| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Custom activation functions.""" |
| 4 | |
| 5 | import math |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | import torch.nn.functional as F |
| 10 | |
| 11 | from vllm import ir |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.model_executor.custom_op import CustomOp |
| 14 | from vllm.platforms import CpuArchEnum, current_platform |
| 15 | from vllm.triton_utils import tl, triton |
| 16 | from vllm.utils.collection_utils import LazyDict |
| 17 | |
| 18 | logger = init_logger(__name__) |
| 19 | |
| 20 | |
| 21 | @triton.jit |
| 22 | def _swiglustep_and_mul_kernel( |
| 23 | o_ptr, |
| 24 | o_stride, |
| 25 | x_ptr, |
| 26 | x_stride, |
| 27 | limit: tl.constexpr, |
| 28 | d: tl.constexpr, |
| 29 | BLOCK_SIZE: tl.constexpr, |
| 30 | ) -> None: |
| 31 | i = tl.program_id(axis=0).to(tl.int64) |
| 32 | j = tl.program_id(axis=1) |
| 33 | o_row_ptr = o_ptr + o_stride * i |
| 34 | x_row_ptr = x_ptr + x_stride * i |
| 35 | offsets = j * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) |
| 36 | mask = offsets < d |
| 37 | |
| 38 | gate = tl.load(x_row_ptr + offsets, mask=mask).to(tl.float32) |
| 39 | up = tl.load(x_row_ptr + offsets + d, mask=mask).to(tl.float32) |
| 40 | |
| 41 | gate_silu = tl.sigmoid(gate) * gate |
| 42 | gate_clamped = tl.minimum(gate_silu, limit) |
| 43 | up_clamped = tl.minimum(tl.maximum(up, -limit), limit) |
| 44 | |
| 45 | result = gate_clamped * up_clamped |
| 46 | result = result.to(x_ptr.dtype.element_ty) |
| 47 | tl.store(o_row_ptr + offsets, result, mask=mask) |
| 48 | |
| 49 | |
| 50 | def swiglustep_and_mul_triton( |
| 51 | output: torch.Tensor, input: torch.Tensor, limit: float = 7.0 |
| 52 | ): |
| 53 | b, n = input.shape |
| 54 | assert input.ndim == 2 |
| 55 | assert n % 2 == 0 |
| 56 | d = n // 2 |
| 57 | |
| 58 | def grid(meta): |
| 59 | return (b, triton.cdiv(d, meta["BLOCK_SIZE"])) |
| 60 | |
| 61 | _swiglustep_and_mul_kernel[grid]( |
| 62 | output, |
| 63 | output.stride(0), |
| 64 | input, |
| 65 | input.stride(0), |
| 66 | limit=limit, |
| 67 | d=d, |
| 68 | BLOCK_SIZE=1024, |
| 69 | ) |
| 70 | |
| 71 | |
| 72 | # --8<-- [start:fatrelu_and_mul] |
| 73 | @CustomOp.register("fatrelu_and_mul") |
| 74 | class FatreluAndMul(CustomOp): |
| 75 | """An activation function for FATReLU. |
| 76 | |
| 77 | The function computes x -> FATReLU(x[:d]) * x[d:] where |
| 78 | d = x.shape[-1] // 2. |
| 79 | This is used in openbmb/MiniCPM-S-1B-sft. |
| 80 | |
| 81 | Shapes: |
| 82 | x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d) |
| 83 | return: (num_tokens, d) or (batch_size, seq_len, d) |
| 84 | """ |
| 85 | |
| 86 | # --8<-- [end:fatrelu_and_mul] |
| 87 | |
| 88 | def __init__(self, threshold: float = 0.0): |
| 89 | super().__init__() |
| 90 | self.threshold = threshold |
| 91 | if current_platform.is_cuda_alike() or current_platform.is_xpu(): |
| 92 | self.op = torch.ops._C.fatrelu_and_mul |
| 93 | elif current_platform.is_cpu(): |
| 94 | self._forward_method = self.forward_native |
| 95 | |
| 96 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 97 | d = x.shape[-1] // 2 |
| 98 | x1 = x[..., :d] |
| 99 | x2 = x[..., d:] |
| 100 | x1 = F.threshold(x1, self.threshold, 0.0) |
| 101 | return x1 * x2 |
| 102 | |
| 103 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 104 | d = x.shape[-1] // 2 |
| 105 | output_shape = x.shape[:-1] + (d,) |
| 106 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 107 | self.op(out, x, self.threshold) |
| 108 | return out |
| 109 | |
| 110 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 111 | return self.forward_cuda(x) |
| 112 | |
| 113 | |
| 114 | # --8<-- [start:silu_and_mul] |
| 115 | @CustomOp.register("silu_and_mul") |
| 116 | class SiluAndMul(CustomOp): |
| 117 | """An activation function for SwiGLU. |
| 118 | |
| 119 | The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[-1] // 2. |
| 120 | |
| 121 | Shapes: |
| 122 | x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d) |
| 123 | return: (num_tokens, d) or (batch_size, seq_len, d) |
| 124 | """ |
| 125 | |
| 126 | # --8<-- [end:silu_and_mul] |
| 127 | |
| 128 | def __init__(self, *, compile_native: bool = True): |
| 129 | super().__init__(compile_native=compile_native) |
| 130 | if ( |
| 131 | current_platform.is_cuda_alike() |
| 132 | or current_platform.is_xpu() |
| 133 | or ( |
| 134 | current_platform.is_cpu() |
| 135 | and current_platform.get_cpu_architecture() == CpuArchEnum.POWERPC |
| 136 | ) |
| 137 | ): |
| 138 | self.op = torch.ops._C.silu_and_mul |
| 139 | |
| 140 | @staticmethod |
| 141 | def forward_native(x: torch.Tensor) -> torch.Tensor: |
| 142 | """PyTorch-native implementation equivalent to forward().""" |
| 143 | d = x.shape[-1] // 2 |
| 144 | return F.silu(x[..., :d]) * x[..., d:] |
| 145 | |
| 146 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 147 | d = x.shape[-1] // 2 |
| 148 | output_shape = x.shape[:-1] + (d,) |
| 149 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 150 | self.op(out, x) |
| 151 | return out |
| 152 | |
| 153 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 154 | return self.forward_cuda(x) |
| 155 | |
| 156 | def forward_cpu(self, x: torch.Tensor) -> torch.Tensor: |
| 157 | if current_platform.get_cpu_architecture() == CpuArchEnum.POWERPC: |
| 158 | return self.forward_cuda(x) |
| 159 | return self.forward_native(x) |
| 160 | |
| 161 | |
| 162 | @CustomOp.register("situ_and_mul") |
| 163 | class SituAndMul(CustomOp): |
| 164 | """SituGLU activation used by Kimi models. |
| 165 | |
| 166 | Computes beta * tanh(gate / beta) * sigmoid(gate) * up. When |
| 167 | ``linear_beta`` is set, the up projection is also softly clipped with |
| 168 | linear_beta * tanh(up / linear_beta). |
| 169 | """ |
| 170 | |
| 171 | def __init__( |
| 172 | self, |
| 173 | beta: float = 1.0, |
| 174 | linear_beta: float | None = None, |
| 175 | *, |
| 176 | compile_native: bool = True, |
| 177 | ): |
| 178 | super().__init__(compile_native=compile_native) |
| 179 | self.beta = float(beta) |
| 180 | self.linear_beta = None if linear_beta is None else float(linear_beta) |
| 181 | if current_platform.is_cuda_alike() or current_platform.is_xpu(): |
| 182 | self.op = torch.ops._C.situ_and_mul |
| 183 | |
| 184 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 185 | d = x.shape[-1] // 2 |
| 186 | gate = x[..., :d].float() |
| 187 | up = x[..., d:].float() |
| 188 | gate = self.beta * torch.tanh(gate / self.beta) * torch.sigmoid(gate) |
| 189 | if self.linear_beta is not None: |
| 190 | up = self.linear_beta * torch.tanh(up / self.linear_beta) |
| 191 | return (gate * up).to(x.dtype) |
| 192 | |
| 193 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 194 | # Fused CUDA/XPU kernel: writes straight to `out`, no fp32 temporaries. |
| 195 | # linear_beta<=0 signals "unset" to the kernel (up passed through). |
| 196 | d = x.shape[-1] // 2 |
| 197 | out = torch.empty(x.shape[:-1] + (d,), dtype=x.dtype, device=x.device) |
| 198 | self.op( |
| 199 | out, x, self.beta, -1.0 if self.linear_beta is None else self.linear_beta |
| 200 | ) |
| 201 | return out |
| 202 | |
| 203 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 204 | return self.forward_cuda(x) |
| 205 | |
| 206 | |
| 207 | @CustomOp.register("silu_and_mul_with_clamp") |
| 208 | class SiluAndMulWithClamp(CustomOp): |
| 209 | """SwiGLU activation with input clamping (used by some MoE shared experts). |
| 210 | |
| 211 | Computes: |
| 212 | gate = clamp(x[..., :d], max=swiglu_limit) |
| 213 | up = clamp(x[..., d:], min=-swiglu_limit, max=swiglu_limit) |
| 214 | out = gate * sigmoid(alpha * gate) * (up + beta) |
| 215 | where d = x.shape[-1] // 2. The defaults alpha=1.0, beta=0.0 reduce this to |
| 216 | ``silu(gate) * up``; SwiGLU-OAI style models pass alpha (sigmoid scale) and |
| 217 | beta=1.0 (up bias). |
| 218 | |
| 219 | Shapes: |
| 220 | x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d) |
| 221 | return: (num_tokens, d) or (batch_size, seq_len, d) |
| 222 | """ |
| 223 | |
| 224 | def __init__( |
| 225 | self, |
| 226 | swiglu_limit: float, |
| 227 | alpha: float = 1.0, |
| 228 | beta: float = 0.0, |
| 229 | *, |
| 230 | compile_native: bool = True, |
| 231 | ): |
| 232 | super().__init__(compile_native=compile_native) |
| 233 | self.swiglu_limit = float(swiglu_limit) |
| 234 | self.alpha = float(alpha) |
| 235 | self.beta = float(beta) |
| 236 | if current_platform.is_cuda_alike(): |
| 237 | self.op = torch.ops._C.silu_and_mul_with_clamp |
| 238 | |
| 239 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 240 | d = x.shape[-1] // 2 |
| 241 | gate = torch.clamp(x[..., :d], max=self.swiglu_limit) |
| 242 | up = torch.clamp(x[..., d:], min=-self.swiglu_limit, max=self.swiglu_limit) |
| 243 | return gate * torch.sigmoid(self.alpha * gate) * (up + self.beta) |
| 244 | |
| 245 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 246 | d = x.shape[-1] // 2 |
| 247 | output_shape = x.shape[:-1] + (d,) |
| 248 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 249 | self.op(out, x, self.swiglu_limit, self.alpha, self.beta) |
| 250 | return out |
| 251 | |
| 252 | def forward_hip(self, x: torch.Tensor) -> torch.Tensor: |
| 253 | # Limit the ROCm _C path to the default case to avoid the |
| 254 | # precision loss MiniMax saw without fp32 intermediates. |
| 255 | if self.alpha == 1.0 and self.beta == 0.0: |
| 256 | return self.forward_cuda(x) |
| 257 | return self.forward_native(x) |
| 258 | |
| 259 | def extra_repr(self) -> str: |
| 260 | return ( |
| 261 | f"swiglu_limit={self.swiglu_limit!r}, " |
| 262 | f"alpha={self.alpha!r}, beta={self.beta!r}" |
| 263 | ) |
| 264 | |
| 265 | |
| 266 | # --8<-- [start:mul_and_silu] |
| 267 | @CustomOp.register("mul_and_silu") |
| 268 | class MulAndSilu(CustomOp): |
| 269 | """An activation function for SwiGLU. |
| 270 | |
| 271 | The function computes x -> x[:d] * silu(x[d:]) where d = x.shape[-1] // 2. |
| 272 | |
| 273 | Shapes: |
| 274 | x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d) |
| 275 | return: (num_tokens, d) or (batch_size, seq_len, d) |
| 276 | """ |
| 277 | |
| 278 | # --8<-- [end:mul_and_silu] |
| 279 | |
| 280 | def __init__(self): |
| 281 | super().__init__() |
| 282 | if current_platform.is_cuda_alike() or current_platform.is_xpu(): |
| 283 | self.op = torch.ops._C.mul_and_silu |
| 284 | elif current_platform.is_cpu(): |
| 285 | self._forward_method = self.forward_native |
| 286 | |
| 287 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 288 | """PyTorch-native implementation equivalent to forward().""" |
| 289 | d = x.shape[-1] // 2 |
| 290 | return x[..., :d] * F.silu(x[..., d:]) |
| 291 | |
| 292 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 293 | d = x.shape[-1] // 2 |
| 294 | output_shape = x.shape[:-1] + (d,) |
| 295 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 296 | self.op(out, x) |
| 297 | return out |
| 298 | |
| 299 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 300 | return self.forward_cuda(x) |
| 301 | |
| 302 | |
| 303 | # --8<-- [start:gelu_and_mul_sparse] |
| 304 | @CustomOp.register("gelu_and_mul_sparse") |
| 305 | class GeluAndMulSparse(CustomOp): |
| 306 | """An activation function for GeluAndMulSparse. |
| 307 | This activation function is used in Gemma3n. It computes: |
| 308 | up_proj = self.up_proj(x) |
| 309 | gate_proj = self.gate_proj(x) |
| 310 | gate_proj = self._gaussian_topk(gate_proj) # sparsity |
| 311 | activations = self.act_fn(gate_proj) # gelu |
| 312 | down_proj = self.down_proj(activations * up_proj) |
| 313 | Shapes: |
| 314 | x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d) |
| 315 | return: (num_tokens, d) or (batch_size, seq_len, d) |
| 316 | """ |
| 317 | |
| 318 | # --8<-- [end:gelu_and_mul_sparse] |
| 319 | |
| 320 | def __init__(self, activation_sparsity: float, approximate: str = "none"): |
| 321 | super().__init__() |
| 322 | # Gelu. |
| 323 | self.approximate = approximate |
| 324 | if approximate not in ("none", "tanh"): |
| 325 | raise ValueError(f"Unknown approximate mode: {approximate}") |
| 326 | if current_platform.is_rocm() and approximate == "tanh": |
| 327 | # TODO:[ROCm] PyTorch native GELU with tanh is unstable with torch.compile |
| 328 | logger.warning_once( |
| 329 | "[ROCm] Pytorch's native GELU with tanh approximation is currently " |
| 330 | "unstable and produces garbage. Fallback to 'none' approximation." |
| 331 | ) |
| 332 | self.approximate = "none" |
| 333 | |
| 334 | # Sparsity. |
| 335 | if activation_sparsity == 0.0: |
| 336 | raise ValueError("activation_sparsity is 0.0. Please use GeluAndMul.") |
| 337 | target_sparsity_tensor = torch.tensor(activation_sparsity, dtype=torch.float32) |
| 338 | normal_dist = torch.distributions.normal.Normal(0, 1) |
| 339 | self.std_multiplier = normal_dist.icdf(target_sparsity_tensor).item() |
| 340 | |
| 341 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 342 | """PyTorch-native implementation equivalent to forward().""" |
| 343 | return ir.ops.gelu_and_mul_sparse(x, self.std_multiplier, self.approximate) |
| 344 | |
| 345 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 346 | return self.forward_native(x) |
| 347 | |
| 348 | |
| 349 | # --8<-- [start:gelu] |
| 350 | @CustomOp.register("gelu") |
| 351 | class GELU(CustomOp): |
| 352 | # --8<-- [end:gelu] |
| 353 | |
| 354 | def __init__(self): |
| 355 | super().__init__() |
| 356 | if ( |
| 357 | current_platform.is_cpu() |
| 358 | and current_platform.get_cpu_architecture() == CpuArchEnum.ARM |
| 359 | and hasattr(torch.ops._C, "activation_lut_bf16") |
| 360 | ): |
| 361 | self.op = torch.ops._C.activation_lut_bf16 |
| 362 | else: |
| 363 | self.op = None |
| 364 | |
| 365 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 366 | return F.gelu(x, approximate="none") |
| 367 | |
| 368 | def forward_cpu(self, x: torch.Tensor) -> torch.Tensor: |
| 369 | if self.op and x.dtype == torch.bfloat16 and x.is_contiguous(): |
| 370 | out = torch.empty_like(x) |
| 371 | self.op(out, x, "gelu") |
| 372 | return out |
| 373 | return self.forward_native(x) |
| 374 | |
| 375 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 376 | return self.forward_native(x) |
| 377 | |
| 378 | |
| 379 | # --8<-- [start:gelu_tanh] |
| 380 | @CustomOp.register("gelu_tanh") |
| 381 | class GELUTanh(CustomOp): |
| 382 | # --8<-- [end:gelu_tanh] |
| 383 | |
| 384 | def __init__(self): |
| 385 | super().__init__() |
| 386 | if ( |
| 387 | current_platform.is_cpu() |
| 388 | and current_platform.get_cpu_architecture() == CpuArchEnum.ARM |
| 389 | and hasattr(torch.ops._C, "gelu_tanh") |
| 390 | ): |
| 391 | self.op = torch.ops._C.gelu_tanh |
| 392 | else: |
| 393 | self.op = None |
| 394 | |
| 395 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 396 | return F.gelu(x, approximate="tanh") |
| 397 | |
| 398 | def forward_cpu(self, x: torch.Tensor) -> torch.Tensor: |
| 399 | if self.op: |
| 400 | out = torch.empty_like(x) |
| 401 | self.op(out, x) |
| 402 | return out |
| 403 | return self.forward_native(x) |
| 404 | |
| 405 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 406 | return self.forward_native(x) |
| 407 | |
| 408 | |
| 409 | # --8<-- [start:gelu_and_mul] |
| 410 | @CustomOp.register("gelu_and_mul") |
| 411 | class GeluAndMul(CustomOp): |
| 412 | """An activation function for GeGLU. |
| 413 | |
| 414 | The function computes x -> GELU(x[:d]) * x[d:] where d = x.shape[-1] // 2. |
| 415 | |
| 416 | Shapes: |
| 417 | x: (batch_size, seq_len, 2 * d) or (num_tokens, 2 * d) |
| 418 | return: (batch_size, seq_len, d) or (num_tokens, d) |
| 419 | """ |
| 420 | |
| 421 | # --8<-- [end:gelu_and_mul] |
| 422 | |
| 423 | def __init__(self, approximate: str = "none"): |
| 424 | super().__init__() |
| 425 | self.approximate = approximate |
| 426 | if approximate not in ("none", "tanh"): |
| 427 | raise ValueError(f"Unknown approximate mode: {approximate}") |
| 428 | if ( |
| 429 | current_platform.is_cuda_alike() |
| 430 | or current_platform.is_xpu() |
| 431 | or ( |
| 432 | current_platform.is_cpu() |
| 433 | and current_platform.get_cpu_architecture() == CpuArchEnum.POWERPC |
| 434 | ) |
| 435 | ): |
| 436 | if approximate == "none": |
| 437 | self.op = torch.ops._C.gelu_and_mul |
| 438 | elif approximate == "tanh": |
| 439 | self.op = torch.ops._C.gelu_tanh_and_mul |
| 440 | if current_platform.is_rocm() and approximate == "tanh": |
| 441 | logger.warning_once( |
| 442 | "[ROCm] PyTorch's native GELU with tanh approximation is unstable " |
| 443 | "with torch.compile. For native implementation, fallback to 'none' " |
| 444 | "approximation. The custom kernel implementation is unaffected." |
| 445 | ) |
| 446 | |
| 447 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 448 | """PyTorch-native implementation equivalent to forward().""" |
| 449 | # TODO: [ROCm] PyTorch's native GELU with tanh is unstable with torch.compile |
| 450 | approximate = self.approximate |
| 451 | if current_platform.is_rocm() and approximate == "tanh": |
| 452 | approximate = "none" |
| 453 | d = x.shape[-1] // 2 |
| 454 | return F.gelu(x[..., :d], approximate=approximate) * x[..., d:] |
| 455 | |
| 456 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 457 | d = x.shape[-1] // 2 |
| 458 | output_shape = x.shape[:-1] + (d,) |
| 459 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 460 | self.op(out, x) |
| 461 | return out |
| 462 | |
| 463 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 464 | return self.forward_cuda(x) |
| 465 | |
| 466 | def forward_cpu(self, x: torch.Tensor) -> torch.Tensor: |
| 467 | if current_platform.get_cpu_architecture() == CpuArchEnum.POWERPC: |
| 468 | return self.forward_cuda(x) |
| 469 | return self.forward_native(x) |
| 470 | |
| 471 | def extra_repr(self) -> str: |
| 472 | return f"approximate={repr(self.approximate)}" |
| 473 | |
| 474 | |
| 475 | # --8<-- [start:swigluoai_and_mul] |
| 476 | @CustomOp.register("swigluoai_and_mul") |
| 477 | class SwigluOAIAndMul(CustomOp): |
| 478 | # https://github.com/huggingface/transformers/blob/v4.55.0/src/transformers/models/gpt_oss/modeling_gpt_oss.py#L106-L110 |
| 479 | # --8<-- [end:swigluoai_and_mul] |
| 480 | |
| 481 | def __init__(self, alpha: float = 1.702, limit: float = 7.0): |
| 482 | super().__init__() |
| 483 | self.alpha = alpha |
| 484 | self.limit = limit |
| 485 | |
| 486 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 487 | """PyTorch-native implementation equivalent to forward().""" |
| 488 | gate, up = x[..., ::2], x[..., 1::2] |
| 489 | gate = gate.clamp(min=None, max=self.limit) |
| 490 | up = up.clamp(min=-self.limit, max=self.limit) |
| 491 | glu = gate * torch.sigmoid(gate * self.alpha) |
| 492 | gated_output = (up + 1) * glu |
| 493 | return gated_output |
| 494 | |
| 495 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 496 | d = x.shape[-1] // 2 |
| 497 | output_shape = x.shape[:-1] + (d,) |
| 498 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 499 | torch.ops._C.swigluoai_and_mul(out, x, self.alpha, self.limit) |
| 500 | return out |
| 501 | |
| 502 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 503 | return self.forward_cuda(x) |
| 504 | |
| 505 | def extra_repr(self) -> str: |
| 506 | return f"alpha={repr(self.alpha)}, limit={repr(self.limit)}" |
| 507 | |
| 508 | |
| 509 | # --8<-- [start:swiglustep_and_mul] |
| 510 | @CustomOp.register("swiglustep_and_mul") |
| 511 | class SwigluStepAndMul(CustomOp): |
| 512 | """An activation function for SwiGLU with clamping. |
| 513 | |
| 514 | Computes x -> silu(x[:d]).clamp(max=limit) * x[d:].clamp(-limit, limit) |
| 515 | where d = x.shape[-1] // 2. |
| 516 | |
| 517 | Shapes: |
| 518 | x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d) |
| 519 | return: (num_tokens, d) or (batch_size, seq_len, d) |
| 520 | """ |
| 521 | |
| 522 | def __init__(self, limit: float = 7.0): |
| 523 | super().__init__() |
| 524 | if limit is None: |
| 525 | raise ValueError("SwigluStepAndMul requires limit to be set.") |
| 526 | self.limit = limit |
| 527 | |
| 528 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 529 | """PyTorch-native implementation equivalent to forward().""" |
| 530 | gate, up = x.chunk(2, dim=-1) |
| 531 | gate = F.silu(gate) |
| 532 | gate = gate.clamp(max=self.limit) |
| 533 | up = up.clamp(min=-self.limit, max=self.limit) |
| 534 | return gate * up |
| 535 | |
| 536 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 537 | d = x.shape[-1] // 2 |
| 538 | output_shape = x.shape[:-1] + (d,) |
| 539 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 540 | swiglustep_and_mul_triton(out, x, self.limit) |
| 541 | return out |
| 542 | |
| 543 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 544 | d = x.shape[-1] // 2 |
| 545 | output_shape = x.shape[:-1] + (d,) |
| 546 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 547 | torch.ops._C.swiglustep_and_mul(out, x, self.limit) |
| 548 | return out |
| 549 | |
| 550 | def extra_repr(self) -> str: |
| 551 | return f"limit={repr(self.limit)}" |
| 552 | |
| 553 | |
| 554 | # --8<-- [start:gelu_new] |
| 555 | @CustomOp.register("gelu_new") |
| 556 | class NewGELU(CustomOp): |
| 557 | # --8<-- [end:gelu_new] |
| 558 | |
| 559 | def __init__(self): |
| 560 | super().__init__() |
| 561 | if ( |
| 562 | current_platform.is_cuda_alike() |
| 563 | or current_platform.is_cpu() |
| 564 | or current_platform.is_xpu() |
| 565 | ): |
| 566 | self.op = torch.ops._C.gelu_new |
| 567 | |
| 568 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 569 | """PyTorch-native implementation equivalent to forward().""" |
| 570 | c = math.sqrt(2.0 / math.pi) |
| 571 | return 0.5 * x * (1.0 + torch.tanh(c * (x + 0.044715 * torch.pow(x, 3.0)))) |
| 572 | |
| 573 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 574 | out = torch.empty_like(x) |
| 575 | self.op(out, x) |
| 576 | return out |
| 577 | |
| 578 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 579 | return self.forward_cuda(x) |
| 580 | |
| 581 | def forward_cpu(self, x: torch.Tensor) -> torch.Tensor: |
| 582 | if current_platform.get_cpu_architecture() == CpuArchEnum.POWERPC: |
| 583 | return self.forward_cuda(x) |
| 584 | return self.forward_native(x) |
| 585 | |
| 586 | |
| 587 | # --8<-- [start:gelu_fast] |
| 588 | @CustomOp.register("gelu_fast") |
| 589 | class FastGELU(CustomOp): |
| 590 | # --8<-- [end:gelu_fast] |
| 591 | |
| 592 | def __init__(self): |
| 593 | super().__init__() |
| 594 | if ( |
| 595 | current_platform.is_cuda_alike() |
| 596 | or current_platform.is_cpu() |
| 597 | or current_platform.is_xpu() |
| 598 | ): |
| 599 | self.op = torch.ops._C.gelu_fast |
| 600 | |
| 601 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 602 | """PyTorch-native implementation equivalent to forward().""" |
| 603 | return 0.5 * x * (1.0 + torch.tanh(x * 0.7978845608 * (1.0 + 0.044715 * x * x))) |
| 604 | |
| 605 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 606 | out = torch.empty_like(x) |
| 607 | self.op(out, x) |
| 608 | return out |
| 609 | |
| 610 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 611 | return self.forward_cuda(x) |
| 612 | |
| 613 | def forward_cpu(self, x: torch.Tensor) -> torch.Tensor: |
| 614 | if current_platform.get_cpu_architecture() == CpuArchEnum.POWERPC: |
| 615 | return self.forward_cuda(x) |
| 616 | return self.forward_native(x) |
| 617 | |
| 618 | |
| 619 | # --8<-- [start:quick_gelu] |
| 620 | @CustomOp.register("quick_gelu") |
| 621 | class QuickGELU(CustomOp): |
| 622 | # https://github.com/huggingface/transformers/blob/main/src/transformers/activations.py#L90 |
| 623 | # --8<-- [end:quick_gelu] |
| 624 | |
| 625 | def __init__(self): |
| 626 | super().__init__() |
| 627 | if ( |
| 628 | current_platform.is_cuda_alike() |
| 629 | or current_platform.is_cpu() |
| 630 | or current_platform.is_xpu() |
| 631 | ): |
| 632 | self.op = torch.ops._C.gelu_quick |
| 633 | |
| 634 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 635 | """PyTorch-native implementation equivalent to forward().""" |
| 636 | return x * torch.sigmoid(1.702 * x) |
| 637 | |
| 638 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 639 | out = torch.empty_like(x) |
| 640 | self.op(out, x) |
| 641 | return out |
| 642 | |
| 643 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 644 | return self.forward_cuda(x) |
| 645 | |
| 646 | def forward_cpu(self, x: torch.Tensor) -> torch.Tensor: |
| 647 | if current_platform.get_cpu_architecture() == CpuArchEnum.POWERPC: |
| 648 | return self.forward_cuda(x) |
| 649 | return self.forward_native(x) |
| 650 | |
| 651 | |
| 652 | # --8<-- [start:relu2] |
| 653 | @CustomOp.register("relu2") |
| 654 | class ReLUSquaredActivation(CustomOp): |
| 655 | """Applies the relu^2 activation introduced in https://arxiv.org/abs/2109.08668v2.""" |
| 656 | |
| 657 | # --8<-- [end:relu2] |
| 658 | |
| 659 | def __init__(self, *, compile_native: bool = True): |
| 660 | super().__init__(compile_native=compile_native) |
| 661 | if current_platform.is_cuda_alike(): |
| 662 | self.op = torch.ops._C.relu_squared |
| 663 | elif current_platform.is_xpu(): |
| 664 | self.op = torch.ops._C.relu2_no_mul |
| 665 | |
| 666 | def forward_native(self, x: torch.Tensor) -> torch.Tensor: |
| 667 | """PyTorch-native implementation equivalent to forward().""" |
| 668 | return torch.square(F.relu(x)) |
| 669 | |
| 670 | def forward_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 671 | out = torch.empty_like(x) |
| 672 | self.op(out, x) |
| 673 | return out |
| 674 | |
| 675 | def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: |
| 676 | return self.forward_cuda(x) |
| 677 | |
| 678 | |
| 679 | # --8<-- [start:xielu] |
| 680 | @CustomOp.register("xielu") |
| 681 | class XIELU(CustomOp): |
| 682 | """Applies the xIELU activation function introduced in https://arxiv.org/abs/2411.13010 |
| 683 | If the user has installed the nickjbrowning/XIELU, we import xIELU CUDA |
| 684 | Otherwise, we emit a single warning and use xIELU Python |
| 685 | """ |
| 686 | |
| 687 | # --8<-- [end:xielu] |
| 688 | |
| 689 | def __init__( |
| 690 | self, |
| 691 | alpha_p_init: float = 0.8, |
| 692 | alpha_n_init: float = 0.8, |
| 693 | beta: float = 0.5, |
| 694 | eps: float = -1e-6, |
| 695 | dtype: torch.dtype = torch.bfloat16, |
| 696 | with_vector_loads: bool = False, |
| 697 | ): |
| 698 | super().__init__() |
| 699 | self.alpha_p = nn.Parameter( |
| 700 | torch.log(torch.exp(torch.tensor(alpha_p_init, dtype=dtype)) - 1).unsqueeze( |
| 701 | 0 |
| 702 | ) |
| 703 | ) |
| 704 | self.alpha_n = nn.Parameter( |
| 705 | torch.log( |
| 706 | torch.exp(torch.tensor(alpha_n_init - beta, dtype=dtype)) - 1 |
| 707 | ).unsqueeze(0) |
| 708 | ) |
| 709 | self.register_buffer("beta", torch.tensor(beta, dtype=dtype)) |
| 710 | self.register_buffer("eps", torch.tensor(eps, dtype=dtype)) |
| 711 | self.with_vector_loads = with_vector_loads |
| 712 | # Temporary until xIELU CUDA fully implemented |
| 713 | self._beta_scalar = float(self.beta.detach().cpu().float().item()) |
| 714 | self._eps_scalar = float(self.eps.detach().cpu().float().item()) |
| 715 | |
| 716 | self._xielu_cuda_obj = None |
| 717 | try: |
| 718 | import xielu.ops # noqa: F401 |
| 719 | |
| 720 | self._xielu_cuda_obj = torch.classes.xielu.XIELU() |
| 721 | msg = "Using experimental xIELU CUDA." |
| 722 | try: |
| 723 | from torch._dynamo import allow_in_graph |
| 724 | |
| 725 | self._xielu_cuda_fn = allow_in_graph(self._xielu_cuda) |
| 726 | msg += " Enabled torch._dynamo for xIELU CUDA." |
| 727 | except Exception as err: |
| 728 | msg += ( |
| 729 | f" Could not enable torch._dynamo for xIELU ({err}) - " |
| 730 | "this may result in slower performance." |
| 731 | ) |
| 732 | self._xielu_cuda_fn = self._xielu_cuda |
| 733 | logger.warning_once(msg) |
| 734 | except Exception as err: |
| 735 | logger.warning_once( |
| 736 | "CUDA-fused xIELU not available (%s) –" |
| 737 | " falling back to a Python version.\n" |
| 738 | "For CUDA xIELU (experimental), `pip install git+https://github.com/nickjbrowning/XIELU`", |
| 739 | str(err), |
| 740 | ) |
| 741 | |
| 742 | def _xielu_python(self, x: torch.Tensor) -> torch.Tensor: |
| 743 | alpha_p = nn.functional.softplus(self.alpha_p) |
| 744 | alpha_n = self.beta + nn.functional.softplus(self.alpha_n) |
| 745 | return torch.where( |
| 746 | x > 0, |
| 747 | alpha_p * x * x + self.beta * x, |
| 748 | (torch.expm1(torch.min(x, self.eps)) - x) * alpha_n + self.beta * x, |
| 749 | ) |
| 750 | |
| 751 | def _xielu_cuda(self, x: torch.Tensor) -> torch.Tensor: |
| 752 | """Firewall function to prevent torch.compile from seeing .item().""" |
| 753 | assert self._xielu_cuda_obj is not None, "XIELU CUDA object must not be None" |
| 754 | original_shape = x.shape |
| 755 | # CUDA kernel expects 3D tensors, reshape if needed |
| 756 | while x.dim() < 3: |
| 757 | x = x.unsqueeze(0) |
| 758 | if x.dim() > 3: |
| 759 | x = x.view(-1, 1, x.size(-1)) |
| 760 | if original_shape != x.shape: |
| 761 | logger.warning_once( |
| 762 | "Warning: xIELU input tensor expects 3 dimensions" |
| 763 | " but got (shape: %s). Reshaping to (shape: %s).", |
| 764 | original_shape, |
| 765 | x.shape, |
| 766 | ) |
| 767 | result = self._xielu_cuda_obj.forward( |
| 768 | x, |
| 769 | self.alpha_p, |
| 770 | self.alpha_n, |
| 771 | # Temporary until xIELU CUDA fully implemented -> |
| 772 | # self.{beta,eps}.item() |
| 773 | self._beta_scalar, |
| 774 | self._eps_scalar, |
| 775 | self.with_vector_loads, |
| 776 | ) |
| 777 | return result.view(original_shape) |
| 778 | |
| 779 | def forward_native(self, input: torch.Tensor) -> torch.Tensor: |
| 780 | if self._xielu_cuda_obj is not None and input.is_cuda: |
| 781 | if not torch._dynamo.is_compiling(): |
| 782 | return self._xielu_cuda_fn(input) |
| 783 | else: |
| 784 | logger.warning_once( |
| 785 | "torch._dynamo is compiling, using Python version of xIELU." |
| 786 | ) |
| 787 | return self._xielu_python(input) |
| 788 | |
| 789 | def forward_cuda(self, input: torch.Tensor) -> torch.Tensor: |
| 790 | return self.forward_native(input) |
| 791 | |
| 792 | |
| 793 | _ACTIVATION_REGISTRY = LazyDict( |
| 794 | { |
| 795 | "gelu": lambda: GELU(), |
| 796 | "gelu_fast": lambda: FastGELU(), |
| 797 | "gelu_new": lambda: NewGELU(), |
| 798 | "gelu_pytorch_tanh": lambda: _get_gelu_pytorch_tanh(), |
| 799 | "relu": lambda: nn.ReLU(), |
| 800 | "relu2": lambda: ReLUSquaredActivation(), |
| 801 | "silu": lambda: nn.SiLU(), |
| 802 | "swish": lambda: nn.SiLU(), |
| 803 | "quick_gelu": lambda: QuickGELU(), |
| 804 | "tanh": lambda: nn.Tanh(), |
| 805 | "sigmoid": lambda: nn.Sigmoid(), |
| 806 | "xielu": lambda: XIELU(), |
| 807 | } |
| 808 | ) |
| 809 | |
| 810 | |
| 811 | def _get_gelu_pytorch_tanh() -> nn.Module: |
| 812 | """Get PyTorch GELU with tanh approximation, with ROCm fallback |
| 813 | and fast GELU for ARM.""" |
| 814 | if current_platform.is_rocm(): |
| 815 | # TODO:[ROCm] PyTorch native GELU with tanh is unstable with torch.compile |
| 816 | logger.warning_once( |
| 817 | "[ROCm] PyTorch's native GELU with tanh approximation is unstable. " |
| 818 | "Falling back to GELU(approximate='none')." |
| 819 | ) |
| 820 | return nn.GELU(approximate="none") |
| 821 | if ( |
| 822 | current_platform.is_cpu() |
| 823 | and current_platform.get_cpu_architecture() == CpuArchEnum.ARM |
| 824 | ): |
| 825 | return GELUTanh() |
| 826 | return nn.GELU(approximate="tanh") |
| 827 | |
| 828 | |
| 829 | def get_act_fn(act_fn_name: str) -> nn.Module: |
| 830 | """Get an activation function by name.""" |
| 831 | act_fn_name = act_fn_name.lower() |
| 832 | |
| 833 | if act_fn_name.startswith("torch.nn.modules."): |
| 834 | activation_name = act_fn_name.split(".")[-1] |
| 835 | if activation_name == "identity": |
| 836 | return nn.Identity() |
| 837 | act_fn_name = activation_name |
| 838 | |
| 839 | if act_fn_name not in _ACTIVATION_REGISTRY: |
| 840 | raise ValueError(f"Activation function {act_fn_name!r} is not supported.") |
| 841 | |
| 842 | return _ACTIVATION_REGISTRY[act_fn_name] |
| 843 | |
| 844 | |
| 845 | _ACTIVATION_AND_MUL_REGISTRY: LazyDict[nn.Module] = LazyDict( |
| 846 | { |
| 847 | "gelu": lambda: GeluAndMul(), |
| 848 | "gelu_pytorch_tanh": lambda: GeluAndMul(approximate="tanh"), |
| 849 | "silu": lambda: SiluAndMul(), |
| 850 | "swish": lambda: SiluAndMul(), |
| 851 | "geglu": lambda: GeluAndMul(), |
| 852 | "swigluoai": lambda: SwigluOAIAndMul(), |
| 853 | } |
| 854 | ) |
| 855 | |
| 856 | |
| 857 | def get_act_and_mul_fn(act_fn_name: str, *, compile_native: bool = True) -> nn.Module: |
| 858 | """Get an activation-and-mul (i.e. SiluAndMul) function by name.""" |
| 859 | act_fn_name = act_fn_name.lower() |
| 860 | |
| 861 | if not compile_native and act_fn_name in ("silu", "swish"): |
| 862 | return SiluAndMul(compile_native=False) |
| 863 | |
| 864 | try: |
| 865 | return _ACTIVATION_AND_MUL_REGISTRY[act_fn_name] |
| 866 | except KeyError: |
| 867 | raise ValueError( |
| 868 | f"Activation function {act_fn_name!r} is not supported." |
| 869 | ) from None |