| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for the Transformers modeling backend's MoE fuser.""" |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | import torch.nn as nn |
| 8 | import torch.nn.functional as F |
| 9 | |
| 10 | from vllm.model_executor.models.transformers.fusers import MoEBlockFuser |
| 11 | |
| 12 | from .test_linear import GLUMLP |
| 13 | |
| 14 | |
| 15 | class TopKRouter(nn.Module): |
| 16 | """HF v5 top-k router: `linear -> softmax -> topk (-> renorm)`.""" |
| 17 | |
| 18 | def __init__(self, num_experts=8, hidden=16, top_k=2, sigmoid=False): |
| 19 | super().__init__() |
| 20 | self.top_k = top_k |
| 21 | self.sigmoid = sigmoid |
| 22 | self.weight = nn.Parameter(torch.zeros(num_experts, hidden)) |
| 23 | |
| 24 | def forward(self, hidden_states): |
| 25 | logits = F.linear(hidden_states, self.weight) |
| 26 | scores = torch.sigmoid(logits) if self.sigmoid else F.softmax(logits, dim=-1) |
| 27 | value, index = torch.topk(scores, self.top_k, dim=-1) |
| 28 | value = value / value.sum(dim=-1, keepdim=True) |
| 29 | return logits, value, index |
| 30 | |
| 31 | |
| 32 | class ScaledRouter(TopKRouter): |
| 33 | """Greedy router scaling its top-k weights (DeepSeek `routed_scaling_factor`).""" |
| 34 | |
| 35 | def forward(self, hidden_states): |
| 36 | logits = F.linear(hidden_states, self.weight) |
| 37 | scores = F.softmax(logits, dim=-1) |
| 38 | value, index = torch.topk(scores, self.top_k, dim=-1) |
| 39 | value = value * 16.0 |
| 40 | return logits, value, index |
| 41 | |
| 42 | |
| 43 | class Fp32Router(TopKRouter): |
| 44 | """Router that computes its logits in fp32 (DeepSeek/GLM style).""" |
| 45 | |
| 46 | def forward(self, hidden_states): |
| 47 | logits = F.linear( |
| 48 | hidden_states.type(torch.float32), self.weight.type(torch.float32) |
| 49 | ) |
| 50 | scores = F.softmax(logits, dim=-1) |
| 51 | value, index = torch.topk(scores, self.top_k, dim=-1) |
| 52 | return logits, value.to(hidden_states.dtype), index |
| 53 | |
| 54 | |
| 55 | class GroupedRouter(TopKRouter): |
| 56 | """Group-limited router (DeepSeek `group_limited_greedy`), scaled weights.""" |
| 57 | |
| 58 | def forward(self, hidden_states): |
| 59 | logits = F.linear(hidden_states, self.weight) |
| 60 | scores = F.softmax(logits, dim=-1) |
| 61 | group_scores = scores.view(-1, 4, 2).max(dim=-1).values |
| 62 | group_idx = torch.topk(group_scores, k=2, dim=-1)[1] |
| 63 | group_mask = torch.zeros_like(group_scores) |
| 64 | group_mask.scatter_(1, group_idx, 1) |
| 65 | score_mask = group_mask.unsqueeze(-1).expand(-1, 4, 2).reshape(-1, 8) |
| 66 | scores = scores.masked_fill(~score_mask.bool(), 0.0) |
| 67 | value, index = torch.topk(scores, self.top_k, dim=-1) |
| 68 | value = value * 16.0 |
| 69 | return logits, value, index |
| 70 | |
| 71 | |
| 72 | class CorrectionRouter(nn.Module): |
| 73 | """Router with a score-correction bias buffer (DeepSeek-V3 noaux) -> matched.""" |
| 74 | |
| 75 | def __init__(self, num_experts=8, hidden=16): |
| 76 | super().__init__() |
| 77 | self.weight = nn.Parameter(torch.zeros(num_experts, hidden)) |
| 78 | self.register_buffer("e_score_correction_bias", torch.zeros(num_experts)) |
| 79 | |
| 80 | def forward(self, hidden_states): |
| 81 | logits = F.linear(hidden_states, self.weight) |
| 82 | scores = torch.sigmoid(logits) + self.e_score_correction_bias |
| 83 | _, index = torch.topk(scores, 2, dim=-1) |
| 84 | return logits, scores, index |
| 85 | |
| 86 | |
| 87 | class BiasedRouter(TopKRouter): |
| 88 | """A valid top-k router but not `weight`-only (extra `bias` param) -> declined.""" |
| 89 | |
| 90 | def __init__(self): |
| 91 | super().__init__() |
| 92 | self.bias = nn.Parameter(torch.zeros(8)) |
| 93 | |
| 94 | def forward(self, hidden_states): |
| 95 | logits = F.linear(hidden_states, self.weight) + self.bias |
| 96 | scores = F.softmax(logits, dim=-1) |
| 97 | value, index = torch.topk(scores, self.top_k, dim=-1) |
| 98 | return logits, value, index |
| 99 | |
| 100 | |
| 101 | class DisconnectedRouter(TopKRouter): |
| 102 | """linear+softmax+top-k present but top-k ignores the logits -> not a router.""" |
| 103 | |
| 104 | def forward(self, hidden_states): |
| 105 | logits = F.linear(hidden_states, self.weight) |
| 106 | _ = F.softmax(logits, dim=-1) # scored, but not consumed by top-k |
| 107 | value, index = torch.topk(hidden_states, self.top_k, dim=-1) |
| 108 | return logits, value, index |
| 109 | |
| 110 | |
| 111 | class MoEExperts(nn.Module): |
| 112 | """Packed experts (3D weights); only its name (`experts`) matters here.""" |
| 113 | |
| 114 | def __init__(self, num_experts=8, hidden=16, inter=32): |
| 115 | super().__init__() |
| 116 | self.gate_up_proj = nn.Parameter(torch.zeros(num_experts, 2 * inter, hidden)) |
| 117 | self.down_proj = nn.Parameter(torch.zeros(num_experts, hidden, inter)) |
| 118 | |
| 119 | def forward(self, hidden_states, index, weights): |
| 120 | return hidden_states |
| 121 | |
| 122 | |
| 123 | class MoEBlock(nn.Module): |
| 124 | """Single-tensor MoE block (Qwen3-style); subclasses override `_shared`.""" |
| 125 | |
| 126 | def __init__(self, router_cls=TopKRouter): |
| 127 | super().__init__() |
| 128 | self.experts = MoEExperts() |
| 129 | self.gate = router_cls() |
| 130 | |
| 131 | def _shared(self, x, logits): |
| 132 | """The term added to the experts' output (none for a plain block).""" |
| 133 | return 0 |
| 134 | |
| 135 | def forward(self, hidden_states): |
| 136 | x = hidden_states.reshape(-1, hidden_states.shape[-1]) |
| 137 | logits, weights, index = self.gate(x) |
| 138 | out = self.experts(x, index, weights) + self._shared(x, logits) |
| 139 | return out.reshape(hidden_states.shape) |
| 140 | |
| 141 | |
| 142 | class MoEBlockNoShared(MoEBlock): |
| 143 | """No shared-expert child but a gate-derived add -> trace skipped, still fuses.""" |
| 144 | |
| 145 | def _shared(self, x, logits): |
| 146 | return logits.sum() |
| 147 | |
| 148 | |
| 149 | class MoEBlockShared(MoEBlock): |
| 150 | """A block with a shared expert and its sigmoid gate (Qwen2-style).""" |
| 151 | |
| 152 | def __init__(self): |
| 153 | super().__init__() |
| 154 | self.shared_expert = GLUMLP() |
| 155 | self.shared_expert_gate = nn.Linear(16, 1, bias=False) |
| 156 | |
| 157 | def _shared(self, x, logits): |
| 158 | return torch.sigmoid(self.shared_expert_gate(x)) * self.shared_expert(x) |
| 159 | |
| 160 | |
| 161 | class MoEBlockSharedNoGate(MoEBlock): |
| 162 | """A block with an ungated shared expert -> native, shared passed through.""" |
| 163 | |
| 164 | def __init__(self): |
| 165 | super().__init__() |
| 166 | self.shared_expert = GLUMLP() |
| 167 | |
| 168 | def _shared(self, x, logits): |
| 169 | return self.shared_expert(x) |
| 170 | |
| 171 | |
| 172 | class MoEBlockTuple(MoEBlock): |
| 173 | """A tuple-returning block (gpt-oss-style) -> must decline.""" |
| 174 | |
| 175 | def forward(self, hidden_states): |
| 176 | x = hidden_states.reshape(-1, hidden_states.shape[-1]) |
| 177 | _, weights, index = self.gate(x) |
| 178 | return self.experts(x, index, weights), index |
| 179 | |
| 180 | |
| 181 | class MoEBlockTupleVar(MoEBlock): |
| 182 | """Returns a name bound to a tuple, not a literal tuple -> must still decline.""" |
| 183 | |
| 184 | def forward(self, hidden_states): |
| 185 | x = hidden_states.reshape(-1, hidden_states.shape[-1]) |
| 186 | _, weights, index = self.gate(x) |
| 187 | result = self.experts(x, index, weights), index |
| 188 | return result |
| 189 | |
| 190 | |
| 191 | class MoEBlockNestedTupleReturn(MoEBlock): |
| 192 | """Tuple `return` in a nested helper; block returns one tensor -> still fuses.""" |
| 193 | |
| 194 | def forward(self, hidden_states): |
| 195 | def keep(a, b): |
| 196 | return a, b |
| 197 | |
| 198 | x = hidden_states.reshape(-1, hidden_states.shape[-1]) |
| 199 | _, weights, index = self.gate(x) |
| 200 | out, _ = keep(self.experts(x, index, weights), index) |
| 201 | return out.reshape(hidden_states.shape) |
| 202 | |
| 203 | |
| 204 | class PlainMLP(nn.Module): |
| 205 | """A non-GLU FFN: `down(act(up(x)))`, no gating multiply.""" |
| 206 | |
| 207 | def __init__(self, hidden: int = 16, inter: int = 32): |
| 208 | super().__init__() |
| 209 | self.up_proj = nn.Linear(hidden, inter, bias=False) |
| 210 | self.down_proj = nn.Linear(inter, hidden, bias=False) |
| 211 | self.act_fn = nn.SiLU() |
| 212 | |
| 213 | def forward(self, x): |
| 214 | return self.down_proj(self.act_fn(self.up_proj(x))) |
| 215 | |
| 216 | |
| 217 | class MoEBlockSharedNonGLU(MoEBlock): |
| 218 | """A non-GLU shared expert -> detected by dataflow (no gate/up merge).""" |
| 219 | |
| 220 | def __init__(self): |
| 221 | super().__init__() |
| 222 | self.shared_expert = PlainMLP() |
| 223 | |
| 224 | def _shared(self, x, logits): |
| 225 | return self.shared_expert(x) |
| 226 | |
| 227 | |
| 228 | class MoEBlockUnaccounted(MoEBlock): |
| 229 | """A weight-bearing child outside the fused dataflow (pre-router) -> declined.""" |
| 230 | |
| 231 | def __init__(self): |
| 232 | super().__init__() |
| 233 | self.extra = nn.Linear(16, 16, bias=False) |
| 234 | |
| 235 | def forward(self, hidden_states): |
| 236 | x = self.extra(hidden_states.reshape(-1, hidden_states.shape[-1])) |
| 237 | _, weights, index = self.gate(x) |
| 238 | return self.experts(x, index, weights).reshape(hidden_states.shape) |
| 239 | |
| 240 | |
| 241 | class BufferScale(nn.Module): |
| 242 | """A stateful child carrying only a buffer (no parameters).""" |
| 243 | |
| 244 | def __init__(self, hidden: int = 16): |
| 245 | super().__init__() |
| 246 | self.register_buffer("scale", torch.ones(hidden)) |
| 247 | |
| 248 | def forward(self, x): |
| 249 | return x * self.scale |
| 250 | |
| 251 | |
| 252 | class MoEBlockUnaccountedBuffer(MoEBlockUnaccounted): |
| 253 | """Like `MoEBlockUnaccounted`, but the extra child holds only a buffer.""" |
| 254 | |
| 255 | def __init__(self): |
| 256 | super().__init__() |
| 257 | self.extra = BufferScale() |
| 258 | |
| 259 | |
| 260 | @pytest.mark.parametrize("sigmoid", [False, True]) |
| 261 | def test_moe_fuser_detects_router(sigmoid): |
| 262 | with torch.device("meta"): |
| 263 | block = MoEBlock(lambda: TopKRouter(sigmoid=sigmoid)) |
| 264 | fuser = MoEBlockFuser.match(block, "experts") |
| 265 | assert isinstance(fuser, MoEBlockFuser) |
| 266 | assert fuser.gate_name == "gate" |
| 267 | assert fuser.scoring_func == ("sigmoid" if sigmoid else "softmax") |
| 268 | assert fuser.shared_name is None and fuser.shared_gate_name is None |
| 269 | |
| 270 | |
| 271 | def test_moe_fuser_matches_scaled_router(): |
| 272 | """Weight scaling after the top-k (DeepSeek style) does not break matching.""" |
| 273 | with torch.device("meta"): |
| 274 | block = MoEBlock(ScaledRouter) |
| 275 | assert isinstance(MoEBlockFuser.match(block, "experts"), MoEBlockFuser) |
| 276 | |
| 277 | |
| 278 | def test_moe_fuser_matches_grouped_router(): |
| 279 | """Group masking between the score and the routing top-k still matches: the |
| 280 | scorer is anchored on the routing (last) top-k, not the group one.""" |
| 281 | with torch.device("meta"): |
| 282 | block = MoEBlock(GroupedRouter) |
| 283 | fuser = MoEBlockFuser.match(block, "experts") |
| 284 | assert isinstance(fuser, MoEBlockFuser) |
| 285 | assert fuser.scoring_func == "softmax" |
| 286 | |
| 287 | |
| 288 | def test_moe_fuser_matches_correction_router(): |
| 289 | """A score-correction bias buffer (DeepSeek-V3 noaux) is allowed; the |
| 290 | rebuilt gate carries it in fp32 for FusedMoE's biased routers.""" |
| 291 | with torch.device("meta"): |
| 292 | block = MoEBlock(CorrectionRouter) |
| 293 | fuser = MoEBlockFuser.match(block, "experts") |
| 294 | assert isinstance(fuser, MoEBlockFuser) |
| 295 | assert fuser.scoring_func == "sigmoid" |
| 296 | |
| 297 | |
| 298 | def test_moe_fuser_reads_router_dtype_from_the_gate(): |
| 299 | """A router that computes its logits in fp32 must keep routing in fp32 when |
| 300 | rebuilt, even though no config field names the dtype. The cast back to the |
| 301 | activation dtype after the top-k must not be mistaken for the routing dtype.""" |
| 302 | with torch.device("meta"): |
| 303 | fp32 = MoEBlockFuser.match(MoEBlock(Fp32Router), "experts") |
| 304 | default = MoEBlockFuser.match(MoEBlock(TopKRouter), "experts") |
| 305 | assert fp32.router_dtype == torch.float32 |
| 306 | assert default.router_dtype is None |
| 307 | |
| 308 | |
| 309 | def test_moe_fuser_detects_shared_experts(): |
| 310 | with torch.device("meta"): |
| 311 | block = MoEBlockShared() |
| 312 | fuser = MoEBlockFuser.match(block, "experts") |
| 313 | assert isinstance(fuser, MoEBlockFuser) |
| 314 | assert fuser.shared_name == "shared_expert" |
| 315 | assert fuser.shared_gate_name == "shared_expert_gate" |
| 316 | |
| 317 | |
| 318 | def test_moe_fuser_skips_shared_detection_without_extra_children(): |
| 319 | """With only experts and gate, shared-expert detection (and its block trace) |
| 320 | is skipped, so a gate-derived add is not misread as a shared expert.""" |
| 321 | with torch.device("meta"): |
| 322 | block = MoEBlockNoShared() |
| 323 | fuser = MoEBlockFuser.match(block, "experts") |
| 324 | assert isinstance(fuser, MoEBlockFuser) |
| 325 | assert fuser.shared_name is None and fuser.shared_gate_name is None |
| 326 | |
| 327 | |
| 328 | def test_moe_fuser_shared_without_gate(): |
| 329 | with torch.device("meta"): |
| 330 | block = MoEBlockSharedNoGate() |
| 331 | fuser = MoEBlockFuser.match(block, "experts") |
| 332 | assert isinstance(fuser, MoEBlockFuser) |
| 333 | assert fuser.shared_name == "shared_expert" |
| 334 | assert fuser.shared_gate_name is None |
| 335 | |
| 336 | |
| 337 | def test_moe_fuser_detects_non_glu_shared_expert(): |
| 338 | with torch.device("meta"): |
| 339 | block = MoEBlockSharedNonGLU() |
| 340 | fuser = MoEBlockFuser.match(block, "experts") |
| 341 | assert isinstance(fuser, MoEBlockFuser) |
| 342 | # Recognised by dataflow (added to the experts' output), though not a GLU. |
| 343 | assert fuser.shared_name == "shared_expert" |
| 344 | assert fuser.shared_gate_name is None |
| 345 | |
| 346 | |
| 347 | @pytest.mark.parametrize( |
| 348 | "block_cls", |
| 349 | [ |
| 350 | lambda: MoEBlock(BiasedRouter), # router with an unrecognized extra param |
| 351 | MoEBlockTuple, # tuple-returning block (e.g. gpt-oss) |
| 352 | MoEBlockTupleVar, # tuple returned via a name binding, not a literal |
| 353 | MoEBlockUnaccounted, # weight-bearing child outside the fused dataflow |
| 354 | MoEBlockUnaccountedBuffer, # buffer-only child outside the fused dataflow |
| 355 | ], |
| 356 | ) |
| 357 | def test_moe_fuser_declines_unsupported(block_cls): |
| 358 | with torch.device("meta"): |
| 359 | block = block_cls() |
| 360 | assert MoEBlockFuser.match(block, "experts") is None |
| 361 | |
| 362 | |
| 363 | def test_moe_fuser_ignores_nested_returns(): |
| 364 | """A tuple `return` inside a nested helper must not decline a block whose own |
| 365 | forward returns a single tensor.""" |
| 366 | with torch.device("meta"): |
| 367 | block = MoEBlockNestedTupleReturn() |
| 368 | assert isinstance(MoEBlockFuser.match(block, "experts"), MoEBlockFuser) |
| 369 | |
| 370 | |
| 371 | def test_moe_fuser_router_requires_connected_dataflow(): |
| 372 | """A gate with linear + softmax + top-k present but not wired as a router |
| 373 | (top-k selects over the input, not the scored logits) is not detected.""" |
| 374 | with torch.device("meta"): |
| 375 | block = MoEBlock(DisconnectedRouter) |
| 376 | assert MoEBlockFuser.match(block, "experts") is None |