| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """DMA copy backend for GPU<->CPU block transfers.""" |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | import queue |
| 8 | import threading |
| 9 | |
| 10 | import torch |
| 11 | |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.platforms import current_platform |
| 14 | from vllm.v1.simple_kv_offload.cuda_mem_ops import ( |
| 15 | CU_MEMCPY_SRC_ACCESS_ORDER_ANY, |
| 16 | CU_MEMCPY_SRC_ACCESS_ORDER_STREAM, |
| 17 | BatchMemcpyParams, |
| 18 | build_params, |
| 19 | copy_blocks, |
| 20 | ) |
| 21 | |
| 22 | logger = init_logger(__name__) |
| 23 | |
| 24 | |
| 25 | class DmaCopyBackend: |
| 26 | """cuMemcpyBatchAsync copy backend (background thread).""" |
| 27 | |
| 28 | def __init__(self) -> None: |
| 29 | self._store_params: BatchMemcpyParams | None = None |
| 30 | self._load_params: BatchMemcpyParams | None = None |
| 31 | self._load_stream: torch.cuda.Stream | None = None |
| 32 | self._store_stream: torch.cuda.Stream | None = None |
| 33 | self._queue: queue.SimpleQueue | None = None |
| 34 | self._thread: threading.Thread | None = None |
| 35 | self._shutdown: bool = False |
| 36 | |
| 37 | def init( |
| 38 | self, |
| 39 | gpu_caches: dict[str, torch.Tensor], |
| 40 | cpu_caches: dict[str, torch.Tensor], |
| 41 | device: torch.device, |
| 42 | load_stream: torch.cuda.Stream, |
| 43 | store_stream: torch.cuda.Stream, |
| 44 | ) -> None: |
| 45 | self._load_stream = load_stream |
| 46 | self._store_stream = store_stream |
| 47 | |
| 48 | # Stores read the live KV cache -> STREAM (paired with the compute-done |
| 49 | # wait in get_finished); loads read stable pinned host memory -> ANY. |
| 50 | self._store_params = build_params( |
| 51 | gpu_caches, |
| 52 | cpu_caches, |
| 53 | store_stream, |
| 54 | src_access_order=CU_MEMCPY_SRC_ACCESS_ORDER_STREAM, |
| 55 | ) |
| 56 | self._load_params = build_params( |
| 57 | cpu_caches, |
| 58 | gpu_caches, |
| 59 | load_stream, |
| 60 | src_access_order=CU_MEMCPY_SRC_ACCESS_ORDER_ANY, |
| 61 | ) |
| 62 | |
| 63 | self._queue = queue.SimpleQueue() |
| 64 | self._thread = threading.Thread( |
| 65 | target=self._copy_loop, |
| 66 | args=(self._queue, device, load_stream, store_stream), |
| 67 | daemon=True, |
| 68 | ) |
| 69 | self._thread.start() |
| 70 | |
| 71 | def launch_copy( |
| 72 | self, |
| 73 | src_blocks: list[int], |
| 74 | dst_blocks: list[int], |
| 75 | is_store: bool, |
| 76 | event_idx: int, |
| 77 | events_list: list[tuple[int, torch.Event]], |
| 78 | wait_event: torch.Event | None = None, |
| 79 | ) -> None: |
| 80 | params = self._store_params if is_store else self._load_params |
| 81 | assert params is not None and self._queue is not None |
| 82 | self._queue.put( |
| 83 | ( |
| 84 | src_blocks, |
| 85 | dst_blocks, |
| 86 | params, |
| 87 | is_store, |
| 88 | event_idx, |
| 89 | events_list, |
| 90 | wait_event, |
| 91 | ) |
| 92 | ) |
| 93 | |
| 94 | def shutdown(self) -> None: |
| 95 | if self._shutdown: |
| 96 | return |
| 97 | self._shutdown = True |
| 98 | if self._queue is not None: |
| 99 | self._queue.put(None) |
| 100 | if self._thread is not None: |
| 101 | self._thread.join(timeout=5.0) |
| 102 | |
| 103 | @staticmethod |
| 104 | def _copy_loop( |
| 105 | q: queue.SimpleQueue, |
| 106 | device: torch.device, |
| 107 | load_stream: torch.cuda.Stream, |
| 108 | store_stream: torch.cuda.Stream, |
| 109 | ) -> None: |
| 110 | current_platform.set_device(device) |
| 111 | while True: |
| 112 | item = q.get() |
| 113 | if item is None: |
| 114 | return |
| 115 | ( |
| 116 | src_blocks, |
| 117 | dst_blocks, |
| 118 | params, |
| 119 | is_store, |
| 120 | event_idx, |
| 121 | events_list, |
| 122 | wait_event, |
| 123 | ) = item |
| 124 | stream = store_stream if is_store else load_stream |
| 125 | if wait_event is not None: |
| 126 | stream.wait_event(wait_event) |
| 127 | copy_blocks(src_blocks, dst_blocks, params) |
| 128 | event = torch.Event() |
| 129 | event.record(stream) |
| 130 | events_list.append((event_idx, event)) |