| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import numpy as np |
| 4 | |
| 5 | from vllm.v1.kv_offload.base import BlockIDsLoadStoreSpec |
| 6 | |
| 7 | |
| 8 | class CPUOffloadingMetrics: |
| 9 | STORES_SKIPPED = "vllm:kv_offload_stores_skipped" |
| 10 | CPU_CACHE_USAGE_PERC = "vllm:kv_offload_cpu_cache_usage_perc" |
| 11 | CPU_ALLOCATION_SIZE = "vllm:kv_offload_cpu_allocation_size" |
| 12 | CPU_CACHE_WRITE_USAGE_PERC = "vllm:kv_offload_cpu_cache_write_usage_perc" |
| 13 | CPU_CACHE_READ_USAGE_PERC = "vllm:kv_offload_cpu_cache_read_usage_perc" |
| 14 | |
| 15 | |
| 16 | class CPULoadStoreSpec(BlockIDsLoadStoreSpec): |
| 17 | """Spec for loading/storing KV chunks to/from CPU memory. |
| 18 | |
| 19 | The inherited block_ids field holds chunk indices into the |
| 20 | CPU cache (not GPU block IDs). The chunk_ids alias exposes |
| 21 | the same array under the name used by the tiering layer. |
| 22 | """ |
| 23 | |
| 24 | @property |
| 25 | def chunk_ids(self) -> np.ndarray: |
| 26 | return self.block_ids |