| 1 | #include "cpu_types.hpp" |
| 2 | |
| 3 | namespace { |
| 4 | template <typename scalar_t> |
| 5 | void rotary_embedding_impl( |
| 6 | const int64_t* __restrict__ positions, // [batch_size, seq_len] or |
| 7 | // [num_tokens] |
| 8 | scalar_t* __restrict__ query, /// [batch_size, seq_len, num_heads, |
| 9 | /// head_size] or [num_tokens, num_heads, |
| 10 | /// head_size] |
| 11 | scalar_t* __restrict__ key, // nullptr (optional) or |
| 12 | // [batch_size, seq_len, num_kv_heads, |
| 13 | // head_size] or [num_tokens, num_kv_heads, |
| 14 | // head_size] |
| 15 | const scalar_t* __restrict__ cos_sin_cache, // [max_position, 2, rot_dim // |
| 16 | // 2] |
| 17 | const int rot_dim, const int64_t query_stride, const int64_t key_stride, |
| 18 | const int num_heads, const int num_kv_heads, const int head_size, |
| 19 | const int num_tokens) { |
| 20 | using scalar_vec_t = vec_op::vec_t<scalar_t>; |
| 21 | constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num(); |
| 22 | |
| 23 | const int embed_dim = rot_dim / 2; |
| 24 | bool flag = (embed_dim % VEC_ELEM_NUM == 0); |
| 25 | const int loop_upper = flag ? embed_dim : embed_dim - VEC_ELEM_NUM; |
| 26 | |
| 27 | auto compute_loop = [&](const int64_t token_head, const scalar_t* cache_ptr, |
| 28 | scalar_t* qk) { |
| 29 | int j = 0; |
| 30 | for (; j < loop_upper; j += VEC_ELEM_NUM) { |
| 31 | const int rot_offset = j; |
| 32 | const int x_index = rot_offset; |
| 33 | const int y_index = embed_dim + rot_offset; |
| 34 | |
| 35 | const int64_t out_x = token_head + x_index; |
| 36 | const int64_t out_y = token_head + y_index; |
| 37 | |
| 38 | const scalar_vec_t cos(cache_ptr + x_index); |
| 39 | const scalar_vec_t sin(cache_ptr + y_index); |
| 40 | |
| 41 | const scalar_vec_t q_x(qk + out_x); |
| 42 | const scalar_vec_t q_y(qk + out_y); |
| 43 | |
| 44 | vec_op::FP32Vec8 fp32_cos(cos); |
| 45 | vec_op::FP32Vec8 fp32_sin(sin); |
| 46 | |
| 47 | vec_op::FP32Vec8 fp32_q_x(q_x); |
| 48 | vec_op::FP32Vec8 fp32_q_y(q_y); |
| 49 | |
| 50 | auto out1 = fp32_q_x * fp32_cos - fp32_q_y * fp32_sin; |
| 51 | scalar_vec_t(out1).save(qk + out_x); |
| 52 | |
| 53 | auto out2 = fp32_q_y * fp32_cos + fp32_q_x * fp32_sin; |
| 54 | scalar_vec_t(out2).save(qk + out_y); |
| 55 | } |
| 56 | if (!flag) { |
| 57 | for (; j < embed_dim; ++j) { |
| 58 | const int x_index = j; |
| 59 | const int y_index = embed_dim + j; |
| 60 | |
| 61 | const int64_t out_x = token_head + x_index; |
| 62 | const int64_t out_y = token_head + y_index; |
| 63 | |
| 64 | const float fp32_cos = cache_ptr[x_index]; |
| 65 | const float fp32_sin = cache_ptr[y_index]; |
| 66 | |
| 67 | const float fp32_q_x = qk[out_x]; |
| 68 | const float fp32_q_y = qk[out_y]; |
| 69 | |
| 70 | qk[out_x] = fp32_q_x * fp32_cos - fp32_q_y * fp32_sin; |
| 71 | qk[out_y] = fp32_q_y * fp32_cos + fp32_q_x * fp32_sin; |
| 72 | } |
| 73 | } |
| 74 | }; |
| 75 | |
| 76 | #pragma omp parallel for |
| 77 | for (int token_idx = 0; token_idx < num_tokens; ++token_idx) { |
| 78 | int64_t pos = positions[token_idx]; |
| 79 | const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim; |
| 80 | |
| 81 | for (int i = 0; i < num_heads; ++i) { |
| 82 | const int head_idx = i; |
| 83 | const int64_t token_head = |
| 84 | token_idx * query_stride + head_idx * head_size; |
| 85 | compute_loop(token_head, cache_ptr, query); |
| 86 | } |
| 87 | |
| 88 | if (key != nullptr) { |
| 89 | for (int i = 0; i < num_kv_heads; ++i) { |
| 90 | const int head_idx = i; |
| 91 | const int64_t token_head = |
| 92 | token_idx * key_stride + head_idx * head_size; |
| 93 | compute_loop(token_head, cache_ptr, key); |
| 94 | } |
| 95 | } |
| 96 | } |
| 97 | } |
| 98 | |
| 99 | template <> |
| 100 | void rotary_embedding_impl<c10::Half>( |
| 101 | const int64_t* __restrict__ positions, c10::Half* __restrict__ query, |
| 102 | c10::Half* __restrict__ key, const c10::Half* __restrict__ cos_sin_cache, |
| 103 | const int rot_dim, const int64_t query_stride, const int64_t key_stride, |
| 104 | const int num_heads, const int num_kv_heads, const int head_size, |
| 105 | const int num_tokens) { |
| 106 | using scalar_vec_t = vec_op::FP16Vec8; |
| 107 | constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num(); |
| 108 | |
| 109 | const int embed_dim = rot_dim / 2; |
| 110 | bool flag = (embed_dim % VEC_ELEM_NUM == 0); |
| 111 | const int loop_upper = flag ? embed_dim : embed_dim - VEC_ELEM_NUM; |
| 112 | |
| 113 | auto compute_loop = [&](const int64_t token_head, const c10::Half* cache_ptr, |
| 114 | c10::Half* qk) { |
| 115 | int j = 0; |
| 116 | for (; j < loop_upper; j += VEC_ELEM_NUM) { |
| 117 | const int rot_offset = j; |
| 118 | const int x_index = rot_offset; |
| 119 | const int y_index = embed_dim + rot_offset; |
| 120 | |
| 121 | const int64_t out_x = token_head + x_index; |
| 122 | const int64_t out_y = token_head + y_index; |
| 123 | |
| 124 | const vec_op::FP16Vec8 cos_fp16(cache_ptr + x_index); |
| 125 | const vec_op::FP16Vec8 sin_fp16(cache_ptr + y_index); |
| 126 | const vec_op::FP16Vec8 q_x_fp16(qk + out_x); |
| 127 | const vec_op::FP16Vec8 q_y_fp16(qk + out_y); |
| 128 | |
| 129 | const vec_op::FP32Vec8 fp32_cos(cos_fp16); |
| 130 | const vec_op::FP32Vec8 fp32_sin(sin_fp16); |
| 131 | const vec_op::FP32Vec8 fp32_q_x(q_x_fp16); |
| 132 | const vec_op::FP32Vec8 fp32_q_y(q_y_fp16); |
| 133 | |
| 134 | auto out1 = fp32_q_x * fp32_cos - fp32_q_y * fp32_sin; |
| 135 | auto out2 = fp32_q_y * fp32_cos + fp32_q_x * fp32_sin; |
| 136 | |
| 137 | vec_op::FP16Vec8(out1).save(qk + out_x); |
| 138 | vec_op::FP16Vec8(out2).save(qk + out_y); |
| 139 | } |
| 140 | if (!flag) { |
| 141 | for (; j < embed_dim; ++j) { |
| 142 | const int x_index = j; |
| 143 | const int y_index = embed_dim + j; |
| 144 | |
| 145 | const int64_t out_x = token_head + x_index; |
| 146 | const int64_t out_y = token_head + y_index; |
| 147 | |
| 148 | const float fp32_cos = static_cast<float>(cache_ptr[x_index]); |
| 149 | const float fp32_sin = static_cast<float>(cache_ptr[y_index]); |
| 150 | const float fp32_q_x = static_cast<float>(qk[out_x]); |
| 151 | const float fp32_q_y = static_cast<float>(qk[out_y]); |
| 152 | |
| 153 | qk[out_x] = |
| 154 | static_cast<c10::Half>(fp32_q_x * fp32_cos - fp32_q_y * fp32_sin); |
| 155 | qk[out_y] = |
| 156 | static_cast<c10::Half>(fp32_q_y * fp32_cos + fp32_q_x * fp32_sin); |
| 157 | } |
| 158 | } |
| 159 | }; |
| 160 | |
| 161 | #pragma omp parallel for |
| 162 | for (int token_idx = 0; token_idx < num_tokens; ++token_idx) { |
| 163 | int64_t pos = positions[token_idx]; |
| 164 | const c10::Half* cache_ptr = cos_sin_cache + pos * rot_dim; |
| 165 | |
| 166 | for (int i = 0; i < num_heads; ++i) { |
| 167 | const int head_idx = i; |
| 168 | const int64_t token_head = |
| 169 | token_idx * query_stride + head_idx * head_size; |
| 170 | compute_loop(token_head, cache_ptr, query); |
| 171 | } |
| 172 | |
| 173 | if (key != nullptr) { |
| 174 | for (int i = 0; i < num_kv_heads; ++i) { |
| 175 | const int head_idx = i; |
| 176 | const int64_t token_head = |
| 177 | token_idx * key_stride + head_idx * head_size; |
| 178 | compute_loop(token_head, cache_ptr, key); |
| 179 | } |
| 180 | } |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | template <typename scalar_t> |
| 185 | void rotary_embedding_gptj_impl( |
| 186 | const int64_t* __restrict__ positions, // [batch_size, seq_len] or |
| 187 | // [num_tokens] |
| 188 | scalar_t* __restrict__ query, /// [batch_size, seq_len, num_heads, |
| 189 | /// head_size] or [num_tokens, num_heads, |
| 190 | /// head_size] |
| 191 | scalar_t* __restrict__ key, // nullptr (optional) or |
| 192 | // [batch_size, seq_len, num_kv_heads, |
| 193 | // head_size] or [num_tokens, num_kv_heads, |
| 194 | // head_size] |
| 195 | const scalar_t* __restrict__ cos_sin_cache, // [max_position, 2, rot_dim // |
| 196 | // 2] |
| 197 | const int rot_dim, const int64_t query_stride, const int64_t key_stride, |
| 198 | const int num_heads, const int num_kv_heads, const int head_size, |
| 199 | const int num_tokens) { |
| 200 | const int embed_dim = rot_dim / 2; |
| 201 | |
| 202 | #pragma omp parallel for collapse(2) |
| 203 | for (int token_idx = 0; token_idx < num_tokens; ++token_idx) { |
| 204 | for (int i = 0; i < num_heads; ++i) { |
| 205 | int64_t pos = positions[token_idx]; |
| 206 | const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim; |
| 207 | const scalar_t* cos_cache_ptr = cache_ptr; |
| 208 | const scalar_t* sin_cache_ptr = cache_ptr + embed_dim; |
| 209 | const int head_idx = i; |
| 210 | const int64_t token_head = |
| 211 | token_idx * query_stride + head_idx * head_size; |
| 212 | scalar_t* head_query = token_head + query; |
| 213 | for (int j = 0; j < embed_dim; j += 1) { |
| 214 | const int rot_offset = j; |
| 215 | const int x_index = 2 * rot_offset; |
| 216 | const int y_index = 2 * rot_offset + 1; |
| 217 | |
| 218 | const float cos = cos_cache_ptr[rot_offset]; |
| 219 | const float sin = sin_cache_ptr[rot_offset]; |
| 220 | |
| 221 | const float x = head_query[x_index]; |
| 222 | const float y = head_query[y_index]; |
| 223 | |
| 224 | head_query[x_index] = x * cos - y * sin; |
| 225 | head_query[y_index] = y * cos + x * sin; |
| 226 | } |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | if (key == nullptr) { |
| 231 | return; |
| 232 | } |
| 233 | |
| 234 | #pragma omp parallel for collapse(2) |
| 235 | for (int token_idx = 0; token_idx < num_tokens; ++token_idx) { |
| 236 | for (int i = 0; i < num_kv_heads; ++i) { |
| 237 | int64_t pos = positions[token_idx]; |
| 238 | const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim; |
| 239 | const scalar_t* cos_cache_ptr = cache_ptr; |
| 240 | const scalar_t* sin_cache_ptr = cache_ptr + embed_dim; |
| 241 | const int head_idx = i; |
| 242 | const int64_t token_head = token_idx * key_stride + head_idx * head_size; |
| 243 | scalar_t* head_key = key + token_head; |
| 244 | for (int j = 0; j < embed_dim; j += 1) { |
| 245 | const int rot_offset = j; |
| 246 | const int x_index = 2 * rot_offset; |
| 247 | const int y_index = 2 * rot_offset + 1; |
| 248 | |
| 249 | const float cos = cos_cache_ptr[rot_offset]; |
| 250 | const float sin = sin_cache_ptr[rot_offset]; |
| 251 | |
| 252 | const float x = head_key[x_index]; |
| 253 | const float y = head_key[y_index]; |
| 254 | |
| 255 | head_key[x_index] = x * cos - y * sin; |
| 256 | head_key[y_index] = y * cos + x * sin; |
| 257 | } |
| 258 | } |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | template <> |
| 263 | void rotary_embedding_gptj_impl<c10::Half>( |
| 264 | const int64_t* __restrict__ positions, c10::Half* __restrict__ query, |
| 265 | c10::Half* __restrict__ key, const c10::Half* __restrict__ cos_sin_cache, |
| 266 | const int rot_dim, const int64_t query_stride, const int64_t key_stride, |
| 267 | const int num_heads, const int num_kv_heads, const int head_size, |
| 268 | const int num_tokens) { |
| 269 | const int embed_dim = rot_dim / 2; |
| 270 | |
| 271 | #pragma omp parallel for collapse(2) |
| 272 | for (int token_idx = 0; token_idx < num_tokens; ++token_idx) { |
| 273 | for (int i = 0; i < num_heads; ++i) { |
| 274 | int64_t pos = positions[token_idx]; |
| 275 | const c10::Half* cache_ptr = cos_sin_cache + pos * rot_dim; |
| 276 | const c10::Half* cos_cache_ptr = cache_ptr; |
| 277 | const c10::Half* sin_cache_ptr = cache_ptr + embed_dim; |
| 278 | const int head_idx = i; |
| 279 | const int64_t token_head = |
| 280 | token_idx * query_stride + head_idx * head_size; |
| 281 | c10::Half* head_query = token_head + query; |
| 282 | for (int j = 0; j < embed_dim; j += 1) { |
| 283 | const int rot_offset = j; |
| 284 | const int x_index = 2 * rot_offset; |
| 285 | const int y_index = 2 * rot_offset + 1; |
| 286 | |
| 287 | const float cos = static_cast<float>(cos_cache_ptr[rot_offset]); |
| 288 | const float sin = static_cast<float>(sin_cache_ptr[rot_offset]); |
| 289 | |
| 290 | const float x = static_cast<float>(head_query[x_index]); |
| 291 | const float y = static_cast<float>(head_query[y_index]); |
| 292 | |
| 293 | head_query[x_index] = static_cast<c10::Half>(x * cos - y * sin); |
| 294 | head_query[y_index] = static_cast<c10::Half>(y * cos + x * sin); |
| 295 | } |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | if (key == nullptr) { |
| 300 | return; |
| 301 | } |
| 302 | |
| 303 | #pragma omp parallel for collapse(2) |
| 304 | for (int token_idx = 0; token_idx < num_tokens; ++token_idx) { |
| 305 | for (int i = 0; i < num_kv_heads; ++i) { |
| 306 | int64_t pos = positions[token_idx]; |
| 307 | const c10::Half* cache_ptr = cos_sin_cache + pos * rot_dim; |
| 308 | const c10::Half* cos_cache_ptr = cache_ptr; |
| 309 | const c10::Half* sin_cache_ptr = cache_ptr + embed_dim; |
| 310 | const int head_idx = i; |
| 311 | const int64_t token_head = token_idx * key_stride + head_idx * head_size; |
| 312 | c10::Half* head_key = key + token_head; |
| 313 | for (int j = 0; j < embed_dim; j += 1) { |
| 314 | const int rot_offset = j; |
| 315 | const int x_index = 2 * rot_offset; |
| 316 | const int y_index = 2 * rot_offset + 1; |
| 317 | |
| 318 | const float cos = static_cast<float>(cos_cache_ptr[rot_offset]); |
| 319 | const float sin = static_cast<float>(sin_cache_ptr[rot_offset]); |
| 320 | |
| 321 | const float x = static_cast<float>(head_key[x_index]); |
| 322 | const float y = static_cast<float>(head_key[y_index]); |
| 323 | |
| 324 | head_key[x_index] = static_cast<c10::Half>(x * cos - y * sin); |
| 325 | head_key[y_index] = static_cast<c10::Half>(y * cos + x * sin); |
| 326 | } |
| 327 | } |
| 328 | } |
| 329 | } |
| 330 | }; // namespace |
| 331 | |
| 332 | void rotary_embedding(torch::Tensor& positions, torch::Tensor& query, |
| 333 | std::optional<torch::Tensor> key, int64_t head_size, |
| 334 | torch::Tensor& cos_sin_cache, bool is_neox, |
| 335 | int64_t rope_dim_offset, bool inverse) { |
| 336 | TORCH_CHECK(rope_dim_offset == 0, |
| 337 | "rope_dim_offset != 0 is not supported on CPU"); |
| 338 | TORCH_CHECK(!inverse, "inverse rotary embedding is not supported on CPU"); |
| 339 | |
| 340 | int num_tokens = positions.numel(); |
| 341 | int rot_dim = cos_sin_cache.size(1); |
| 342 | int num_heads = query.size(-1) / head_size; |
| 343 | int num_kv_heads = key.has_value() ? key->size(-1) / head_size : num_heads; |
| 344 | int64_t key_stride = key.has_value() ? key->stride(-2) : 0; |
| 345 | int64_t query_stride = query.stride(-2); |
| 346 | |
| 347 | VLLM_DISPATCH_FLOATING_TYPES( |
| 348 | query.scalar_type(), "rotary_embedding_impl", [&] { |
| 349 | CPU_KERNEL_GUARD_IN(rotary_embedding_impl) |
| 350 | if (is_neox) { |
| 351 | rotary_embedding_impl( |
| 352 | positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(), |
| 353 | key.has_value() ? key->data_ptr<scalar_t>() : nullptr, |
| 354 | cos_sin_cache.data_ptr<scalar_t>(), rot_dim, query_stride, |
| 355 | key_stride, num_heads, num_kv_heads, head_size, num_tokens); |
| 356 | } else { |
| 357 | rotary_embedding_gptj_impl( |
| 358 | positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(), |
| 359 | key.has_value() ? key->data_ptr<scalar_t>() : nullptr, |
| 360 | cos_sin_cache.data_ptr<scalar_t>(), rot_dim, query_stride, |
| 361 | key_stride, num_heads, num_kv_heads, head_size, num_tokens); |
| 362 | } |
| 363 | |
| 364 | CPU_KERNEL_GUARD_OUT(rotary_embedding_impl) |
| 365 | }); |
| 366 | } |