| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | |
| 5 | import numpy as np |
| 6 | import torch |
| 7 | |
| 8 | from vllm.platforms import current_platform |
| 9 | from vllm.triton_utils import tl, triton |
| 10 | from vllm.utils.torch_utils import direct_register_custom_op |
| 11 | |
| 12 | from .base import RotaryEmbeddingBase |
| 13 | from .yarn_scaling_rope import YaRNScalingRotaryEmbedding, yarn_get_mscale |
| 14 | |
| 15 | |
| 16 | @triton.jit |
| 17 | def _triton_mrope_forward( |
| 18 | q_ptr, |
| 19 | k_ptr, |
| 20 | cos, |
| 21 | sin, |
| 22 | num_tokens, |
| 23 | q_token_stride, |
| 24 | k_token_stride, |
| 25 | n_qh: tl.constexpr, |
| 26 | n_kh: tl.constexpr, |
| 27 | hd: tl.constexpr, |
| 28 | rd: tl.constexpr, |
| 29 | pad_n_qh: tl.constexpr, |
| 30 | pad_n_kh: tl.constexpr, |
| 31 | pad_rd: tl.constexpr, |
| 32 | mrope_section_t: tl.constexpr, |
| 33 | mrope_section_h: tl.constexpr, |
| 34 | mrope_section_w: tl.constexpr, |
| 35 | is_interleaved: tl.constexpr, |
| 36 | is_neox_style: tl.constexpr, |
| 37 | ): |
| 38 | # Adapted from |
| 39 | # https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/qwen2vl_mrope.py |
| 40 | # This version supports flatten input tensors from vllm |
| 41 | # and supports cos and sin cache with shape (3, num_tokens, rotary_dim // 2) |
| 42 | # instead of (3, bsz, seq_len, head_dim), also supports interleaved rotary |
| 43 | pid = tl.program_id(0) |
| 44 | # locate start address (int64: num_tokens * row_stride exceeds 2^31 for |
| 45 | # worst-case ViT forwards, e.g. MiniMax M3 packs ~768k patches/item) |
| 46 | q_ptr = q_ptr + pid.to(tl.int64) * q_token_stride |
| 47 | k_ptr = k_ptr + pid.to(tl.int64) * k_token_stride |
| 48 | |
| 49 | # #################################################################### |
| 50 | # get the cos(mθ_{i...d/2}) and sin(mθ_{i...d/2}) for token position |
| 51 | # m of this program instance |
| 52 | # #################################################################### |
| 53 | # Note: cos and sin now have shape (3, num_tokens, rotary_dim // 2) |
| 54 | |
| 55 | # Updated stride calculation for half rotary_dim |
| 56 | half_rd = rd // 2 |
| 57 | t_cos = cos + pid * half_rd |
| 58 | h_cos = t_cos + num_tokens * half_rd |
| 59 | w_cos = h_cos + num_tokens * half_rd |
| 60 | t_sin = sin + pid * half_rd |
| 61 | h_sin = t_sin + num_tokens * half_rd |
| 62 | w_sin = h_sin + num_tokens * half_rd |
| 63 | |
| 64 | # Updated offsets for half rotary_dim |
| 65 | cos_offsets = tl.arange(0, pad_rd // 2) |
| 66 | if is_interleaved: |
| 67 | valid_mask = cos_offsets < half_rd |
| 68 | h_mask = ( |
| 69 | valid_mask & ((cos_offsets % 3) == 1) & (cos_offsets <= 3 * mrope_section_h) |
| 70 | ) |
| 71 | w_mask = ( |
| 72 | valid_mask & ((cos_offsets % 3) == 2) & (cos_offsets <= 3 * mrope_section_w) |
| 73 | ) |
| 74 | t_mask = valid_mask & ~(h_mask | w_mask) |
| 75 | else: |
| 76 | t_end = mrope_section_t |
| 77 | h_end = t_end + mrope_section_h |
| 78 | t_mask = cos_offsets < mrope_section_t |
| 79 | h_mask = (t_end <= cos_offsets) & (cos_offsets < h_end) |
| 80 | w_mask = (h_end <= cos_offsets) & (cos_offsets < half_rd) |
| 81 | |
| 82 | t_cos_row = tl.load(t_cos + cos_offsets, mask=t_mask, other=0) |
| 83 | h_cos_row = tl.load(h_cos + cos_offsets, mask=h_mask, other=0) |
| 84 | w_cos_row = tl.load(w_cos + cos_offsets, mask=w_mask, other=0) |
| 85 | t_sin_row = tl.load(t_sin + cos_offsets, mask=t_mask, other=0) |
| 86 | h_sin_row = tl.load(h_sin + cos_offsets, mask=h_mask, other=0) |
| 87 | w_sin_row = tl.load(w_sin + cos_offsets, mask=w_mask, other=0) |
| 88 | |
| 89 | cos_row = t_cos_row + h_cos_row + w_cos_row |
| 90 | sin_row = t_sin_row + h_sin_row + w_sin_row |
| 91 | |
| 92 | # #################################################################### |
| 93 | # Load the two values in each rotary pair for the current token. |
| 94 | # NeoX pairs the first and second halves, while GPT-J pairs |
| 95 | # adjacent values. |
| 96 | # #################################################################### |
| 97 | if is_neox_style: |
| 98 | rotary_offsets = tl.arange(0, pad_rd // 2) |
| 99 | first_q_offsets = tl.arange(0, pad_n_qh)[:, None] * hd + rotary_offsets[None, :] |
| 100 | first_k_offsets = tl.arange(0, pad_n_kh)[:, None] * hd + rotary_offsets[None, :] |
| 101 | first_q_mask = (tl.arange(0, pad_n_qh)[:, None] < n_qh) & ( |
| 102 | rotary_offsets[None, :] < rd // 2 |
| 103 | ) |
| 104 | first_k_mask = (tl.arange(0, pad_n_kh)[:, None] < n_kh) & ( |
| 105 | rotary_offsets[None, :] < rd // 2 |
| 106 | ) |
| 107 | |
| 108 | q_tile_1 = tl.load(q_ptr + first_q_offsets, mask=first_q_mask, other=0).to( |
| 109 | sin_row.dtype |
| 110 | ) |
| 111 | k_tile_1 = tl.load(k_ptr + first_k_offsets, mask=first_k_mask, other=0).to( |
| 112 | sin_row.dtype |
| 113 | ) |
| 114 | |
| 115 | second_q_offsets = first_q_offsets + (rd // 2) |
| 116 | second_k_offsets = first_k_offsets + (rd // 2) |
| 117 | q_tile_2 = tl.load(q_ptr + second_q_offsets, mask=first_q_mask, other=0).to( |
| 118 | sin_row.dtype |
| 119 | ) |
| 120 | k_tile_2 = tl.load(k_ptr + second_k_offsets, mask=first_k_mask, other=0).to( |
| 121 | sin_row.dtype |
| 122 | ) |
| 123 | |
| 124 | new_q_tile_1 = q_tile_1 * cos_row - q_tile_2 * sin_row |
| 125 | tl.store(q_ptr + first_q_offsets, new_q_tile_1, mask=first_q_mask) |
| 126 | new_q_tile_2 = q_tile_2 * cos_row + q_tile_1 * sin_row |
| 127 | tl.store(q_ptr + second_q_offsets, new_q_tile_2, mask=first_q_mask) |
| 128 | |
| 129 | new_k_tile_1 = k_tile_1 * cos_row - k_tile_2 * sin_row |
| 130 | tl.store(k_ptr + first_k_offsets, new_k_tile_1, mask=first_k_mask) |
| 131 | new_k_tile_2 = k_tile_2 * cos_row + k_tile_1 * sin_row |
| 132 | tl.store(k_ptr + second_k_offsets, new_k_tile_2, mask=first_k_mask) |
| 133 | else: |
| 134 | # Load and store adjacent rotary pairs contiguously. Using stride-two |
| 135 | # even/odd offsets makes Triton emit scalar 16-bit memory operations on |
| 136 | # AMD, while split/interleave only rearranges values in registers. |
| 137 | rotary_offsets = tl.arange(0, pad_rd) |
| 138 | q_offsets = tl.arange(0, pad_n_qh)[:, None] * hd + rotary_offsets[None, :] |
| 139 | k_offsets = tl.arange(0, pad_n_kh)[:, None] * hd + rotary_offsets[None, :] |
| 140 | q_mask = (tl.arange(0, pad_n_qh)[:, None] < n_qh) & ( |
| 141 | rotary_offsets[None, :] < rd |
| 142 | ) |
| 143 | k_mask = (tl.arange(0, pad_n_kh)[:, None] < n_kh) & ( |
| 144 | rotary_offsets[None, :] < rd |
| 145 | ) |
| 146 | |
| 147 | q_tile = tl.load(q_ptr + q_offsets, mask=q_mask, other=0).to(sin_row.dtype) |
| 148 | k_tile = tl.load(k_ptr + k_offsets, mask=k_mask, other=0).to(sin_row.dtype) |
| 149 | q_tile_1, q_tile_2 = tl.split(tl.reshape(q_tile, (pad_n_qh, pad_rd // 2, 2))) |
| 150 | k_tile_1, k_tile_2 = tl.split(tl.reshape(k_tile, (pad_n_kh, pad_rd // 2, 2))) |
| 151 | |
| 152 | new_q_tile_1 = q_tile_1 * cos_row - q_tile_2 * sin_row |
| 153 | new_q_tile_2 = q_tile_2 * cos_row + q_tile_1 * sin_row |
| 154 | new_q_tile = tl.interleave(new_q_tile_1, new_q_tile_2) |
| 155 | tl.store(q_ptr + q_offsets, new_q_tile, mask=q_mask) |
| 156 | |
| 157 | new_k_tile_1 = k_tile_1 * cos_row - k_tile_2 * sin_row |
| 158 | new_k_tile_2 = k_tile_2 * cos_row + k_tile_1 * sin_row |
| 159 | new_k_tile = tl.interleave(new_k_tile_1, new_k_tile_2) |
| 160 | tl.store(k_ptr + k_offsets, new_k_tile, mask=k_mask) |
| 161 | |
| 162 | |
| 163 | def triton_mrope( |
| 164 | q: torch.Tensor, |
| 165 | k: torch.Tensor, |
| 166 | cos: torch.Tensor, |
| 167 | sin: torch.Tensor, |
| 168 | mrope_section: list[int], |
| 169 | head_size: int, |
| 170 | rotary_dim: int, |
| 171 | mrope_interleaved: bool, |
| 172 | is_neox_style: bool, |
| 173 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 174 | """Qwen2VL mrope kernel. |
| 175 | |
| 176 | Args: |
| 177 | q: [num_tokens, num_heads * head_size] |
| 178 | k: [num_tokens, num_kv_heads * head_size] |
| 179 | cos: [3, num_tokens, rotary_dim // 2] |
| 180 | (T/H/W positions with multimodal inputs) |
| 181 | sin: [3, num_tokens, rotary_dim // 2] |
| 182 | (T/H/W positions with multimodal inputs) |
| 183 | mrope_section: [t, h, w] |
| 184 | head_size: int |
| 185 | rotary_dim: Number of leading dimensions rotary is applied to. |
| 186 | mrope_interleaved: Whether the T/H/W sections are interleaved rather |
| 187 | than concatenated. |
| 188 | is_neox_style: Whether rotary pairs use split-half (NeoX) or |
| 189 | adjacent (GPT-J) layout. |
| 190 | |
| 191 | """ |
| 192 | # The kernel indexes each token row through an explicit row stride and |
| 193 | # requires dense rows. |
| 194 | assert q.dim() == 2 and q.stride(-1) == 1 |
| 195 | assert k.dim() == 2 and k.stride(-1) == 1 |
| 196 | n_row, n_q_head_head_dim = q.shape |
| 197 | n_q_head = n_q_head_head_dim // head_size |
| 198 | n_kv_head = k.shape[1] // head_size |
| 199 | pad_rd = triton.next_power_of_2(rotary_dim) |
| 200 | pad_n_q_head = triton.next_power_of_2(n_q_head) |
| 201 | pad_n_kv_head = triton.next_power_of_2(n_kv_head) |
| 202 | |
| 203 | cos = cos.contiguous() |
| 204 | sin = sin.contiguous() |
| 205 | |
| 206 | # Small adjacent-pair tiles perform best with one wave per program on |
| 207 | # ROCm. Keep the existing launch shape for larger rotary dimensions, |
| 208 | # NeoX, and other backends. |
| 209 | use_single_wave = current_platform.is_rocm() and not is_neox_style and pad_rd <= 64 |
| 210 | num_warps = 1 if use_single_wave else 4 |
| 211 | _triton_mrope_forward[(n_row,)]( |
| 212 | q, |
| 213 | k, |
| 214 | cos, |
| 215 | sin, |
| 216 | n_row, |
| 217 | q.stride(0), |
| 218 | k.stride(0), |
| 219 | n_q_head, |
| 220 | n_kv_head, |
| 221 | head_size, |
| 222 | rotary_dim, |
| 223 | pad_n_q_head, |
| 224 | pad_n_kv_head, |
| 225 | pad_rd, |
| 226 | mrope_section[0], |
| 227 | mrope_section[1], |
| 228 | mrope_section[2], |
| 229 | mrope_interleaved, |
| 230 | is_neox_style, |
| 231 | num_warps=num_warps, |
| 232 | ) |
| 233 | return q, k |
| 234 | |
| 235 | |
| 236 | def apply_interleaved_rope(x: torch.Tensor, mrope_section: list[int]) -> torch.Tensor: |
| 237 | """Apply interleaved MRoPE to 3D rotary embeddings. |
| 238 | Reorganizes frequency layout from chunked [TTT...HHH...WWW] to |
| 239 | interleaved [THWTHWTHW...TT], preserving frequency continuity. |
| 240 | """ |
| 241 | channels = torch.arange(x.shape[-1], device=x.device) |
| 242 | is_height = (channels % 3 == 1) & (channels < mrope_section[1] * 3) |
| 243 | is_width = (channels % 3 == 2) & (channels < mrope_section[2] * 3) |
| 244 | |
| 245 | result = torch.where(is_height, x[1], x[0]) |
| 246 | return torch.where(is_width, x[2], result) |
| 247 | |
| 248 | |
| 249 | def _mrope_apply( |
| 250 | positions: torch.Tensor, |
| 251 | query: torch.Tensor, |
| 252 | key: torch.Tensor, |
| 253 | cos_sin_cache: torch.Tensor, |
| 254 | head_size: int, |
| 255 | rotary_dim: int, |
| 256 | mrope_section_t: int, |
| 257 | mrope_section_h: int, |
| 258 | mrope_section_w: int, |
| 259 | mrope_interleaved: bool, |
| 260 | is_neox_style: bool, |
| 261 | ) -> None: |
| 262 | cos_sin = cos_sin_cache[positions] |
| 263 | cos, sin = cos_sin.chunk(2, dim=-1) |
| 264 | query_out, key_out = triton_mrope( |
| 265 | query, |
| 266 | key, |
| 267 | cos, |
| 268 | sin, |
| 269 | [mrope_section_t, mrope_section_h, mrope_section_w], |
| 270 | head_size, |
| 271 | rotary_dim, |
| 272 | mrope_interleaved, |
| 273 | is_neox_style, |
| 274 | ) |
| 275 | if query_out.data_ptr() != query.data_ptr(): |
| 276 | query.copy_(query_out.view_as(query)) |
| 277 | if key_out.data_ptr() != key.data_ptr(): |
| 278 | key.copy_(key_out.view_as(key)) |
| 279 | |
| 280 | |
| 281 | def _mrope_apply_fake( |
| 282 | positions: torch.Tensor, |
| 283 | query: torch.Tensor, |
| 284 | key: torch.Tensor, |
| 285 | cos_sin_cache: torch.Tensor, |
| 286 | head_size: int, |
| 287 | rotary_dim: int, |
| 288 | mrope_section_t: int, |
| 289 | mrope_section_h: int, |
| 290 | mrope_section_w: int, |
| 291 | mrope_interleaved: bool, |
| 292 | is_neox_style: bool, |
| 293 | ) -> None: |
| 294 | return |
| 295 | |
| 296 | |
| 297 | direct_register_custom_op( |
| 298 | op_name="mrope", |
| 299 | op_func=_mrope_apply, |
| 300 | mutates_args=["query", "key"], |
| 301 | fake_impl=_mrope_apply_fake, |
| 302 | ) |
| 303 | |
| 304 | |
| 305 | class MRotaryEmbedding(RotaryEmbeddingBase): |
| 306 | """Rotary Embedding with Multimodal Sections.""" |
| 307 | |
| 308 | def __init__( |
| 309 | self, |
| 310 | head_size: int, |
| 311 | rotary_dim: int, |
| 312 | max_position_embeddings: int, |
| 313 | base: float, |
| 314 | is_neox_style: bool, |
| 315 | dtype: torch.dtype, |
| 316 | mrope_section: list[int] | None = None, |
| 317 | mrope_interleaved: bool = False, |
| 318 | # YaRN parameters. |
| 319 | *, |
| 320 | scaling_factor: float | None = None, |
| 321 | beta_fast: int = 32, |
| 322 | beta_slow: int = 1, |
| 323 | mscale: float | None = None, |
| 324 | mscale_all_dim: float | None = None, |
| 325 | attention_factor: float | None = None, |
| 326 | truncate: bool = True, |
| 327 | ) -> None: |
| 328 | self.scaling_factor = scaling_factor |
| 329 | self.beta_fast = beta_fast |
| 330 | self.beta_slow = beta_slow |
| 331 | self.truncate = truncate |
| 332 | if self.scaling_factor is not None: |
| 333 | # Get n-d magnitude scaling corrected for interpolation |
| 334 | if attention_factor is not None: |
| 335 | self.mscale = float(attention_factor) |
| 336 | elif mscale and mscale_all_dim: |
| 337 | self.mscale = float( |
| 338 | yarn_get_mscale(self.scaling_factor, mscale) |
| 339 | / yarn_get_mscale(self.scaling_factor, mscale_all_dim) |
| 340 | ) |
| 341 | else: |
| 342 | self.mscale = float(yarn_get_mscale(self.scaling_factor)) |
| 343 | else: |
| 344 | self.mscale = 1.0 |
| 345 | |
| 346 | # In Qwen2.5-VL, the maximum index value is related to the duration of |
| 347 | # the input video. We enlarge max_position_embeddings to 4 times to get |
| 348 | # a larger the cos and sin cache. |
| 349 | self.cache_max_position_num = max_position_embeddings * 4 |
| 350 | super().__init__( |
| 351 | head_size, |
| 352 | rotary_dim, |
| 353 | self.cache_max_position_num, |
| 354 | base, |
| 355 | is_neox_style, |
| 356 | dtype, |
| 357 | ) |
| 358 | |
| 359 | self.mrope_section = mrope_section |
| 360 | self.mrope_interleaved = mrope_interleaved |
| 361 | if self.mrope_section: |
| 362 | assert sum(self.mrope_section) == rotary_dim // 2 |
| 363 | |
| 364 | def _compute_inv_freq(self, base: float) -> torch.Tensor: |
| 365 | if self.scaling_factor is None: |
| 366 | return super()._compute_inv_freq(base) |
| 367 | return YaRNScalingRotaryEmbedding._compute_inv_freq(self, base) |
| 368 | |
| 369 | def _compute_cos_sin_cache(self) -> torch.Tensor: |
| 370 | if self.scaling_factor is None: |
| 371 | return super()._compute_cos_sin_cache() |
| 372 | return YaRNScalingRotaryEmbedding._compute_cos_sin_cache(self) |
| 373 | |
| 374 | def forward_native( |
| 375 | self, |
| 376 | positions: torch.Tensor, |
| 377 | query: torch.Tensor, |
| 378 | key: torch.Tensor | None = None, |
| 379 | offsets: torch.Tensor | None = None, |
| 380 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 381 | """PyTorch-native implementation equivalent to forward(). |
| 382 | |
| 383 | Args: |
| 384 | positions: |
| 385 | [num_tokens,] (text only) or |
| 386 | [3, num_tokens] (T/H/W positions with multimodal inputs) |
| 387 | query: [num_tokens, num_heads * head_size] |
| 388 | key: [num_tokens, num_kv_heads * head_size] |
| 389 | offsets: Optional per-token position offsets added to positions. |
| 390 | |
| 391 | """ |
| 392 | assert positions.ndim == 1 or positions.ndim == 2 |
| 393 | assert key is not None |
| 394 | |
| 395 | cos_sin_cache = self._match_cos_sin_cache_dtype(query) |
| 396 | num_tokens = positions.shape[-1] |
| 397 | cos_sin = cos_sin_cache[positions] |
| 398 | cos, sin = cos_sin.chunk(2, dim=-1) |
| 399 | if positions.ndim == 2: |
| 400 | assert self.mrope_section |
| 401 | if self.mrope_interleaved: |
| 402 | cos = apply_interleaved_rope(cos, self.mrope_section) |
| 403 | sin = apply_interleaved_rope(sin, self.mrope_section) |
| 404 | else: |
| 405 | cos = torch.cat( |
| 406 | [m[i] for i, m in enumerate(cos.split(self.mrope_section, dim=-1))], |
| 407 | dim=-1, |
| 408 | ) |
| 409 | sin = torch.cat( |
| 410 | [m[i] for i, m in enumerate(sin.split(self.mrope_section, dim=-1))], |
| 411 | dim=-1, |
| 412 | ) |
| 413 | |
| 414 | query_shape = query.shape |
| 415 | query = query.view(num_tokens, -1, self.head_size) |
| 416 | query_rot = query[..., : self.rotary_dim] |
| 417 | query_pass = query[..., self.rotary_dim :] |
| 418 | query_rot = self.apply_rotary_emb.forward_native( |
| 419 | query_rot, |
| 420 | cos, |
| 421 | sin, |
| 422 | ) |
| 423 | query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape) |
| 424 | |
| 425 | key_shape = key.shape |
| 426 | key = key.view(num_tokens, -1, self.head_size) |
| 427 | key_rot = key[..., : self.rotary_dim] |
| 428 | key_pass = key[..., self.rotary_dim :] |
| 429 | key_rot = self.apply_rotary_emb.forward_native( |
| 430 | key_rot, |
| 431 | cos, |
| 432 | sin, |
| 433 | ) |
| 434 | key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape) |
| 435 | return query, key |
| 436 | |
| 437 | def forward_cuda( |
| 438 | self, |
| 439 | positions: torch.Tensor, |
| 440 | query: torch.Tensor, |
| 441 | key: torch.Tensor | None = None, |
| 442 | offsets: torch.Tensor | None = None, |
| 443 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 444 | assert positions.ndim == 1 or positions.ndim == 2 |
| 445 | assert key is not None |
| 446 | |
| 447 | cos_sin_cache = self._match_cos_sin_cache_dtype(query) |
| 448 | num_tokens = positions.shape[-1] |
| 449 | cos_sin = cos_sin_cache[positions] |
| 450 | cos, sin = cos_sin.chunk(2, dim=-1) |
| 451 | query_shape = query.shape |
| 452 | key_shape = key.shape |
| 453 | if positions.ndim == 2: |
| 454 | assert self.mrope_section |
| 455 | |
| 456 | q, k = triton_mrope( |
| 457 | query, |
| 458 | key, |
| 459 | cos, |
| 460 | sin, |
| 461 | self.mrope_section, |
| 462 | self.head_size, |
| 463 | self.rotary_dim, |
| 464 | self.mrope_interleaved, |
| 465 | self.is_neox_style, |
| 466 | ) |
| 467 | |
| 468 | return q.reshape(query_shape), k.reshape(key_shape) |
| 469 | |
| 470 | query = query.view(num_tokens, -1, self.head_size) |
| 471 | query_rot = query[..., : self.rotary_dim] |
| 472 | query_pass = query[..., self.rotary_dim :] |
| 473 | query_rot = self.apply_rotary_emb( |
| 474 | query_rot, |
| 475 | cos, |
| 476 | sin, |
| 477 | ) |
| 478 | query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape) |
| 479 | |
| 480 | key = key.view(num_tokens, -1, self.head_size) |
| 481 | key_rot = key[..., : self.rotary_dim] |
| 482 | key_pass = key[..., self.rotary_dim :] |
| 483 | key_rot = self.apply_rotary_emb( |
| 484 | key_rot, |
| 485 | cos, |
| 486 | sin, |
| 487 | ) |
| 488 | key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape) |
| 489 | return query, key |
| 490 | |
| 491 | def forward_hip( |
| 492 | self, |
| 493 | positions: torch.Tensor, |
| 494 | query: torch.Tensor, |
| 495 | key: torch.Tensor | None = None, |
| 496 | offsets: torch.Tensor | None = None, |
| 497 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 498 | if positions.ndim == 2: |
| 499 | assert key is not None |
| 500 | assert self.mrope_section |
| 501 | query_shape = query.shape |
| 502 | key_shape = key.shape |
| 503 | cos_sin_cache = self._match_cos_sin_cache_dtype(query) |
| 504 | torch.ops.vllm.mrope( |
| 505 | positions, |
| 506 | query, |
| 507 | key, |
| 508 | cos_sin_cache, |
| 509 | self.head_size, |
| 510 | self.rotary_dim, |
| 511 | self.mrope_section[0], |
| 512 | self.mrope_section[1], |
| 513 | self.mrope_section[2], |
| 514 | self.mrope_interleaved, |
| 515 | self.is_neox_style, |
| 516 | ) |
| 517 | return query.reshape(query_shape), key.reshape(key_shape) |
| 518 | return super().forward_hip(positions, query, key) |
| 519 | |
| 520 | def forward_cpu( |
| 521 | self, |
| 522 | positions: torch.Tensor, |
| 523 | query: torch.Tensor, |
| 524 | key: torch.Tensor | None = None, |
| 525 | offsets: torch.Tensor | None = None, |
| 526 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 527 | return self.forward_native(positions, query, key, offsets) |
| 528 | |
| 529 | def forward_xpu( |
| 530 | self, |
| 531 | positions: torch.Tensor, |
| 532 | query: torch.Tensor, |
| 533 | key: torch.Tensor | None = None, |
| 534 | offsets: torch.Tensor | None = None, |
| 535 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 536 | return self.forward_cuda(positions, query, key, offsets) |
| 537 | |
| 538 | @staticmethod |
| 539 | def get_next_input_positions( |
| 540 | mrope_position_delta: int, |
| 541 | context_len: int, |
| 542 | seq_len: int, |
| 543 | ) -> list[list[int]]: |
| 544 | return [ |
| 545 | list( |
| 546 | range( |
| 547 | context_len + mrope_position_delta, seq_len + mrope_position_delta |
| 548 | ) |
| 549 | ) |
| 550 | for _ in range(3) |
| 551 | ] |
| 552 | |
| 553 | @staticmethod |
| 554 | def get_next_input_positions_tensor( |
| 555 | out: np.ndarray, |
| 556 | out_offset: int, |
| 557 | mrope_position_delta: int, |
| 558 | context_len: int, |
| 559 | num_new_tokens: int, |
| 560 | ): |
| 561 | values = np.arange( |
| 562 | mrope_position_delta + context_len, |
| 563 | mrope_position_delta + context_len + num_new_tokens, |
| 564 | dtype=out.dtype, |
| 565 | ) |
| 566 | out[:, out_offset : out_offset + num_new_tokens] = values |