| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | |
| 6 | from vllm.triton_utils import tl, triton |
| 7 | |
| 8 | from .mrope import MRotaryEmbedding |
| 9 | |
| 10 | |
| 11 | def _validate_bailing_mrope_config( |
| 12 | head_size: int, |
| 13 | rotary_dim: int, |
| 14 | mrope_section: list[int], |
| 15 | ) -> int: |
| 16 | if rotary_dim <= 0 or rotary_dim % 2 != 0 or rotary_dim > head_size: |
| 17 | raise ValueError( |
| 18 | "Bailing M-RoPE requires an even rotary_dim in (0, head_size], " |
| 19 | f"got rotary_dim={rotary_dim} and head_size={head_size}" |
| 20 | ) |
| 21 | if len(mrope_section) != 3: |
| 22 | raise ValueError( |
| 23 | "Bailing M-RoPE expects [T, H, W] sections, " |
| 24 | f"got mrope_section={mrope_section}" |
| 25 | ) |
| 26 | |
| 27 | temporal_size, height_size, width_size = mrope_section |
| 28 | if any(section_size < 0 for section_size in mrope_section): |
| 29 | raise ValueError( |
| 30 | "Bailing M-RoPE sections must be non-negative, " |
| 31 | f"got mrope_section={mrope_section}" |
| 32 | ) |
| 33 | if height_size != width_size: |
| 34 | raise ValueError( |
| 35 | "Bailing M-RoPE requires equal H and W sections, " |
| 36 | f"got height_size={height_size} and width_size={width_size}" |
| 37 | ) |
| 38 | spatial_size = height_size + width_size |
| 39 | if spatial_size + temporal_size != rotary_dim // 2: |
| 40 | raise ValueError( |
| 41 | "Bailing M-RoPE sections must sum to rotary_dim // 2, got " |
| 42 | f"mrope_section={mrope_section} and rotary_dim={rotary_dim}" |
| 43 | ) |
| 44 | return spatial_size |
| 45 | |
| 46 | |
| 47 | def _validate_bailing_positions( |
| 48 | positions: torch.Tensor, |
| 49 | *, |
| 50 | require_2d: bool = False, |
| 51 | ) -> None: |
| 52 | if positions.dtype not in (torch.int32, torch.int64): |
| 53 | raise ValueError( |
| 54 | "Bailing M-RoPE positions must have int32 or int64 dtype, got " |
| 55 | f"dtype={positions.dtype}" |
| 56 | ) |
| 57 | if positions.ndim == 1 and not require_2d: |
| 58 | return |
| 59 | if positions.ndim == 2: |
| 60 | if positions.shape[0] == 3: |
| 61 | return |
| 62 | raise ValueError( |
| 63 | "Bailing M-RoPE expects T/H/W positions with shape [3, N], " |
| 64 | f"got shape={tuple(positions.shape)}" |
| 65 | ) |
| 66 | if require_2d: |
| 67 | raise ValueError( |
| 68 | "Bailing M-RoPE expects T/H/W positions with shape [3, N], " |
| 69 | f"got shape={tuple(positions.shape)}" |
| 70 | ) |
| 71 | raise ValueError( |
| 72 | f"Bailing M-RoPE positions must be 1D or 2D, got shape={tuple(positions.shape)}" |
| 73 | ) |
| 74 | |
| 75 | |
| 76 | @triton.jit |
| 77 | def _triton_bailing_mrope_forward( |
| 78 | q_ptr, |
| 79 | k_ptr, |
| 80 | positions_ptr, |
| 81 | cos_sin_cache_ptr, |
| 82 | q_stride_token, |
| 83 | q_stride_head, |
| 84 | q_stride_dim, |
| 85 | k_stride_token, |
| 86 | k_stride_head, |
| 87 | k_stride_dim, |
| 88 | positions_stride_axis, |
| 89 | positions_stride_token, |
| 90 | cache_stride_position, |
| 91 | cache_stride_dim, |
| 92 | n_qh: tl.constexpr, |
| 93 | n_kh: tl.constexpr, |
| 94 | rotary_dim: tl.constexpr, |
| 95 | pad_n_qh: tl.constexpr, |
| 96 | pad_n_kh: tl.constexpr, |
| 97 | pad_half_rotary_dim: tl.constexpr, |
| 98 | spatial_size: tl.constexpr, |
| 99 | ): |
| 100 | """Apply GPT-J-style H/W-interleaved, T-tail M-RoPE in place.""" |
| 101 | token_idx = tl.program_id(0) |
| 102 | position_offset = token_idx * positions_stride_token |
| 103 | temporal_position = tl.load(positions_ptr + position_offset).to(tl.int64) |
| 104 | height_position = tl.load( |
| 105 | positions_ptr + positions_stride_axis + position_offset |
| 106 | ).to(tl.int64) |
| 107 | width_position = tl.load( |
| 108 | positions_ptr + 2 * positions_stride_axis + position_offset |
| 109 | ).to(tl.int64) |
| 110 | |
| 111 | frequency_offsets = tl.arange(0, pad_half_rotary_dim) |
| 112 | frequency_mask = frequency_offsets < rotary_dim // 2 |
| 113 | spatial_position = tl.where( |
| 114 | frequency_offsets % 2 == 0, |
| 115 | height_position, |
| 116 | width_position, |
| 117 | ) |
| 118 | selected_position = tl.where( |
| 119 | frequency_offsets < spatial_size, |
| 120 | spatial_position, |
| 121 | temporal_position, |
| 122 | ) |
| 123 | cache_offsets = ( |
| 124 | selected_position * cache_stride_position + frequency_offsets * cache_stride_dim |
| 125 | ) |
| 126 | cos_row = tl.load( |
| 127 | cos_sin_cache_ptr + cache_offsets, |
| 128 | mask=frequency_mask, |
| 129 | other=0.0, |
| 130 | ).to(tl.float32) |
| 131 | sin_row = tl.load( |
| 132 | cos_sin_cache_ptr + cache_offsets + (rotary_dim // 2) * cache_stride_dim, |
| 133 | mask=frequency_mask, |
| 134 | other=0.0, |
| 135 | ).to(tl.float32) |
| 136 | |
| 137 | q_head_offsets = tl.arange(0, pad_n_qh)[:, None] |
| 138 | k_head_offsets = tl.arange(0, pad_n_kh)[:, None] |
| 139 | pair_offsets = frequency_offsets[None, :] |
| 140 | q_mask = (q_head_offsets < n_qh) & frequency_mask[None, :] |
| 141 | k_mask = (k_head_offsets < n_kh) & frequency_mask[None, :] |
| 142 | |
| 143 | q_base = q_ptr + token_idx * q_stride_token |
| 144 | q_even_offsets = q_head_offsets * q_stride_head + 2 * pair_offsets * q_stride_dim |
| 145 | q_odd_offsets = q_even_offsets + q_stride_dim |
| 146 | q_even = tl.load(q_base + q_even_offsets, mask=q_mask, other=0.0).to(tl.float32) |
| 147 | q_odd = tl.load(q_base + q_odd_offsets, mask=q_mask, other=0.0).to(tl.float32) |
| 148 | q_even_cos = (q_even * cos_row[None, :]).to(q_ptr.type.element_ty).to(tl.float32) |
| 149 | q_odd_sin = (q_odd * sin_row[None, :]).to(q_ptr.type.element_ty).to(tl.float32) |
| 150 | q_odd_cos = (q_odd * cos_row[None, :]).to(q_ptr.type.element_ty).to(tl.float32) |
| 151 | q_even_sin = (q_even * sin_row[None, :]).to(q_ptr.type.element_ty).to(tl.float32) |
| 152 | tl.store( |
| 153 | q_base + q_even_offsets, |
| 154 | q_even_cos - q_odd_sin, |
| 155 | mask=q_mask, |
| 156 | ) |
| 157 | tl.store( |
| 158 | q_base + q_odd_offsets, |
| 159 | q_odd_cos + q_even_sin, |
| 160 | mask=q_mask, |
| 161 | ) |
| 162 | |
| 163 | k_base = k_ptr + token_idx * k_stride_token |
| 164 | k_even_offsets = k_head_offsets * k_stride_head + 2 * pair_offsets * k_stride_dim |
| 165 | k_odd_offsets = k_even_offsets + k_stride_dim |
| 166 | k_even = tl.load(k_base + k_even_offsets, mask=k_mask, other=0.0).to(tl.float32) |
| 167 | k_odd = tl.load(k_base + k_odd_offsets, mask=k_mask, other=0.0).to(tl.float32) |
| 168 | k_even_cos = (k_even * cos_row[None, :]).to(k_ptr.type.element_ty).to(tl.float32) |
| 169 | k_odd_sin = (k_odd * sin_row[None, :]).to(k_ptr.type.element_ty).to(tl.float32) |
| 170 | k_odd_cos = (k_odd * cos_row[None, :]).to(k_ptr.type.element_ty).to(tl.float32) |
| 171 | k_even_sin = (k_even * sin_row[None, :]).to(k_ptr.type.element_ty).to(tl.float32) |
| 172 | tl.store( |
| 173 | k_base + k_even_offsets, |
| 174 | k_even_cos - k_odd_sin, |
| 175 | mask=k_mask, |
| 176 | ) |
| 177 | tl.store( |
| 178 | k_base + k_odd_offsets, |
| 179 | k_odd_cos + k_even_sin, |
| 180 | mask=k_mask, |
| 181 | ) |
| 182 | |
| 183 | |
| 184 | def triton_bailing_mrope( |
| 185 | query: torch.Tensor, |
| 186 | key: torch.Tensor, |
| 187 | positions: torch.Tensor, |
| 188 | cos_sin_cache: torch.Tensor, |
| 189 | spatial_size: int, |
| 190 | head_size: int, |
| 191 | rotary_dim: int, |
| 192 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 193 | """Fused Bailing M-RoPE for packed or strided Q/K.""" |
| 194 | _validate_bailing_positions(positions, require_2d=True) |
| 195 | num_tokens = positions.shape[-1] |
| 196 | if num_tokens == 0: |
| 197 | return query, key |
| 198 | |
| 199 | query_3d = query.view(num_tokens, -1, head_size) |
| 200 | key_3d = key.view(num_tokens, -1, head_size) |
| 201 | n_q_heads = query_3d.shape[1] |
| 202 | n_k_heads = key_3d.shape[1] |
| 203 | pad_n_q_heads = triton.next_power_of_2(n_q_heads) |
| 204 | pad_n_k_heads = triton.next_power_of_2(n_k_heads) |
| 205 | pad_half_rotary_dim = triton.next_power_of_2(rotary_dim // 2) |
| 206 | _triton_bailing_mrope_forward[(num_tokens,)]( |
| 207 | query_3d, |
| 208 | key_3d, |
| 209 | positions, |
| 210 | cos_sin_cache, |
| 211 | query_3d.stride(0), |
| 212 | query_3d.stride(1), |
| 213 | query_3d.stride(2), |
| 214 | key_3d.stride(0), |
| 215 | key_3d.stride(1), |
| 216 | key_3d.stride(2), |
| 217 | positions.stride(0), |
| 218 | positions.stride(1), |
| 219 | cos_sin_cache.stride(0), |
| 220 | cos_sin_cache.stride(1), |
| 221 | n_q_heads, |
| 222 | n_k_heads, |
| 223 | rotary_dim, |
| 224 | pad_n_q_heads, |
| 225 | pad_n_k_heads, |
| 226 | pad_half_rotary_dim, |
| 227 | spatial_size, |
| 228 | num_warps=4, |
| 229 | ) |
| 230 | return query, key |
| 231 | |
| 232 | |
| 233 | class BailingMRotaryEmbedding(MRotaryEmbedding): |
| 234 | """Bailing M-RoPE with alternating H/W frequencies and a temporal tail.""" |
| 235 | |
| 236 | def __init__( |
| 237 | self, |
| 238 | head_size: int, |
| 239 | rotary_dim: int, |
| 240 | max_position_embeddings: int, |
| 241 | base: float, |
| 242 | is_neox_style: bool, |
| 243 | dtype: torch.dtype, |
| 244 | mrope_section: list[int], |
| 245 | ) -> None: |
| 246 | self.spatial_size = _validate_bailing_mrope_config( |
| 247 | head_size, rotary_dim, mrope_section |
| 248 | ) |
| 249 | super().__init__( |
| 250 | head_size, |
| 251 | rotary_dim, |
| 252 | max_position_embeddings, |
| 253 | base, |
| 254 | is_neox_style, |
| 255 | dtype, |
| 256 | mrope_section=mrope_section, |
| 257 | ) |
| 258 | |
| 259 | def select_cos_sin( |
| 260 | self, |
| 261 | positions: torch.Tensor, |
| 262 | cos_sin_cache: torch.Tensor | None = None, |
| 263 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 264 | _validate_bailing_positions(positions) |
| 265 | if cos_sin_cache is None: |
| 266 | cos_sin_cache = self.cos_sin_cache |
| 267 | cos_sin = cos_sin_cache[positions.to(torch.long)] |
| 268 | cos, sin = cos_sin.chunk(2, dim=-1) |
| 269 | if positions.ndim == 1: |
| 270 | return cos, sin |
| 271 | |
| 272 | spatial_mask = torch.arange(self.spatial_size, device=cos.device) % 2 == 0 |
| 273 | spatial_mask = spatial_mask.unsqueeze(0) |
| 274 | spatial_cos = torch.where( |
| 275 | spatial_mask, |
| 276 | cos[1, :, : self.spatial_size], |
| 277 | cos[2, :, : self.spatial_size], |
| 278 | ) |
| 279 | spatial_sin = torch.where( |
| 280 | spatial_mask, |
| 281 | sin[1, :, : self.spatial_size], |
| 282 | sin[2, :, : self.spatial_size], |
| 283 | ) |
| 284 | temporal_slice = slice(self.spatial_size, self.rotary_dim // 2) |
| 285 | return ( |
| 286 | torch.cat((spatial_cos, cos[0, :, temporal_slice]), dim=-1), |
| 287 | torch.cat((spatial_sin, sin[0, :, temporal_slice]), dim=-1), |
| 288 | ) |
| 289 | |
| 290 | def forward_native( |
| 291 | self, |
| 292 | positions: torch.Tensor, |
| 293 | query: torch.Tensor, |
| 294 | key: torch.Tensor | None = None, |
| 295 | offsets: torch.Tensor | None = None, |
| 296 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 297 | if key is None: |
| 298 | raise ValueError("Bailing M-RoPE requires key states") |
| 299 | |
| 300 | cos_sin_cache = self._match_cos_sin_cache_dtype(query) |
| 301 | cos, sin = self.select_cos_sin(positions, cos_sin_cache) |
| 302 | num_tokens = positions.shape[-1] |
| 303 | query_shape = query.shape |
| 304 | query = query.view(num_tokens, -1, self.head_size) |
| 305 | query_rot = self.apply_rotary_emb.forward_native( |
| 306 | query[..., : self.rotary_dim], |
| 307 | cos, |
| 308 | sin, |
| 309 | ) |
| 310 | query = torch.cat((query_rot, query[..., self.rotary_dim :]), dim=-1) |
| 311 | |
| 312 | key_shape = key.shape |
| 313 | key = key.view(num_tokens, -1, self.head_size) |
| 314 | key_rot = self.apply_rotary_emb.forward_native( |
| 315 | key[..., : self.rotary_dim], |
| 316 | cos, |
| 317 | sin, |
| 318 | ) |
| 319 | key = torch.cat((key_rot, key[..., self.rotary_dim :]), dim=-1) |
| 320 | return query.reshape(query_shape), key.reshape(key_shape) |
| 321 | |
| 322 | def forward_cuda( |
| 323 | self, |
| 324 | positions: torch.Tensor, |
| 325 | query: torch.Tensor, |
| 326 | key: torch.Tensor | None = None, |
| 327 | offsets: torch.Tensor | None = None, |
| 328 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 329 | if positions.ndim == 2 and not self.is_neox_style: |
| 330 | if key is None: |
| 331 | raise ValueError("Bailing M-RoPE requires key states") |
| 332 | cos_sin_cache = self._match_cos_sin_cache_dtype(query) |
| 333 | return triton_bailing_mrope( |
| 334 | query, |
| 335 | key, |
| 336 | positions, |
| 337 | cos_sin_cache, |
| 338 | self.spatial_size, |
| 339 | self.head_size, |
| 340 | self.rotary_dim, |
| 341 | ) |
| 342 | return self.forward_native(positions, query, key, offsets) |
| 343 | |
| 344 | def forward_cpu( |
| 345 | self, |
| 346 | positions: torch.Tensor, |
| 347 | query: torch.Tensor, |
| 348 | key: torch.Tensor | None = None, |
| 349 | offsets: torch.Tensor | None = None, |
| 350 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 351 | return self.forward_native(positions, query, key, offsets) |