| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """MoE activation function enum and utilities.""" |
| 4 | |
| 5 | from dataclasses import dataclass |
| 6 | from enum import Enum |
| 7 | from typing import TYPE_CHECKING |
| 8 | |
| 9 | import torch |
| 10 | import torch.nn.functional as F |
| 11 | |
| 12 | from vllm.platforms import current_platform |
| 13 | |
| 14 | if TYPE_CHECKING: |
| 15 | from vllm.model_executor.layers.fused_moe.config import ( |
| 16 | FusedMoEConfig, |
| 17 | FusedMoEQuantConfig, |
| 18 | ) |
| 19 | |
| 20 | |
| 21 | class MoEActivation(Enum): |
| 22 | """Activation functions for MoE layers.""" |
| 23 | |
| 24 | # Gated activations (gate * activation(up)) expect input of shape [..., 2*d] |
| 25 | # and produce output of shape [..., d] |
| 26 | SILU = "silu" |
| 27 | GELU = "gelu" |
| 28 | GELU_TANH = "gelu_tanh" |
| 29 | RELU2 = "relu2" |
| 30 | # SWIGLUOAI expects gate/up *interleaved* in w13 ([gate0, up0, gate1, ...]), |
| 31 | # as in gpt-oss checkpoints. SWIGLUOAI_UNINTERLEAVE has identical math but |
| 32 | # expects the *packed* layout ([all gates; all ups]), as produced by a |
| 33 | # MergedColumnParallelLinear gate_up_proj (e.g. MiniMax-M3). |
| 34 | SWIGLUOAI = "swigluoai" |
| 35 | SITU = "situ" |
| 36 | SWIGLUOAI_UNINTERLEAVE = "swigluoai_uninterleave" |
| 37 | SWIGLUSTEP = "swiglustep" |
| 38 | |
| 39 | # Non-gated activations (no mul with gate) expect input of shape [..., d] |
| 40 | # and produce output of shape [..., d]. |
| 41 | # NOTE: Non-gated activations require the "_no_mul" suffix to be present. |
| 42 | SILU_NO_MUL = "silu_no_mul" |
| 43 | GELU_NO_MUL = "gelu_no_mul" |
| 44 | GELU_TANH_NO_MUL = "gelu_tanh_no_mul" |
| 45 | RELU2_NO_MUL = "relu2_no_mul" |
| 46 | |
| 47 | @property |
| 48 | def is_gated(self) -> bool: |
| 49 | """Returns True if activation expects gate*activation(up) pattern. |
| 50 | |
| 51 | Gated activations expect input tensor with 2x the output size, |
| 52 | where the first half is the gate and second half is the up projection. |
| 53 | """ |
| 54 | return not self.value.endswith("_no_mul") |
| 55 | |
| 56 | @property |
| 57 | def custom_op_name(self) -> str: |
| 58 | """Maps to the CustomOp name of activations |
| 59 | in vllm/model_executor/layers/activation.py.""" |
| 60 | return _CUSTOM_OP_NAMES[self] |
| 61 | |
| 62 | def without_mul(self) -> "MoEActivation": |
| 63 | """Get the non-gated variant of this activation. |
| 64 | |
| 65 | For activations that have a _no_mul variant, returns that variant. |
| 66 | For activations without a _no_mul variant (or already _no_mul), |
| 67 | returns self. |
| 68 | """ |
| 69 | return _WITHOUT_MUL.get(self, self) |
| 70 | |
| 71 | @classmethod |
| 72 | def from_str(cls, s: str) -> "MoEActivation": |
| 73 | """Parse from string for backward compatibility.""" |
| 74 | s = _STR_ALIASES.get(s, s) |
| 75 | for member in cls: |
| 76 | if member.value == s: |
| 77 | return member |
| 78 | valid = [m.value for m in cls] |
| 79 | raise ValueError(f"Unknown MoE activation: {s!r}. Valid activations: {valid}") |
| 80 | |
| 81 | |
| 82 | # Module-level lookup tables used by MoEActivation functions. |
| 83 | _STR_ALIASES: dict[str, str] = { |
| 84 | "gelu_pytorch_tanh": "gelu_tanh", |
| 85 | } |
| 86 | |
| 87 | _CUSTOM_OP_NAMES: dict[MoEActivation, str] = { |
| 88 | MoEActivation.SILU: "silu_and_mul", |
| 89 | MoEActivation.GELU: "gelu_and_mul", |
| 90 | MoEActivation.GELU_TANH: "gelu_tanh_and_mul", |
| 91 | MoEActivation.SITU: "situ_and_mul", |
| 92 | MoEActivation.SWIGLUOAI: "swigluoai_and_mul", |
| 93 | MoEActivation.SWIGLUOAI_UNINTERLEAVE: "silu_and_mul_with_clamp", |
| 94 | MoEActivation.SWIGLUSTEP: "swiglustep_and_mul", |
| 95 | MoEActivation.RELU2: "relu2", |
| 96 | MoEActivation.SILU_NO_MUL: "silu_and_mul", |
| 97 | MoEActivation.GELU_NO_MUL: "gelu_and_mul", |
| 98 | MoEActivation.GELU_TANH_NO_MUL: "gelu_tanh_and_mul", |
| 99 | MoEActivation.RELU2_NO_MUL: "relu2", |
| 100 | } |
| 101 | |
| 102 | _WITHOUT_MUL: dict[MoEActivation, MoEActivation] = { |
| 103 | MoEActivation.SILU: MoEActivation.SILU_NO_MUL, |
| 104 | MoEActivation.GELU: MoEActivation.GELU_NO_MUL, |
| 105 | MoEActivation.GELU_TANH: MoEActivation.GELU_TANH_NO_MUL, |
| 106 | MoEActivation.RELU2: MoEActivation.RELU2_NO_MUL, |
| 107 | } |
| 108 | |
| 109 | |
| 110 | def activation_without_mul(activation: str) -> str: |
| 111 | """Get the non-gated variant of an activation function. |
| 112 | |
| 113 | Args: |
| 114 | activation: The activation function name (e.g., "silu", "gelu") |
| 115 | |
| 116 | Returns: |
| 117 | The non-gated activation name (e.g., "silu_no_mul", "gelu_no_mul") |
| 118 | |
| 119 | """ |
| 120 | return MoEActivation.from_str(activation).without_mul().value |
| 121 | |
| 122 | |
| 123 | _APPLY_MOE_ACTIVATIONS = frozenset( |
| 124 | { |
| 125 | MoEActivation.SILU, |
| 126 | MoEActivation.GELU, |
| 127 | MoEActivation.GELU_TANH, |
| 128 | MoEActivation.SITU, |
| 129 | MoEActivation.SWIGLUOAI, |
| 130 | MoEActivation.SWIGLUOAI_UNINTERLEAVE, |
| 131 | MoEActivation.SWIGLUSTEP, |
| 132 | MoEActivation.SILU_NO_MUL, |
| 133 | MoEActivation.GELU_NO_MUL, |
| 134 | MoEActivation.GELU_TANH_NO_MUL, |
| 135 | MoEActivation.RELU2_NO_MUL, |
| 136 | } |
| 137 | ) |
| 138 | |
| 139 | |
| 140 | _MASKED_MOE_ACTIVATION_NAMES: dict[MoEActivation, str] = { |
| 141 | MoEActivation.SILU: "silu", |
| 142 | MoEActivation.GELU: "gelu", |
| 143 | MoEActivation.GELU_TANH: "gelu_tanh", |
| 144 | MoEActivation.SITU: "situ", |
| 145 | MoEActivation.SWIGLUOAI: "swigluoai", |
| 146 | MoEActivation.SWIGLUOAI_UNINTERLEAVE: "swigluoai_uninterleave", |
| 147 | MoEActivation.SWIGLUSTEP: "swiglustep", |
| 148 | MoEActivation.SILU_NO_MUL: "silu_no_mul", |
| 149 | MoEActivation.GELU_NO_MUL: "gelu_no_mul", |
| 150 | MoEActivation.GELU_TANH_NO_MUL: "gelu_tanh_no_mul", |
| 151 | MoEActivation.RELU2_NO_MUL: "relu2_no_mul", |
| 152 | } |
| 153 | |
| 154 | |
| 155 | def apply_moe_activation_supported(activation: MoEActivation) -> bool: |
| 156 | """Whether ``apply_moe_activation`` supports an activation.""" |
| 157 | return activation in _APPLY_MOE_ACTIVATIONS |
| 158 | |
| 159 | |
| 160 | def apply_moe_activation_masked_supported(activation: MoEActivation) -> bool: |
| 161 | """Whether the masked ``apply_moe_activation`` path supports an activation.""" |
| 162 | return activation in _MASKED_MOE_ACTIVATION_NAMES |
| 163 | |
| 164 | |
| 165 | @dataclass(frozen=True) |
| 166 | class ApplyMoEActivationConfig: |
| 167 | """Configuration forwarded to ``apply_moe_activation``.""" |
| 168 | |
| 169 | clamp_limit: float | None = None |
| 170 | alpha: float = 1.0 |
| 171 | beta: float = 0.0 |
| 172 | activation_situ_beta: float | None = None |
| 173 | activation_situ_linear_beta: float | None = None |
| 174 | |
| 175 | @classmethod |
| 176 | def from_configs( |
| 177 | cls, |
| 178 | moe_config: "FusedMoEConfig", |
| 179 | quant_config: "FusedMoEQuantConfig", |
| 180 | ) -> "ApplyMoEActivationConfig": |
| 181 | """Build from the model and quantization configurations.""" |
| 182 | clamp_limit = quant_config.gemm1_clamp_limit |
| 183 | if clamp_limit is None: |
| 184 | clamp_limit = moe_config.swiglu_limit |
| 185 | alpha = quant_config.gemm1_alpha |
| 186 | if alpha is None: |
| 187 | alpha = moe_config.swiglu_alpha |
| 188 | beta = quant_config.gemm1_beta |
| 189 | if beta is None: |
| 190 | beta = moe_config.swiglu_beta |
| 191 | return cls( |
| 192 | clamp_limit=clamp_limit, |
| 193 | alpha=1.0 if alpha is None else alpha, |
| 194 | beta=0.0 if beta is None else beta, |
| 195 | activation_situ_beta=moe_config.activation_situ_beta, |
| 196 | activation_situ_linear_beta=moe_config.activation_situ_linear_beta, |
| 197 | ) |
| 198 | |
| 199 | |
| 200 | _DEFAULT_APPLY_MOE_ACTIVATION_CONFIG = ApplyMoEActivationConfig() |
| 201 | |
| 202 | |
| 203 | def _validate_moe_activation_shapes( |
| 204 | activation: MoEActivation, |
| 205 | output: torch.Tensor, |
| 206 | input: torch.Tensor, |
| 207 | expected_dim: int, |
| 208 | ) -> None: |
| 209 | assert input.dim() == expected_dim, f"Input must be {expected_dim}D" |
| 210 | assert output.dim() == expected_dim, f"Output must be {expected_dim}D" |
| 211 | assert input.shape[:-1] == output.shape[:-1], ( |
| 212 | f"Input/output leading shapes must match: {input.shape} vs {output.shape}" |
| 213 | ) |
| 214 | if activation.is_gated: |
| 215 | assert output.size(-1) * 2 == input.size(-1), ( |
| 216 | f"{activation.value} expects 2x ratio: " |
| 217 | f"{output.size(-1) * 2} vs {input.size(-1)}" |
| 218 | ) |
| 219 | else: |
| 220 | assert output.size(-1) == input.size(-1), ( |
| 221 | f"{activation.value} expects equal sizes: " |
| 222 | f"{output.size(-1)} vs {input.size(-1)}" |
| 223 | ) |
| 224 | |
| 225 | |
| 226 | def _apply_moe_activation_masked( |
| 227 | activation: MoEActivation, |
| 228 | output: torch.Tensor, |
| 229 | input: torch.Tensor, |
| 230 | valid_token_counts: torch.Tensor, |
| 231 | config: ApplyMoEActivationConfig, |
| 232 | ) -> torch.Tensor: |
| 233 | masked_activation = _MASKED_MOE_ACTIVATION_NAMES.get(activation) |
| 234 | if masked_activation is None: |
| 235 | raise NotImplementedError( |
| 236 | f"Masked MoE activation is not implemented for {activation.value}" |
| 237 | ) |
| 238 | |
| 239 | assert input.dim() in (2, 3), "Masked input must be 2D or 3D" |
| 240 | _validate_moe_activation_shapes(activation, output, input, expected_dim=input.dim()) |
| 241 | assert input.dtype == output.dtype, "Input and output dtypes must match" |
| 242 | assert input.device == output.device, "Input and output devices must match" |
| 243 | assert input.is_contiguous(), "Input must be contiguous" |
| 244 | assert output.is_contiguous(), "Output must be contiguous" |
| 245 | assert valid_token_counts.dtype == torch.int32, ( |
| 246 | "valid_token_counts must use torch.int32" |
| 247 | ) |
| 248 | assert valid_token_counts.dim() == 1, "valid_token_counts must be 1D" |
| 249 | assert valid_token_counts.device == input.device, ( |
| 250 | "valid_token_counts must be on the input device" |
| 251 | ) |
| 252 | assert valid_token_counts.is_contiguous(), "valid_token_counts must be contiguous" |
| 253 | expected_counts = input.size(0) if input.dim() == 3 else 1 |
| 254 | assert valid_token_counts.size(0) == expected_counts, ( |
| 255 | f"valid_token_counts must have {expected_counts} element(s) for " |
| 256 | f"{input.dim()}D input" |
| 257 | ) |
| 258 | |
| 259 | masked_clamp_limit = 0.0 if config.clamp_limit is None else config.clamp_limit |
| 260 | masked_alpha = config.alpha |
| 261 | if activation == MoEActivation.SILU and config.clamp_limit is not None: |
| 262 | masked_activation = "silu_with_clamp" |
| 263 | elif activation == MoEActivation.SITU: |
| 264 | assert config.activation_situ_beta is not None, ( |
| 265 | "SITU requires activation_situ_beta from FusedMoEConfig" |
| 266 | ) |
| 267 | elif activation == MoEActivation.SWIGLUOAI: |
| 268 | masked_clamp_limit = 7.0 |
| 269 | masked_alpha = 1.702 |
| 270 | elif activation == MoEActivation.SWIGLUOAI_UNINTERLEAVE: |
| 271 | assert config.clamp_limit is not None, ( |
| 272 | "SWIGLUOAI_UNINTERLEAVE requires clamp_limit" |
| 273 | ) |
| 274 | elif activation == MoEActivation.SWIGLUSTEP: |
| 275 | masked_clamp_limit = 7.0 |
| 276 | |
| 277 | torch.ops._C.masked_moe_activation( |
| 278 | output, |
| 279 | input, |
| 280 | valid_token_counts, |
| 281 | masked_activation, |
| 282 | masked_clamp_limit, |
| 283 | masked_alpha, |
| 284 | config.beta, |
| 285 | 1.0 if config.activation_situ_beta is None else config.activation_situ_beta, |
| 286 | -1.0 |
| 287 | if config.activation_situ_linear_beta is None |
| 288 | else config.activation_situ_linear_beta, |
| 289 | ) |
| 290 | return output |
| 291 | |
| 292 | |
| 293 | def silu_and_mul_with_clamp( |
| 294 | output: torch.Tensor, |
| 295 | input: torch.Tensor, |
| 296 | clamp_limit: float, |
| 297 | topk_ids: torch.Tensor | None = None, |
| 298 | expert_map: torch.Tensor | None = None, |
| 299 | ) -> None: |
| 300 | if current_platform.is_xpu(): |
| 301 | d = input.shape[-1] // 2 |
| 302 | gate = torch.clamp(input[..., :d], max=clamp_limit) |
| 303 | up = torch.clamp(input[..., d:], min=-clamp_limit, max=clamp_limit) |
| 304 | output.copy_(gate * torch.sigmoid(gate) * up) |
| 305 | elif topk_ids is not None and expert_map is not None: |
| 306 | from vllm.model_executor.layers.fused_moe.utils import swiglu_limit_func |
| 307 | |
| 308 | swiglu_limit_func(output, input, clamp_limit, topk_ids, expert_map) |
| 309 | else: |
| 310 | # Fused silu(clamp(gate)) * clamp(up); equivalent to swiglu_limit_func. |
| 311 | torch.ops._C.silu_and_mul_with_clamp(output, input, clamp_limit, 1.0, 0.0) |
| 312 | |
| 313 | |
| 314 | def apply_moe_activation( |
| 315 | activation: MoEActivation, |
| 316 | output: torch.Tensor, |
| 317 | input: torch.Tensor, |
| 318 | *, |
| 319 | activation_config: ApplyMoEActivationConfig | None = None, |
| 320 | topk_ids: torch.Tensor | None = None, |
| 321 | expert_map: torch.Tensor | None = None, |
| 322 | valid_token_counts: torch.Tensor | None = None, |
| 323 | ) -> torch.Tensor: |
| 324 | """Apply MoE activation function. |
| 325 | |
| 326 | The configuration drives specialized activation behavior. Routing tensors |
| 327 | and valid token counts remain per-call inputs because they depend on the |
| 328 | current token assignment. A single token count masks a flat ``[T, D]`` |
| 329 | buffer; one count per expert masks each prefix in a padded ``[E, T, D]`` |
| 330 | buffer. |
| 331 | """ |
| 332 | config = ( |
| 333 | _DEFAULT_APPLY_MOE_ACTIVATION_CONFIG |
| 334 | if activation_config is None |
| 335 | else activation_config |
| 336 | ) |
| 337 | |
| 338 | if valid_token_counts is not None: |
| 339 | return _apply_moe_activation_masked( |
| 340 | activation, |
| 341 | output, |
| 342 | input, |
| 343 | valid_token_counts, |
| 344 | config, |
| 345 | ) |
| 346 | |
| 347 | _validate_moe_activation_shapes(activation, output, input, expected_dim=2) |
| 348 | |
| 349 | # Activations with gated multiplication (gate × activation(up)) |
| 350 | if activation == MoEActivation.SILU: |
| 351 | if config.clamp_limit is not None: |
| 352 | silu_and_mul_with_clamp( |
| 353 | output, input, config.clamp_limit, topk_ids, expert_map |
| 354 | ) |
| 355 | else: |
| 356 | torch.ops._C.silu_and_mul(output, input) |
| 357 | elif activation == MoEActivation.GELU: |
| 358 | torch.ops._C.gelu_and_mul(output, input) |
| 359 | elif activation == MoEActivation.GELU_TANH: |
| 360 | torch.ops._C.gelu_tanh_and_mul(output, input) |
| 361 | elif activation == MoEActivation.SITU: |
| 362 | # Fused CUDA kernel: writes straight to `output`, no fp32 temporaries. |
| 363 | # (The pure-torch fallback below upcast both halves to fp32 and |
| 364 | # allocated ~8 temporaries per call, blowing up MoE memory.) |
| 365 | # Both betas come from FusedMoEConfig; a missing beta means the caller |
| 366 | # bypassed the config plumbing, so fail rather than silently use 1.0. |
| 367 | # linear_beta is genuinely optional: <= 0 signals "unset" to the kernel |
| 368 | # (up passed through), matching SituAndMul(linear_beta=None). |
| 369 | assert config.activation_situ_beta is not None, ( |
| 370 | "SITU requires activation_situ_beta from FusedMoEConfig" |
| 371 | ) |
| 372 | torch.ops._C.situ_and_mul( |
| 373 | output, |
| 374 | input, |
| 375 | config.activation_situ_beta, |
| 376 | -1.0 |
| 377 | if config.activation_situ_linear_beta is None |
| 378 | else config.activation_situ_linear_beta, |
| 379 | ) |
| 380 | elif activation == MoEActivation.SWIGLUOAI: |
| 381 | torch.ops._C.swigluoai_and_mul(output, input) |
| 382 | elif activation == MoEActivation.SWIGLUOAI_UNINTERLEAVE: |
| 383 | # SwiGLU-OAI on packed w13 (gate = first half, up = second half). |
| 384 | assert config.clamp_limit is not None, ( |
| 385 | "SWIGLUOAI_UNINTERLEAVE requires clamp_limit" |
| 386 | ) |
| 387 | torch.ops._C.silu_and_mul_with_clamp( |
| 388 | output, input, config.clamp_limit, config.alpha, config.beta |
| 389 | ) |
| 390 | elif activation == MoEActivation.SWIGLUSTEP: |
| 391 | from vllm.model_executor.layers.activation import swiglustep_and_mul_triton |
| 392 | |
| 393 | swiglustep_and_mul_triton(output, input) |
| 394 | |
| 395 | # Activations without gated multiplication |
| 396 | elif activation == MoEActivation.SILU_NO_MUL: |
| 397 | output.copy_(F.silu(input)) |
| 398 | elif activation == MoEActivation.GELU_NO_MUL: |
| 399 | output.copy_(F.gelu(input)) |
| 400 | elif activation == MoEActivation.GELU_TANH_NO_MUL: |
| 401 | output.copy_(F.gelu(input, approximate="tanh")) |
| 402 | elif activation == MoEActivation.RELU2_NO_MUL: |
| 403 | F.relu(input, inplace=True) |
| 404 | torch.square(input, out=output) |
| 405 | else: |
| 406 | raise ValueError(f"Unsupported FusedMoe activation: {activation}") |
| 407 | |
| 408 | return output |
| 409 | |
| 410 | |
| 411 | def situ_and_mul_quant( |
| 412 | output: torch.Tensor, |
| 413 | scale: torch.Tensor, |
| 414 | input: torch.Tensor, |
| 415 | *, |
| 416 | beta: float, |
| 417 | linear_beta: float | None, |
| 418 | group_size: int = 0, |
| 419 | num_valid_tokens: torch.Tensor | None = None, |
| 420 | topk: int = 1, |
| 421 | ) -> None: |
| 422 | """Fused Kimi SITU activation + dynamic FP8 quantization. |
| 423 | |
| 424 | Writes the quantized fp8 down-projection input into ``output`` [M, d] and its |
| 425 | float32 scale into ``scale`` (dequant = q * scale), fusing the SITU |
| 426 | activation with the FP8 quant that Humming's w2 GEMM would otherwise perform |
| 427 | in a separate pass -- this avoids materializing (and rounding through) the |
| 428 | bf16 activation_output buffer. |
| 429 | |
| 430 | ``group_size`` selects the quant granularity: ``0`` gives per-token scales |
| 431 | ``scale`` [M, 1]; ``128`` gives k-major block-FP8 group scales ``scale`` |
| 432 | [M, d // 128], matching humming ``quant_input(group_size=128, float32)``. |
| 433 | |
| 434 | ``linear_beta`` <= 0 (or None) means "unset" (up passed through), matching |
| 435 | ``SituAndMul(linear_beta=None)``. ``num_valid_tokens`` (int32 scalar tensor) |
| 436 | is the DeepEP v2 contiguous-layout token count (e.g. ``psum[-1:]``); the |
| 437 | kernel multiplies it by ``topk`` to bound rows. Padding rows are skipped and |
| 438 | receive a benign scale of 1.0. Kept in this module so that every |
| 439 | ``torch.ops._C.situ*`` call lives in one file. |
| 440 | """ |
| 441 | torch.ops._C.situ_and_mul_quant( |
| 442 | output, |
| 443 | scale, |
| 444 | input, |
| 445 | beta, |
| 446 | -1.0 if linear_beta is None else linear_beta, |
| 447 | group_size, |
| 448 | num_valid_tokens, |
| 449 | topk, |
| 450 | ) |