| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Router for CPU MoE experts. |
| 4 | |
| 5 | This is the routing logic that used to live inline in each CPU |
| 6 | FusedMoEExpertsMonolithic.apply() (see git history of |
| 7 | experts/cpu_moe.py); factoring it out into a proper FusedMoERouter is what |
| 8 | lets CPU MoE experts be FusedMoEExpertsModular like every other backend. |
| 9 | The math below is unchanged from that prior inline version -- this is a |
| 10 | relocation, not a rewrite. |
| 11 | """ |
| 12 | |
| 13 | from collections.abc import Callable |
| 14 | |
| 15 | import torch |
| 16 | |
| 17 | from vllm._custom_ops import biased_topk_cpu, hash_topk_cpu |
| 18 | from vllm.distributed.eplb.eplb_state import EplbLayerState |
| 19 | from vllm.model_executor.layers.fused_moe.config import ( |
| 20 | RoutingMethodType, |
| 21 | get_routing_method_type, |
| 22 | ) |
| 23 | from vllm.model_executor.layers.fused_moe.router.base_router import BaseRouter |
| 24 | |
| 25 | |
| 26 | def _grouped_topk( |
| 27 | gating_output: torch.Tensor, |
| 28 | topk: int, |
| 29 | renormalize: bool, |
| 30 | num_expert_group: int, |
| 31 | topk_group: int, |
| 32 | scoring_func: str, |
| 33 | routed_scaling_factor: float, |
| 34 | e_score_correction_bias: torch.Tensor | None, |
| 35 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 36 | gating_output = gating_output.float() |
| 37 | if scoring_func == "softmax": |
| 38 | scores = torch.softmax(gating_output, dim=-1) |
| 39 | elif scoring_func == "sigmoid": |
| 40 | scores = gating_output.sigmoid() |
| 41 | else: |
| 42 | raise ValueError(f"Unsupported scoring function: {scoring_func}") |
| 43 | |
| 44 | num_token = scores.shape[0] |
| 45 | if e_score_correction_bias is not None: |
| 46 | original_scores = scores |
| 47 | scores = scores + e_score_correction_bias.unsqueeze(0) |
| 48 | group_scores = ( |
| 49 | scores.view(num_token, num_expert_group, -1).topk(2, dim=-1)[0].sum(dim=-1) |
| 50 | ) |
| 51 | else: |
| 52 | group_scores = ( |
| 53 | scores.view(num_token, num_expert_group, -1).max(dim=-1).values |
| 54 | ) # [n, n_group] |
| 55 | group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[ |
| 56 | 1 |
| 57 | ] # [n, top_k_group] |
| 58 | group_mask = torch.zeros_like(group_scores) # [n, n_group] |
| 59 | group_mask.scatter_(1, group_idx, 1) # [n, n_group] |
| 60 | score_mask = ( |
| 61 | group_mask.unsqueeze(-1) |
| 62 | .expand(num_token, num_expert_group, scores.shape[-1] // num_expert_group) |
| 63 | .reshape(num_token, -1) |
| 64 | ) # [n, e] |
| 65 | tmp_scores = scores.masked_fill(~score_mask.bool(), float("-inf")) # [n, e] |
| 66 | |
| 67 | if e_score_correction_bias is not None: |
| 68 | topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1] |
| 69 | topk_weights = original_scores.gather(1, topk_ids) |
| 70 | else: |
| 71 | topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False) |
| 72 | |
| 73 | if renormalize: |
| 74 | topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) |
| 75 | |
| 76 | if routed_scaling_factor != 1.0: |
| 77 | topk_weights = topk_weights * routed_scaling_factor |
| 78 | return topk_weights, topk_ids.to(torch.int32) |
| 79 | |
| 80 | |
| 81 | def _sqrtsoftplus_bias_topk( |
| 82 | router_logits: torch.Tensor, |
| 83 | top_k: int, |
| 84 | renormalize: bool, |
| 85 | e_score_correction_bias: torch.Tensor | None, |
| 86 | routed_scaling_factor: float, |
| 87 | input_ids: torch.Tensor | None, |
| 88 | hash_indices_table: torch.Tensor | None, |
| 89 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 90 | """DeepSeek V4's routing: weight = sqrt(softplus(logit)); expert |
| 91 | selection uses a bias-corrected score (weight + correction_bias), but |
| 92 | the returned weight for each selected expert is the *unbiased* value |
| 93 | (the "noaux_tc" pattern -- bias steers selection only). Always |
| 94 | renormalizes selected weights to sum to ``routed_scaling_factor``, |
| 95 | matching the reference Triton kernel |
| 96 | (vllm/model_executor/layers/fused_moe/router/dsv4_topk.py). |
| 97 | |
| 98 | When ``hash_indices_table`` is given (hash-routed layers), expert |
| 99 | selection instead comes from ``hash_indices_table[input_ids]`` -- |
| 100 | the bias-corrected-score topk is skipped entirely, matching the CUDA |
| 101 | reference kernel (``dsv4HashTopkSoftplusSqrt``). |
| 102 | |
| 103 | Backed by the ported SGLang AVX512 kernels `biased_topk_cpu` (the |
| 104 | correction-bias path) and `hash_topk_cpu` (the hash-routed-layer path); |
| 105 | see csrc/cpu/sgl-kernels/topk.cpp. |
| 106 | """ |
| 107 | assert renormalize, "DeepseekV4 routing always renormalizes" |
| 108 | gating_output = router_logits.contiguous() |
| 109 | num_experts = gating_output.shape[1] |
| 110 | if hash_indices_table is not None: |
| 111 | assert input_ids is not None |
| 112 | # hash_indices_table is int32 by construction (see cpu/model.py) and |
| 113 | # advanced indexing always allocates a fresh contiguous tensor. |
| 114 | tid2eid = hash_indices_table[input_ids] |
| 115 | return hash_topk_cpu( |
| 116 | gating_output, |
| 117 | tid2eid, |
| 118 | top_k, |
| 119 | "sqrtsoftplus", |
| 120 | 0, # num_fused_shared_experts: never fused into routing on CPU |
| 121 | num_experts, |
| 122 | routed_scaling_factor, |
| 123 | ) |
| 124 | assert e_score_correction_bias is not None |
| 125 | return biased_topk_cpu( |
| 126 | gating_output, # only used for num_tokens/device (dtype forced fp32) |
| 127 | gating_output, |
| 128 | e_score_correction_bias.contiguous(), |
| 129 | top_k, |
| 130 | renormalize, |
| 131 | "sqrtsoftplus", |
| 132 | 0, # num_fused_shared_experts: never fused into routing on CPU |
| 133 | routed_scaling_factor, |
| 134 | True, # apply_routed_scaling_factor_on_output |
| 135 | ) |
| 136 | |
| 137 | |
| 138 | def _softmax_topk( |
| 139 | router_logits: torch.Tensor, |
| 140 | top_k: int, |
| 141 | renormalize: bool, |
| 142 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 143 | topk_logit_vals, topk_idx = torch.topk(router_logits, k=top_k, dim=-1, sorted=False) |
| 144 | if renormalize: |
| 145 | topk_vals = torch.softmax(topk_logit_vals, dim=-1) |
| 146 | else: |
| 147 | logZ = torch.logsumexp(router_logits, dim=-1, keepdim=True) |
| 148 | topk_vals = (topk_logit_vals - logZ).exp() |
| 149 | return topk_vals.to(torch.float32), topk_idx.to(torch.int32) |
| 150 | |
| 151 | |
| 152 | class CPURouter(BaseRouter): |
| 153 | """Router covering every routing scheme CPU MoE experts support: plain |
| 154 | softmax top-k, grouped top-k (with an optional correction bias), an |
| 155 | arbitrary custom_routing_function, and DeepSeek V4's sqrtsoftplus scheme |
| 156 | (bias-corrected or hash-routed via the ported biased_topk_cpu/ |
| 157 | hash_topk_cpu kernels). All of the above except sqrtsoftplus are plain |
| 158 | torch ops so Dynamo traces straight through into them like it did when |
| 159 | this logic lived inline in FusedMoEExpertsMonolithic.apply() -- no new |
| 160 | compile boundary is introduced here. |
| 161 | """ |
| 162 | |
| 163 | def __init__( |
| 164 | self, |
| 165 | top_k: int, |
| 166 | global_num_experts: int, |
| 167 | use_grouped_topk: bool = False, |
| 168 | num_expert_group: int | None = None, |
| 169 | topk_group: int | None = None, |
| 170 | renormalize: bool = True, |
| 171 | scoring_func: str = "softmax", |
| 172 | routed_scaling_factor: float = 1.0, |
| 173 | e_score_correction_bias: torch.Tensor | None = None, |
| 174 | custom_routing_function: Callable | None = None, |
| 175 | hash_indices_table: torch.Tensor | None = None, |
| 176 | eplb_state: EplbLayerState | None = None, |
| 177 | ): |
| 178 | super().__init__( |
| 179 | top_k=top_k, |
| 180 | global_num_experts=global_num_experts, |
| 181 | eplb_state=eplb_state, |
| 182 | ) |
| 183 | self.use_grouped_topk = use_grouped_topk |
| 184 | self.num_expert_group = num_expert_group |
| 185 | self.topk_group = topk_group |
| 186 | self.renormalize = renormalize |
| 187 | self.scoring_func = scoring_func |
| 188 | self.routed_scaling_factor = routed_scaling_factor |
| 189 | self.e_score_correction_bias = e_score_correction_bias |
| 190 | self.custom_routing_function = custom_routing_function |
| 191 | self._hash_indices_table = hash_indices_table |
| 192 | |
| 193 | @property |
| 194 | def routing_method_type(self) -> RoutingMethodType: |
| 195 | return get_routing_method_type( |
| 196 | scoring_func=self.scoring_func, |
| 197 | top_k=self.top_k, |
| 198 | renormalize=self.renormalize, |
| 199 | num_expert_group=self.num_expert_group, |
| 200 | has_e_score_bias=self.e_score_correction_bias is not None, |
| 201 | ) |
| 202 | |
| 203 | def _compute_routing( |
| 204 | self, |
| 205 | hidden_states: torch.Tensor, |
| 206 | router_logits: torch.Tensor, |
| 207 | indices_type: torch.dtype | None, |
| 208 | *, |
| 209 | input_ids: torch.Tensor | None = None, |
| 210 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 211 | if self.use_grouped_topk: |
| 212 | assert self.topk_group is not None |
| 213 | assert self.num_expert_group is not None |
| 214 | return _grouped_topk( |
| 215 | gating_output=router_logits, |
| 216 | topk=self.top_k, |
| 217 | renormalize=self.renormalize, |
| 218 | num_expert_group=self.num_expert_group, |
| 219 | topk_group=self.topk_group, |
| 220 | scoring_func=self.scoring_func, |
| 221 | routed_scaling_factor=self.routed_scaling_factor, |
| 222 | e_score_correction_bias=self.e_score_correction_bias, |
| 223 | ) |
| 224 | elif self.scoring_func == "sqrtsoftplus": |
| 225 | assert ( |
| 226 | self._hash_indices_table is not None |
| 227 | or self.e_score_correction_bias is not None |
| 228 | ) |
| 229 | return _sqrtsoftplus_bias_topk( |
| 230 | router_logits=router_logits, |
| 231 | top_k=self.top_k, |
| 232 | renormalize=self.renormalize, |
| 233 | e_score_correction_bias=self.e_score_correction_bias, |
| 234 | routed_scaling_factor=self.routed_scaling_factor, |
| 235 | input_ids=input_ids, |
| 236 | hash_indices_table=self._hash_indices_table, |
| 237 | ) |
| 238 | elif self.custom_routing_function is None: |
| 239 | assert self.scoring_func == "softmax" |
| 240 | return _softmax_topk(router_logits, self.top_k, self.renormalize) |
| 241 | else: |
| 242 | topk_weights, topk_ids = self.custom_routing_function( |
| 243 | hidden_states=hidden_states, |
| 244 | gating_output=router_logits, |
| 245 | topk=self.top_k, |
| 246 | renormalize=self.renormalize, |
| 247 | ) |
| 248 | # cpu_fused_moe reads routing tensors as contiguous float32/int32 |
| 249 | # buffers and does not account for tensor strides. |
| 250 | topk_weights = topk_weights.to(torch.float32).contiguous() |
| 251 | topk_ids = topk_ids.to(torch.int32).contiguous() |
| 252 | return topk_weights, topk_ids |
| 253 | |
| 254 | |
| 255 | def select_experts( |
| 256 | hidden_states: torch.Tensor, |
| 257 | router_logits: torch.Tensor, |
| 258 | top_k: int, |
| 259 | use_grouped_topk: bool, |
| 260 | renormalize: bool, |
| 261 | topk_group: int | None = None, |
| 262 | num_expert_group: int | None = None, |
| 263 | custom_routing_function: Callable | None = None, |
| 264 | scoring_func: str = "softmax", |
| 265 | routed_scaling_factor: float = 1.0, |
| 266 | e_score_correction_bias: torch.Tensor | None = None, |
| 267 | input_ids: torch.Tensor | None = None, |
| 268 | hash_indices_table: torch.Tensor | None = None, |
| 269 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 270 | """Routing helper for the CPU MoE experts that still use the monolithic |
| 271 | contract and so must compute routing themselves from ``router_logits`` |
| 272 | (the zentorch-accelerated path and the ARM W4A8 dynamic-quant kernel, |
| 273 | kept monolithic since zentorch/Arm hardware isn't available to validate |
| 274 | a modular migration here). Modular CPU experts get routing from |
| 275 | ``CPURouter`` instead; this is a thin wrapper around the same logic for |
| 276 | the two classes that can't use it. |
| 277 | """ |
| 278 | router = CPURouter( |
| 279 | top_k=top_k, |
| 280 | global_num_experts=router_logits.shape[-1], |
| 281 | use_grouped_topk=use_grouped_topk, |
| 282 | num_expert_group=num_expert_group, |
| 283 | topk_group=topk_group, |
| 284 | renormalize=renormalize, |
| 285 | scoring_func=scoring_func, |
| 286 | routed_scaling_factor=routed_scaling_factor, |
| 287 | e_score_correction_bias=e_score_correction_bias, |
| 288 | custom_routing_function=custom_routing_function, |
| 289 | hash_indices_table=hash_indices_table, |
| 290 | ) |
| 291 | return router._compute_routing( |
| 292 | hidden_states, router_logits, None, input_ids=input_ids |
| 293 | ) |