| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | |
| 5 | import torch |
| 6 | |
| 7 | from vllm.platforms import current_platform |
| 8 | |
| 9 | if current_platform.is_cuda_alike(): |
| 10 | from vllm import _custom_ops as ops |
| 11 | elif current_platform.is_xpu(): |
| 12 | from vllm._xpu_ops import xpu_ops as ops # type: ignore[no-redef] |
| 13 | |
| 14 | |
| 15 | class PagedAttention: |
| 16 | @staticmethod |
| 17 | def split_kv_cache( |
| 18 | kv_cache: torch.Tensor, |
| 19 | num_kv_heads: int, |
| 20 | head_size: int, |
| 21 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 22 | x = 16 // kv_cache.element_size() |
| 23 | num_blocks = kv_cache.shape[1] |
| 24 | |
| 25 | key_cache = kv_cache[0] |
| 26 | key_cache = key_cache.view(num_blocks, num_kv_heads, head_size // x, -1, x) |
| 27 | value_cache = kv_cache[1] |
| 28 | value_cache = value_cache.view(num_blocks, num_kv_heads, head_size, -1) |
| 29 | return key_cache, value_cache |
| 30 | |
| 31 | @staticmethod |
| 32 | def write_to_paged_cache( |
| 33 | key: torch.Tensor, |
| 34 | value: torch.Tensor, |
| 35 | key_cache: torch.Tensor, |
| 36 | value_cache: torch.Tensor, |
| 37 | slot_mapping: torch.Tensor, |
| 38 | kv_cache_dtype: str, |
| 39 | k_scale: torch.Tensor, |
| 40 | v_scale: torch.Tensor, |
| 41 | ) -> None: |
| 42 | ops.reshape_and_cache( |
| 43 | key, |
| 44 | value, |
| 45 | key_cache, |
| 46 | value_cache, |
| 47 | slot_mapping.flatten(), |
| 48 | kv_cache_dtype, |
| 49 | k_scale, |
| 50 | v_scale, |
| 51 | ) |