| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from abc import ABC, abstractmethod |
| 4 | from typing import Any |
| 5 | |
| 6 | import torch |
| 7 | from torch._higher_order_ops import auto_functionalized |
| 8 | from torch._ops import OpOverload |
| 9 | |
| 10 | from vllm._aiter_ops import rocm_aiter_ops |
| 11 | from vllm.config import get_current_vllm_config |
| 12 | from vllm.model_executor.layers.activation import SiluAndMul |
| 13 | from vllm.model_executor.layers.layernorm import RMSNormGated |
| 14 | from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8 |
| 15 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 16 | GroupShape, |
| 17 | QuantKey, |
| 18 | _normalize_quant_group_shape, |
| 19 | kFp8Dynamic64Sym, |
| 20 | kFp8Dynamic128Sym, |
| 21 | kFp8DynamicTensorSym, |
| 22 | kFp8DynamicTokenSym, |
| 23 | kFp8StaticTensorSym, |
| 24 | kNvfp4Dynamic, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding |
| 27 | from vllm.model_executor.layers.rotary_embedding.deepseek_scaling_rope import ( |
| 28 | DeepseekScalingRotaryEmbedding, |
| 29 | ) |
| 30 | from vllm.platforms import current_platform |
| 31 | |
| 32 | ROTARY_OP = torch.ops._C.rotary_embedding.default |
| 33 | FLASHINFER_ROTARY_OP = torch.ops.vllm.flashinfer_rotary_embedding.default |
| 34 | MROPE_OP = torch.ops.vllm.mrope.default |
| 35 | |
| 36 | QUANT_OPS: dict[QuantKey, OpOverload] = { |
| 37 | kFp8StaticTensorSym: torch.ops._C.static_scaled_fp8_quant.default, # noqa: E501 |
| 38 | kFp8DynamicTensorSym: torch.ops._C.dynamic_scaled_fp8_quant.default, # noqa: E501 |
| 39 | kFp8DynamicTokenSym: torch.ops._C.dynamic_per_token_scaled_fp8_quant.default, # noqa: E501 |
| 40 | } |
| 41 | |
| 42 | if hasattr(torch.ops._C, "per_token_group_fp8_quant"): |
| 43 | QUANT_OPS[kFp8Dynamic128Sym] = torch.ops._C.per_token_group_fp8_quant.default # noqa: E501 |
| 44 | QUANT_OPS[kFp8Dynamic64Sym] = torch.ops._C.per_token_group_fp8_quant.default # noqa: E501 |
| 45 | |
| 46 | if current_platform.is_cuda() and hasattr(torch.ops._C, "scaled_fp4_quant"): |
| 47 | QUANT_OPS[kNvfp4Dynamic] = torch.ops._C.scaled_fp4_quant.out # noqa: E501 |
| 48 | |
| 49 | |
| 50 | SILU_MUL_OP = torch.ops._C.silu_and_mul.default |
| 51 | |
| 52 | |
| 53 | class MatcherCustomOp(ABC): |
| 54 | def __init__(self, enabled: bool) -> None: |
| 55 | config = get_current_vllm_config() |
| 56 | self.model_dtype = config.model_config.dtype if config.model_config else None |
| 57 | self.device = config.device_config.device if config.device_config else None |
| 58 | |
| 59 | self.enabled = enabled |
| 60 | self.forward = self.forward_custom if enabled else self.forward_native |
| 61 | |
| 62 | @abstractmethod |
| 63 | def forward_custom(self, *args: Any, **kwargs: Any) -> Any: |
| 64 | pass |
| 65 | |
| 66 | @abstractmethod |
| 67 | def forward_native(self, *args: Any, **kwargs: Any) -> Any: |
| 68 | pass |
| 69 | |
| 70 | def __call__(self, *args: Any, **kwargs: Any) -> Any: |
| 71 | return self.forward(*args, **kwargs) |
| 72 | |
| 73 | def empty(self, *args: Any, **kwargs: Any) -> torch.Tensor: |
| 74 | return torch.empty(*args, dtype=self.model_dtype, device=self.device, **kwargs) |
| 75 | |
| 76 | def empty_int64(self, *args: Any, **kwargs: Any) -> torch.Tensor: |
| 77 | return torch.empty(*args, dtype=torch.int64, device=self.device, **kwargs) |
| 78 | |
| 79 | def empty_f32(self, *args: Any, **kwargs: Any) -> torch.Tensor: |
| 80 | return torch.empty(*args, dtype=torch.float32, device=self.device, **kwargs) |
| 81 | |
| 82 | def inputs(self) -> list[torch.Tensor]: |
| 83 | """Utility for inputs to the pattern.""" |
| 84 | raise NotImplementedError |
| 85 | |
| 86 | |
| 87 | class MatcherRotaryEmbedding(MatcherCustomOp): |
| 88 | def __init__( |
| 89 | self, |
| 90 | is_neox: bool, |
| 91 | head_size: int, |
| 92 | num_heads: int, |
| 93 | num_kv_heads: int, |
| 94 | use_flashinfer: bool = False, |
| 95 | match_rocm_aiter: bool | None = None, |
| 96 | enabled: bool | None = None, |
| 97 | ) -> None: |
| 98 | if enabled is None: |
| 99 | enabled = RotaryEmbedding.enabled() |
| 100 | if match_rocm_aiter is None: |
| 101 | match_rocm_aiter = rocm_aiter_ops.is_triton_rotary_embed_enabled() |
| 102 | |
| 103 | super().__init__(enabled) |
| 104 | self.is_neox = is_neox |
| 105 | self.head_size = head_size |
| 106 | self.num_heads = num_heads |
| 107 | self.num_kv_heads = num_kv_heads |
| 108 | self.q_size = self.num_heads * self.head_size |
| 109 | self.kv_size = self.num_kv_heads * self.head_size |
| 110 | self.rotary_dim = head_size |
| 111 | if use_flashinfer: |
| 112 | self.rotary_op = FLASHINFER_ROTARY_OP |
| 113 | elif match_rocm_aiter: |
| 114 | self.rotary_op = rocm_aiter_ops.get_triton_rotary_embedding_op() |
| 115 | else: |
| 116 | self.rotary_op = ROTARY_OP |
| 117 | |
| 118 | def inputs(self) -> list[torch.Tensor]: |
| 119 | positions = self.empty_int64(5) |
| 120 | query = self.empty(5, self.q_size) |
| 121 | key = self.empty(5, self.kv_size) |
| 122 | cos_sin_cache = self.empty(4096, self.rotary_dim) |
| 123 | return [positions, query, key, cos_sin_cache] |
| 124 | |
| 125 | def forward_custom( |
| 126 | self, |
| 127 | positions: torch.Tensor, |
| 128 | query: torch.Tensor, |
| 129 | key: torch.Tensor | None, |
| 130 | cos_sin_cache: torch.Tensor, |
| 131 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 132 | result = auto_functionalized( |
| 133 | self.rotary_op, |
| 134 | positions=positions, |
| 135 | query=query, |
| 136 | key=key, |
| 137 | head_size=self.head_size, |
| 138 | cos_sin_cache=cos_sin_cache, |
| 139 | is_neox=self.is_neox, |
| 140 | ) |
| 141 | query_out = result[1] |
| 142 | key_out = result[2] if len(result) > 2 else None |
| 143 | return query_out, key_out |
| 144 | |
| 145 | def forward_native( |
| 146 | self, |
| 147 | positions: torch.Tensor, |
| 148 | query: torch.Tensor, |
| 149 | key: torch.Tensor | None, |
| 150 | cos_sin_cache: torch.Tensor, |
| 151 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 152 | result: tuple[torch.Tensor, torch.Tensor | None] = ( |
| 153 | RotaryEmbedding.forward_static( |
| 154 | positions, |
| 155 | query, |
| 156 | key, |
| 157 | self.head_size, |
| 158 | self.rotary_dim, |
| 159 | cos_sin_cache, |
| 160 | self.is_neox, |
| 161 | ) |
| 162 | ) |
| 163 | return result |
| 164 | |
| 165 | |
| 166 | class MatcherMRotaryEmbedding(MatcherCustomOp): |
| 167 | def __init__( |
| 168 | self, |
| 169 | is_neox: bool, |
| 170 | head_size: int, |
| 171 | rotary_dim: int, |
| 172 | num_heads: int, |
| 173 | num_kv_heads: int, |
| 174 | mrope_section: tuple[int, int, int], |
| 175 | mrope_interleaved: bool, |
| 176 | enabled: bool | None = None, |
| 177 | ) -> None: |
| 178 | if enabled is None: |
| 179 | enabled = RotaryEmbedding.enabled() |
| 180 | |
| 181 | super().__init__(enabled) |
| 182 | self.is_neox = is_neox |
| 183 | self.head_size = head_size |
| 184 | self.rotary_dim = rotary_dim |
| 185 | self.num_heads = num_heads |
| 186 | self.num_kv_heads = num_kv_heads |
| 187 | self.q_size = self.num_heads * self.head_size |
| 188 | self.kv_size = self.num_kv_heads * self.head_size |
| 189 | self.mrope_section = mrope_section |
| 190 | self.mrope_interleaved = mrope_interleaved |
| 191 | |
| 192 | def inputs(self) -> list[torch.Tensor]: |
| 193 | positions = self.empty_int64(3, 5) |
| 194 | query = self.empty(5, self.q_size) |
| 195 | key = self.empty(5, self.kv_size) |
| 196 | cos_sin_cache = self.empty(4096, self.rotary_dim) |
| 197 | return [positions, query, key, cos_sin_cache] |
| 198 | |
| 199 | def forward_custom( |
| 200 | self, |
| 201 | positions: torch.Tensor, |
| 202 | query: torch.Tensor, |
| 203 | key: torch.Tensor, |
| 204 | cos_sin_cache: torch.Tensor, |
| 205 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 206 | result = auto_functionalized( |
| 207 | MROPE_OP, |
| 208 | positions=positions, |
| 209 | query=query, |
| 210 | key=key, |
| 211 | cos_sin_cache=cos_sin_cache, |
| 212 | head_size=self.head_size, |
| 213 | rotary_dim=self.rotary_dim, |
| 214 | mrope_section_t=self.mrope_section[0], |
| 215 | mrope_section_h=self.mrope_section[1], |
| 216 | mrope_section_w=self.mrope_section[2], |
| 217 | mrope_interleaved=self.mrope_interleaved, |
| 218 | is_neox_style=self.is_neox, |
| 219 | ) |
| 220 | return result[1], result[2] |
| 221 | |
| 222 | def forward_native( |
| 223 | self, |
| 224 | positions: torch.Tensor, |
| 225 | query: torch.Tensor, |
| 226 | key: torch.Tensor, |
| 227 | cos_sin_cache: torch.Tensor, |
| 228 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 229 | raise NotImplementedError( |
| 230 | "MRoPE fusion requires the rotary_embedding custom op" |
| 231 | ) |
| 232 | |
| 233 | |
| 234 | class MatcherRMSNormGated(MatcherCustomOp): |
| 235 | """Matches RMSNormGated with norm_before_gate=True and group_size=None.""" |
| 236 | |
| 237 | def __init__( |
| 238 | self, |
| 239 | epsilon: float, |
| 240 | enabled: bool | None = None, |
| 241 | norm_before_gate: bool = True, |
| 242 | group_size: int | None = None, |
| 243 | ) -> None: |
| 244 | if enabled is None: |
| 245 | enabled = RMSNormGated.enabled() |
| 246 | |
| 247 | super().__init__(enabled) |
| 248 | self.epsilon = epsilon |
| 249 | self.norm_before_gate = norm_before_gate |
| 250 | self.group_size = group_size |
| 251 | |
| 252 | def inputs(self) -> list[torch.Tensor]: |
| 253 | x = self.empty(5, 16) |
| 254 | z = self.empty(5, 16) |
| 255 | weight = self.empty(16) |
| 256 | return [x, z, weight] |
| 257 | |
| 258 | def forward_custom( |
| 259 | self, |
| 260 | x: torch.Tensor, |
| 261 | z: torch.Tensor, |
| 262 | weight: torch.Tensor, |
| 263 | ) -> torch.Tensor: |
| 264 | from vllm.third_party.flash_linear_attention.ops.layernorm_guard import ( |
| 265 | rmsnorm_fn, |
| 266 | ) |
| 267 | |
| 268 | return rmsnorm_fn( |
| 269 | x, |
| 270 | weight, |
| 271 | bias=None, |
| 272 | z=z, |
| 273 | eps=self.epsilon, |
| 274 | group_size=self.group_size, |
| 275 | norm_before_gate=self.norm_before_gate, |
| 276 | ) |
| 277 | |
| 278 | def forward_native( |
| 279 | self, |
| 280 | x: torch.Tensor, |
| 281 | z: torch.Tensor, |
| 282 | weight: torch.Tensor, |
| 283 | ) -> torch.Tensor: |
| 284 | return RMSNormGated.forward_static( |
| 285 | x, |
| 286 | z, |
| 287 | weight, |
| 288 | self.epsilon, |
| 289 | self.model_dtype, |
| 290 | group_size=self.group_size, |
| 291 | norm_before_gate=self.norm_before_gate, |
| 292 | ) |
| 293 | |
| 294 | |
| 295 | class MatcherDeepseekScalingRotaryEmbedding(MatcherCustomOp): |
| 296 | def __init__( |
| 297 | self, |
| 298 | is_neox: bool, |
| 299 | head_size: int, |
| 300 | num_heads: int, |
| 301 | num_kv_heads: int, |
| 302 | use_flashinfer: bool = False, |
| 303 | enabled: bool | None = None, |
| 304 | ) -> None: |
| 305 | if enabled is None: |
| 306 | enabled = DeepseekScalingRotaryEmbedding.enabled() |
| 307 | |
| 308 | super().__init__(enabled) |
| 309 | self.is_neox = is_neox |
| 310 | self.head_size = head_size |
| 311 | self.num_heads = num_heads |
| 312 | self.num_kv_heads = num_kv_heads |
| 313 | self.q_size = self.num_heads * self.head_size |
| 314 | self.kv_size = self.num_kv_heads * self.head_size |
| 315 | self.rotary_dim = head_size |
| 316 | self.use_flashinfer = use_flashinfer |
| 317 | |
| 318 | def inputs(self) -> list[torch.Tensor]: |
| 319 | positions = self.empty_int64(5) |
| 320 | query = self.empty(5, self.num_heads, self.head_size) |
| 321 | key = self.empty(5, self.num_kv_heads, self.head_size) |
| 322 | cos_sin_cache = self.empty(4096, self.rotary_dim) |
| 323 | return [positions, query, key, cos_sin_cache] |
| 324 | |
| 325 | def forward_custom( |
| 326 | self, |
| 327 | positions: torch.Tensor, |
| 328 | query: torch.Tensor, |
| 329 | key: torch.Tensor | None, |
| 330 | cos_sin_cache: torch.Tensor, |
| 331 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 332 | if self.use_flashinfer: |
| 333 | torch.ops.vllm.flashinfer_rotary_embedding( |
| 334 | positions, |
| 335 | query, |
| 336 | key, |
| 337 | self.head_size, |
| 338 | cos_sin_cache, |
| 339 | self.is_neox, |
| 340 | ) |
| 341 | return query, key |
| 342 | result: tuple[torch.Tensor, torch.Tensor | None] = ( |
| 343 | DeepseekScalingRotaryEmbedding.forward_static( |
| 344 | positions, |
| 345 | query, |
| 346 | key, |
| 347 | self.head_size, |
| 348 | self.rotary_dim, |
| 349 | cos_sin_cache, |
| 350 | self.is_neox, |
| 351 | ) |
| 352 | ) |
| 353 | return result |
| 354 | |
| 355 | def forward_native( |
| 356 | self, |
| 357 | positions: torch.Tensor, |
| 358 | query: torch.Tensor, |
| 359 | key: torch.Tensor | None, |
| 360 | cos_sin_cache: torch.Tensor, |
| 361 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 362 | result: tuple[torch.Tensor, torch.Tensor | None] = ( |
| 363 | DeepseekScalingRotaryEmbedding.forward_static( |
| 364 | positions, |
| 365 | query, |
| 366 | key, |
| 367 | self.head_size, |
| 368 | self.rotary_dim, |
| 369 | cos_sin_cache, |
| 370 | self.is_neox, |
| 371 | ) |
| 372 | ) |
| 373 | return result |
| 374 | |
| 375 | |
| 376 | class MatcherQuantFP8(MatcherCustomOp): |
| 377 | def __init__( |
| 378 | self, |
| 379 | quant_key: QuantKey, |
| 380 | enabled: bool | None = None, |
| 381 | has_col_major_scales: bool = False, |
| 382 | is_e8m0: bool = False, |
| 383 | match_rocm_aiter: bool = False, |
| 384 | is_tma_aligned: bool = False, |
| 385 | ) -> None: |
| 386 | if enabled is None: |
| 387 | enabled = QuantFP8.enabled() |
| 388 | |
| 389 | super().__init__(enabled) |
| 390 | self.quant_key = quant_key |
| 391 | self.has_col_major_scales = has_col_major_scales |
| 392 | self.is_e8m0 = is_e8m0 |
| 393 | self.match_rocm_aiter = match_rocm_aiter |
| 394 | self.is_tma_aligned = is_tma_aligned |
| 395 | |
| 396 | if match_rocm_aiter: |
| 397 | assert not quant_key.scale.group_shape.is_per_tensor(), ( |
| 398 | "ROCm aiter fusion pass does not support per tensor quantization" |
| 399 | ) |
| 400 | if quant_key.scale.group_shape.is_per_token(): |
| 401 | self.QUANT_OP = rocm_aiter_ops.get_per_token_quant_op() |
| 402 | else: |
| 403 | assert quant_key.scale.group_shape.col == 128, ( |
| 404 | "ROCm aiter fusion pass currently supports " |
| 405 | "quantization operation with group_size 128" |
| 406 | ) |
| 407 | self.QUANT_OP = rocm_aiter_ops.get_group_quant_op() |
| 408 | |
| 409 | else: |
| 410 | assert quant_key in QUANT_OPS, ( |
| 411 | f"unsupported quantization scheme {quant_key}" |
| 412 | ) |
| 413 | self.QUANT_OP = QUANT_OPS[quant_key] |
| 414 | |
| 415 | assert quant_key.dtype == current_platform.fp8_dtype(), ( |
| 416 | "Only QuantFP8 supported by" |
| 417 | ) |
| 418 | assert quant_key.scale2 is None |
| 419 | |
| 420 | self.quant_fp8 = QuantFP8( |
| 421 | quant_key.scale.static, |
| 422 | quant_key.scale.group_shape, |
| 423 | column_major_scales=has_col_major_scales, |
| 424 | use_ue8m0=is_e8m0, |
| 425 | tma_aligned_scales=self.is_tma_aligned, |
| 426 | compile_native=False, |
| 427 | ) |
| 428 | |
| 429 | def forward_rocm_aiter( |
| 430 | self, |
| 431 | input: torch.Tensor, |
| 432 | scale: torch.Tensor | None = None, |
| 433 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 434 | quant_key_group_shape = self.quant_key.scale.group_shape |
| 435 | if quant_key_group_shape == GroupShape.PER_TOKEN: |
| 436 | return self.QUANT_OP( # type: ignore[no-any-return] |
| 437 | x=input, |
| 438 | quant_dtype=self.quant_key.dtype, |
| 439 | scale=scale, |
| 440 | ) |
| 441 | else: |
| 442 | return self.QUANT_OP(input, quant_key_group_shape.col) # type: ignore[no-any-return] |
| 443 | |
| 444 | def forward_custom( |
| 445 | self, |
| 446 | input: torch.Tensor, |
| 447 | scale: torch.Tensor | None = None, |
| 448 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 449 | if self.match_rocm_aiter: |
| 450 | return self.forward_rocm_aiter(input, scale) |
| 451 | |
| 452 | result = torch.empty( |
| 453 | input.shape, device=input.device, dtype=self.quant_key.dtype |
| 454 | ) |
| 455 | |
| 456 | if self.quant_key.scale.group_shape.is_per_group(): |
| 457 | # for tma_aligned, the scale must be passed to forward_custom |
| 458 | # tma_aligned fusion then matches by custom op arguments |
| 459 | if not self.is_tma_aligned: |
| 460 | assert scale is None |
| 461 | scale = self.make_scale(input, transposed=self.has_col_major_scales) |
| 462 | |
| 463 | finfo = torch.finfo(self.quant_key.dtype) |
| 464 | fp8_min = finfo.min |
| 465 | fp8_max = finfo.max |
| 466 | |
| 467 | _, result, scale = auto_functionalized( |
| 468 | self.QUANT_OP, |
| 469 | input=input, |
| 470 | output_q=result, |
| 471 | output_s=scale, |
| 472 | group_size=self.quant_key.scale.group_shape[1], |
| 473 | eps=1e-10, |
| 474 | fp8_min=fp8_min, |
| 475 | fp8_max=fp8_max, |
| 476 | scale_ue8m0=self.is_e8m0, |
| 477 | dummy_is_scale_transposed=self.has_col_major_scales, |
| 478 | dummy_is_tma_aligned=self.is_tma_aligned, |
| 479 | ) |
| 480 | return result, scale |
| 481 | |
| 482 | if self.quant_key.scale.static: |
| 483 | assert scale is not None |
| 484 | _, result = auto_functionalized( |
| 485 | self.QUANT_OP, result=result, input=input, scale=scale |
| 486 | ) |
| 487 | return result, scale |
| 488 | else: |
| 489 | assert scale is None |
| 490 | scale = self.make_scale(input) |
| 491 | _, result, scale = auto_functionalized( |
| 492 | self.QUANT_OP, result=result, input=input, scale=scale, scale_ub=None |
| 493 | ) |
| 494 | return result, scale |
| 495 | |
| 496 | def forward_native( |
| 497 | self, |
| 498 | input: torch.Tensor, |
| 499 | scale: torch.Tensor | None = None, |
| 500 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 501 | return self.quant_fp8(input, scale) # type: ignore[no-any-return] |
| 502 | |
| 503 | def make_scale(self, input: torch.Tensor, transposed: bool = False) -> torch.Tensor: |
| 504 | normalized_group_shape = _normalize_quant_group_shape( |
| 505 | input, self.quant_key.scale.group_shape |
| 506 | ) |
| 507 | scale_shape = ( |
| 508 | input.shape[0] // normalized_group_shape[0], |
| 509 | input.shape[1] // normalized_group_shape[1], |
| 510 | ) |
| 511 | if transposed: |
| 512 | scale_shape = tuple(reversed(scale_shape)) |
| 513 | return torch.empty( |
| 514 | scale_shape, device=input.device, dtype=torch.float32 |
| 515 | ).permute(-1, -2) |
| 516 | |
| 517 | return torch.empty(scale_shape, device=input.device, dtype=torch.float32) |
| 518 | |
| 519 | def inputs(self) -> list[torch.Tensor]: |
| 520 | input = self.empty(5, 16) |
| 521 | if self.quant_key.scale.static: |
| 522 | return [input, self.empty_f32(1, 1)] |
| 523 | |
| 524 | return [input] |
| 525 | |
| 526 | |
| 527 | class MatcherSiluAndMul(MatcherCustomOp): |
| 528 | def __init__(self, enabled: bool | None = None) -> None: |
| 529 | if enabled is None: |
| 530 | enabled = SiluAndMul.enabled() |
| 531 | super().__init__(enabled) |
| 532 | |
| 533 | def inputs(self) -> list[torch.Tensor]: |
| 534 | input = self.empty(5, 4) |
| 535 | return [input] |
| 536 | |
| 537 | def forward_custom( |
| 538 | self, |
| 539 | x: torch.Tensor, |
| 540 | ) -> torch.Tensor: |
| 541 | d = x.shape[-1] // 2 |
| 542 | output_shape = x.shape[:-1] + (d,) |
| 543 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 544 | result = auto_functionalized(SILU_MUL_OP, result=out, input=x) |
| 545 | return result[1] |
| 546 | |
| 547 | def forward_native( |
| 548 | self, |
| 549 | x: torch.Tensor, |
| 550 | ) -> torch.Tensor: |
| 551 | return SiluAndMul.forward_native(x) |