip the |
| 716 | # transform in that case so shared experts still see the original input. |
| 717 | if shared_experts_input is None: |
| 718 | hidden_states, shared_experts_input = self.apply_routed_input_transform( |
| 719 | hidden_states |
| 720 | ) |
| 721 | |
| 722 | # Record before `_maybe_pad_hidden_states` pads activations to match |
| 723 | # `moe_config.hidden_dim`, e.g. after `align_fp4_moe_hidden_dim_for_fi` |
| 724 | # so routed output can be trimmed before |
| 725 | # shared+routed add / latent up proj if needed. |
| 726 | |
| 727 | hidden_states, og_hidden_dim_pre_xform, og_hidden_dim_post_xform = ( |
| 728 | self._maybe_pad_hidden_states( |
| 729 | shared_experts_input, |
| 730 | hidden_states, |
| 731 | ) |
| 732 | ) |
| 733 | |
| 734 | result = self._forward_entry( |
| 735 | hidden_states, |
| 736 | router_logits, |
| 737 | shared_experts_input, |
| 738 | input_ids, |
| 739 | self._encode_layer_name(), |
| 740 | self.moe_config.hidden_dim_unpadded |
| 741 | if self._quant_method.has_unpadded_output |
| 742 | else 0, |
| 743 | ) |
| 744 | |
| 745 | # |
| 746 | # Note: there are two all-reduce points below. They are mutually |
| 747 | # exclusive, controlled by _fused_output_is_reduced |
| 748 | # - When True: the combine kernel already reduced fused_output, |
| 749 | # so we reduce shared_output here to match, then skip the |
| 750 | # all-reduce in _maybe_reduce_final_output. |
| 751 | # - When False: neither output is reduced yet, so we combine |
| 752 | # them first and all-reduce the sum in _maybe_reduce_final_output. |
| 753 | |
| 754 | # Extract outputs from result |
| 755 | shared_output, fused_output = _unpack(result) |
| 756 | fused_output = cast(torch.Tensor, fused_output) |
| 757 | |
| 758 | if og_hidden_dim_pre_xform is not None: |
| 759 | fused_output = fused_output[..., :og_hidden_dim_pre_xform] |
| 760 | |
| 761 | fused_output_is_reduced = self._fused_output_is_reduced |
| 762 | |
| 763 | # Latent routed output has to be reduced before output transform, |
| 764 | # because the transform may include non-linear normalization. |
| 765 | fused_output, fused_output_is_reduced = ( |
| 766 | self._maybe_reduce_routed_output_before_transform( |
| 767 | fused_output, |
| 768 | fused_output_is_reduced, |
| 769 | ) |
| 770 | ) |
| 771 | |
| 772 | # If routed output is already reduced, reduce shared to match. |
| 773 | # See note above re: the two all-reduce points. |
| 774 | shared_output = self._maybe_reduce_shared_expert_output( |
| 775 | shared_output, fused_output_is_reduced |
| 776 | ) |
| 777 | |
| 778 | shared_output, fused_output = self._maybe_apply_routed_scale_to_output( |
| 779 | shared_output, fused_output |
| 780 | ) |
| 781 | |
| 782 | # Apply output transform (e.g. latent -> full dim) |
| 783 | fused_output = self.apply_routed_output_transform(fused_output) |
| 784 | |
| 785 | if shared_output is not None: |
| 786 | result = shared_output + fused_output |
| 787 | else: |
| 788 | result = fused_output |
| 789 | |
| 790 | result = self._maybe_reduce_final_output( |
| 791 | result, og_hidden_dim_post_xform, fused_output_is_reduced |
| 792 | ) |
| 793 | |
| 794 | return self._maybe_add_zero_expert_output(result) |
| 795 | |
| 796 | @property |
| 797 | def do_naive_dispatch_combine(self) -> bool: |
| 798 | return ( |
| 799 | self.moe_config.dp_size > 1 or self.moe_config.is_sequence_parallel |
| 800 | ) and not self._quant_method.supports_internal_mk |
| 801 | |
| 802 | def _maybe_dispatch( |
| 803 | self, |
| 804 | hidden_states: torch.Tensor, |
| 805 | router_logits: torch.Tensor, |
| 806 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 807 | # For naive dispatch/combine Dp/Ep, dispatch the hidden states and |
| 808 | # router logits to all experts. |
| 809 | # NOTE: this will be removed once all kernels are migrated into the |
| 810 | # MoEKernel framework. |
| 811 | if self.do_naive_dispatch_combine: |
| 812 | result = get_ep_group().dispatch_router_logits( |
| 813 | hidden_states, |
| 814 | router_logits, |
| 815 | self.moe_config.is_sequence_parallel, |
| 816 | ) |
| 817 | assert len(result) == 2 |
| 818 | hidden_states, router_logits = result |
| 819 | |
| 820 | if ( |
| 821 | self.moe_config.pcp_size > 1 |
| 822 | and not self.moe_config.moe_parallel_config.use_all2all_kernels |
| 823 | ): |
| 824 | hidden_states = get_pcp_group().all_gather(hidden_states, dim=0) |
| 825 | router_logits = get_pcp_group().all_gather(router_logits, dim=0) |
| 826 | |
| 827 | return hidden_states, router_logits |
| 828 | |
| 829 | def _maybe_combine( |
| 830 | self, |
| 831 | shared_output: torch.Tensor | None, |
| 832 | hidden_states: torch.Tensor | UnfinalizedMoEOutput, |
| 833 | ) -> ( |
| 834 | torch.Tensor |
| 835 | | UnfinalizedMoEOutput |
| 836 | | tuple[torch.Tensor | None, torch.Tensor | UnfinalizedMoEOutput] |
| 837 | ): |
| 838 | if self.do_naive_dispatch_combine: |
| 839 | if isinstance(hidden_states, UnfinalizedMoEOutput): |
| 840 | raise RuntimeError( |
| 841 | "Naive expert-parallel combine cannot consume a deferred " |
| 842 | "MoE output." |
| 843 | ) |
| 844 | hidden_states = get_ep_group().combine( |
| 845 | hidden_states, |
| 846 | self.moe_config.is_sequence_parallel, |
| 847 | ) |
| 848 | |
| 849 | if ( |
| 850 | self.moe_config.pcp_size > 1 |
| 851 | and not self.moe_config.moe_parallel_config.use_all2all_kernels |
| 852 | ): |
| 853 | if isinstance(hidden_states, UnfinalizedMoEOutput): |
| 854 | raise RuntimeError( |
| 855 | "PCP reduce-scatter cannot consume a deferred MoE output." |
| 856 | ) |
| 857 | hidden_states = get_pcp_group().reduce_scatter(hidden_states, dim=0) |
| 858 | |
| 859 | if self.shared_experts is not None: |
| 860 | assert shared_output is not None |
| 861 | return shared_output, hidden_states |
| 862 | else: |
| 863 | return hidden_states |
| 864 | |
| 865 | def _forward_impl( |
| 866 | self, |
| 867 | hidden_states: torch.Tensor, |
| 868 | router_logits: torch.Tensor, |
| 869 | shared_experts_input: torch.Tensor | None, |
| 870 | input_ids: torch.Tensor | None = None, |
| 871 | ) -> ( |
| 872 | torch.Tensor |
| 873 | | UnfinalizedMoEOutput |
| 874 | | tuple[torch.Tensor, torch.Tensor | UnfinalizedMoEOutput] |
| 875 | ): |
| 876 | """Entry point called by the custom op to run the MoE computation. |
| 877 | |
| 878 | Handles pre-dispatch setup (gate application, external shared expert |
| 879 | triggering, quant config init) then performs the following steps |
| 880 | within the sequence-parallel context. |
| 881 | |
| 882 | - Performs expert routing |
| 883 | - fused MoE kernel execution |
| 884 | - shared expert computation. |
| 885 | |
| 886 | Returns routed output, optionally paired with shared-expert output. A |
| 887 | fused consumer may request the routed output in deferred-finalize form. |
| 888 | """ |
| 889 | # TODO(bnell): this can be removed after MK migration is complete. |
| 890 | self.routed_experts._ensure_moe_quant_config_init() |
| 891 | |
| 892 | # If using multi-stream overlap for shared experts, we must launch it |
| 893 | # before routed expert dispatch. |
| 894 | shared_experts_overlapping = False |
| 895 | if self._shared_experts is not None: |
| 896 | shared_experts_overlapping = self._shared_experts.maybe_forward_async( |
| 897 | shared_experts_input |
| 898 | ) |
| 899 | |
| 900 | # If the Runner holds the gate, apply it after the stream sync, |
| 901 | # so it can run overlapped with the |
| 902 | # NOTE: in future PR, MoE runner will always hold the gate. |
| 903 | if self.gate is not None: |
| 904 | if self._fse_fuse_gate: |
| 905 | self._maybe_fuse_gate_weights() |
| 906 | router_logits = dispatch_unquantized_gemm()( |
| 907 | self, hidden_states, self._combined_gate_weight, None |
| 908 | ) |
| 909 | else: |
| 910 | router_logits, _ = self.gate(hidden_states) |
| 911 | |
| 912 | with self._sequence_parallel_context(): |
| 913 | # TODO(bnell): parts of the dispatch/combine steps will go away once |
| 914 | # #32567 lands and the remaining kernels are made MKs. The PCP |
| 915 | # code will probably remain |
| 916 | hidden_states, router_logits = self._maybe_dispatch( |
| 917 | hidden_states, |
| 918 | router_logits, |
| 919 | ) |
| 920 | |
| 921 | shared_output, hidden_states = self._apply_quant_method( |
| 922 | hidden_states=hidden_states, |
| 923 | router_logits=router_logits, |
| 924 | shared_experts_input=shared_experts_input, |
| 925 | input_ids=input_ids, |
| 926 | shared_experts_overlapping=shared_experts_overlapping, |
| 927 | ) |
| 928 | |
| 929 | return self._maybe_combine( |
| 930 | shared_output, |
| 931 | hidden_states, |
| 932 | ) |
| 933 | |
| 934 | ######################################################### |
| 935 | # |
| 936 | # Old methods from FusedMoE layer. Remove when possible. |
| 937 | # |
| 938 | ######################################################### |
| 939 | |
| 940 | # |
| 941 | # Properties |
| 942 | # |
| 943 | |
| 944 | @property |
| 945 | def layer_id(self): |
| 946 | # Delayed import to avoid circular dependency |
| 947 | from vllm.model_executor.models.utils import extract_layer_index |
| 948 | |
| 949 | return extract_layer_index(self.layer_name) |
| 950 | |
| 951 | # |
| 952 | # Attributes still needed by models |
| 953 | # |
| 954 | |
| 955 | @property |
| 956 | def is_monolithic(self) -> bool: |
| 957 | return self.routed_experts.quant_method.is_monolithic |
| 958 | |
| 959 | @property |
| 960 | def activation(self) -> MoEActivation: |
| 961 | return self.routed_experts.activation |
| 962 | |
| 963 | # |
| 964 | # Expert maps |
| 965 | # |
| 966 | |
| 967 | @property |
| 968 | def expert_map_manager(self): |
| 969 | """Forward to routed_experts.expert_map_manager for backward compatibility.""" |
| 970 | return self.routed_experts.expert_map_manager |
| 971 | |
| 972 | @property |
| 973 | def expert_placement_strategy(self) -> ExpertPlacementStrategy: |
| 974 | return self.expert_map_manager.placement_strategy |
| 975 | |
| 976 | @property |
| 977 | def expert_global_to_physical(self) -> torch.Tensor | None: |
| 978 | tables = self.expert_map_manager.routing_tables |
| 979 | return tables[0] if tables else None |
| 980 | |
| 981 | @property |
| 982 | def expert_physical_to_global(self) -> torch.Tensor | None: |
| 983 | """Routing table: physical expert ID to global expert ID.""" |
| 984 | tables = self.expert_map_manager.routing_tables |
| 985 | return tables[1] if tables else None |
| 986 | |
| 987 | @property |
| 988 | def expert_local_to_global(self) -> torch.Tensor | None: |
| 989 | """Routing table: local expert ID to global expert ID.""" |
| 990 | tables = self.expert_map_manager.routing_tables |
| 991 | return tables[2] if tables else None |
| 992 | |
| 993 | @property |
| 994 | def expert_map(self) -> torch.Tensor | None: |
| 995 | return self.routed_experts.expert_map |
| 996 | |
| 997 | def _expert_routing_tables( |
| 998 | self, |
| 999 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None: |
| 1000 | return self.routed_experts._expert_routing_tables() |
| 1001 | |
| 1002 | def update_expert_map(self): |
| 1003 | self.routed_experts.update_expert_map() |
| 1004 | |
| 1005 | def _map_global_expert_id_to_local_expert_id(self, expert_id: int) -> int: |
| 1006 | """Map global expert ID to local expert ID.""" |
| 1007 | return self.routed_experts._map_global_expert_id_to_local_expert_id(expert_id) |
| 1008 | |
| 1009 | def get_expert_weights(self) -> Iterable[torch.Tensor]: |
| 1010 | return self.routed_experts.get_expert_weights() |
| 1011 | |
| 1012 | # |
| 1013 | # EPLB |
| 1014 | # |
| 1015 | |
| 1016 | @property |
| 1017 | def eplb_state(self) -> EplbLayerState | None: |
| 1018 | return self.router.eplb_state |
| 1019 | |
| 1020 | def set_eplb_state( |
| 1021 | self, |
| 1022 | moe_layer_idx: int, |
| 1023 | expert_load_view: torch.Tensor, |
| 1024 | logical_to_physical_map: torch.Tensor, |
| 1025 | logical_replica_count: torch.Tensor, |
| 1026 | ) -> None: |
| 1027 | """Register the EPLB state in this layer. |
| 1028 | |
| 1029 | This is used later in forward pass, where we get the expert mapping |
| 1030 | and record the load metrics in `expert_load_view`. |
| 1031 | """ |
| 1032 | if self.router.eplb_state is not None: |
| 1033 | self.router.eplb_state.set_layer_state( |
| 1034 | moe_layer_idx, |
| 1035 | expert_load_view, |
| 1036 | logical_to_physical_map, |
| 1037 | logical_replica_count, |
| 1038 | ) |