# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project from collections import deque from dataclasses import dataclass import numpy as np import torch from vllm import _custom_ops as ops from vllm.logger import init_logger from vllm.utils.platform_utils import is_pin_memory_available from vllm.v1.attention.backend import AttentionBackend from vllm.v1.kv_offload.mediums import BlockIDsLoadStoreSpec from vllm.v1.kv_offload.worker.worker import ( OffloadingHandler, TransferResult, TransferSpec, ) logger = init_logger(__name__) @dataclass class Transfer: job_id: int stream: torch.cuda.Stream start_event: torch.Event end_event: torch.Event num_bytes: int def expand_block_ids( block_ids: np.ndarray, block_size_factor: int, output: np.ndarray, skip_count: int = 0, ): """ Convert a list of block IDs to a list of matching block ids, assuming each block is composed of actual block_size_factor blocks. Outputs to output tensor. The first skip_count blocks will be skipped. Note that skip_count must be less than block_size_factor. For example, if block_ids = [0, 1, 3] and block_size_factor = 4, then it yields [0, 1, 2, 3, 4, 5, 6, 7, 12, 13, 14, 15] since 0 maps to [0, 1, 2, 3] 1 maps to [4, 5, 6, 7] and 3 maps to [12, 13, 14, 15] """ assert skip_count < block_size_factor first_range = np.arange(skip_count, block_size_factor) full_range = np.arange(0, block_size_factor) output_idx = 0 for i, block_id in enumerate(block_ids): base_block_id = block_id * block_size_factor indices = first_range if i == 0 else full_range output_end_idx = output_idx + len(indices) output[output_idx:output_end_idx] = base_block_id + indices output_idx = output_end_idx class SingleDirectionOffloadingHandler(OffloadingHandler): """ SingleDirectionOffloadingHandler handles transfers for a single direction, either CPU->GPU or GPU->CPU. Transfers are guaranteed to be executed in order of their submission. Each transfer uses a unique CUDA stream, and its stream will start executing only after the streams of previous transfers have finished. """ def __init__( self, src_tensors: list[torch.Tensor], dst_tensors: list[torch.Tensor], src_block_size_factor: int, dst_block_size_factor: int, ): """ Initialize a SingleDirectionOffloadingHandler. Args: src_tensors: list of KV cache tensors to copy from. dst_tensors: list of KV cache tensors to copy to. Order should match src_tensors. src_block_size_factor: The number of kernel blocks per KV block in a source tensor. dst_block_size_factor: The number of kernel blocks per KV block in a destination tensor. """ assert len(src_tensors) == len(dst_tensors) self.src_tensors: list[torch.Tensor] = src_tensors self.dst_tensors: list[torch.Tensor] = dst_tensors min_block_size_factor = min(src_block_size_factor, dst_block_size_factor) self.src_block_size_factor: int = src_block_size_factor // min_block_size_factor self.dst_block_size_factor: int = dst_block_size_factor // min_block_size_factor self.block_size_in_bytes = [ tensor.element_size() * tensor.stride(0) * min_block_size_factor for tensor in src_tensors ] self.total_block_size_in_bytes = sum(self.block_size_in_bytes) assert len(src_tensors) > 0 self.gpu_to_cpu: bool = self.src_tensors[0].is_cuda self.transfer_type = ("GPU", "CPU") if self.gpu_to_cpu else ("CPU", "GPU") # job_id -> event self._transfer_events: dict[int, torch.Event] = {} # queue of transfers (job_id, stream, event) self._transfers: deque[Transfer] = deque() # list of CUDA streams available for re-use self._stream_pool: list[torch.cuda.Stream] = [] # list of CUDA events available for re-use self._event_pool: list[torch.Event] = [] def transfer_async(self, job_id: int, transfer_spec: TransferSpec) -> bool: src_spec, dst_spec = transfer_spec assert isinstance(src_spec, BlockIDsLoadStoreSpec) assert isinstance(dst_spec, BlockIDsLoadStoreSpec) src_blocks = src_spec.block_ids dst_blocks = dst_spec.block_ids assert src_blocks.ndim == 1 assert dst_blocks.ndim == 1 src_sub_block_count = src_blocks.size * self.src_block_size_factor dst_sub_block_count = dst_blocks.size * self.dst_block_size_factor src_sub_blocks_to_skip = -dst_blocks.size % self.src_block_size_factor assert dst_sub_block_count == src_sub_block_count - src_sub_blocks_to_skip src_to_dst = np.empty((dst_sub_block_count, 2), dtype=np.int64) expand_block_ids( src_blocks, self.src_block_size_factor, src_to_dst[:, 0], skip_count=src_sub_blocks_to_skip, ) expand_block_ids(dst_blocks, self.dst_block_size_factor, src_to_dst[:, 1]) src_to_dst_tensor = torch.from_numpy(src_to_dst) stream = self._stream_pool.pop() if self._stream_pool else torch.cuda.Stream() start_event = ( self._event_pool.pop() if self._event_pool else torch.Event(enable_timing=True) ) end_event = ( self._event_pool.pop() if self._event_pool else torch.Event(enable_timing=True) ) if self.gpu_to_cpu: # wait for model computation to finish before offloading stream.wait_stream(torch.cuda.current_stream()) if self._transfers: last_transfer: Transfer = self._transfers[-1] last_event = last_transfer.end_event # assure job will start only after the previous one completes stream.wait_event(last_event) with torch.cuda.stream(stream): start_event.record(stream) for src_tensor, dst_tensor, block_size_in_bytes in zip( self.src_tensors, self.dst_tensors, self.block_size_in_bytes, ): ops.swap_blocks( src_tensor, dst_tensor, block_size_in_bytes, src_to_dst_tensor, ) end_event.record(stream) self._transfer_events[job_id] = end_event self._transfers.append( Transfer( job_id=job_id, stream=stream, start_event=start_event, end_event=end_event, num_bytes=dst_sub_block_count * self.total_block_size_in_bytes, ) ) # success return True def get_finished(self) -> list[TransferResult]: results: list[TransferResult] = [] while self._transfers and self._transfers[0].end_event.query(): transfer = self._transfers.popleft() transfer_time = ( transfer.start_event.elapsed_time(transfer.end_event) * 1e-3 ) # elapsed_time is in miliseconds result = TransferResult( job_id=transfer.job_id, success=True, transfer_size=transfer.num_bytes, transfer_time=transfer_time, transfer_type=self.transfer_type, ) results.append(result) self._stream_pool.append(transfer.stream) self._event_pool.append(transfer.end_event) self._event_pool.append(transfer.start_event) del self._transfer_events[transfer.job_id] return results def wait(self, job_ids: set[int]): for job_id in job_ids: event = self._transfer_events.get(job_id) if event is not None: event.synchronize() class CpuGpuOffloadingHandlers: def __init__( self, gpu_block_size: int, cpu_block_size: int, num_cpu_blocks: int, gpu_caches: dict[str, torch.Tensor], attn_backends: dict[str, type[AttentionBackend]], ): assert gpu_caches assert cpu_block_size % gpu_block_size == 0 # find kernel block size and determine layout per each gpu tensor kernel_block_size: int | None = None # list of (gpu_tensor, split_k_and_v) parsed_gpu_tensors: list[tuple[torch.Tensor, bool]] = [] for layer_name, gpu_tensor in gpu_caches.items(): gpu_shape = gpu_tensor.shape attn_backend = attn_backends[layer_name] test_shape = attn_backend.get_kv_cache_shape( num_blocks=1234, block_size=16, num_kv_heads=8, head_size=256 ) has_layers_dim = False split_k_and_v = False if len(gpu_shape) != len(test_shape): # cross-layers tensor # shape is (num_blocks, ...) assert len(gpu_shape) == len(test_shape) + 1 has_layers_dim = True # prepend a dummy num_layers=80 to test_shape test_shape = (80,) + test_shape elif test_shape[0] != 1234: # shape should be (2, num_blocks, ...) assert test_shape[0] == 2 assert test_shape[1] == 1234 assert gpu_shape[0] == 2 split_k_and_v = True try: kv_cache_stride_order = attn_backend.get_kv_cache_stride_order( include_num_layers_dimension=has_layers_dim ) assert len(kv_cache_stride_order) == len(gpu_shape) except (AttributeError, NotImplementedError): kv_cache_stride_order = tuple(range(len(gpu_shape))) # permute test_shape according to stride_order test_shape = tuple(test_shape[i] for i in kv_cache_stride_order) # find block_size (16) dimension index block_size_idx = test_shape.index(16) if kernel_block_size is not None: assert kernel_block_size == gpu_shape[block_size_idx] else: kernel_block_size = gpu_shape[block_size_idx] assert gpu_block_size % kernel_block_size == 0 parsed_gpu_tensors.append((gpu_tensor, split_k_and_v)) assert kernel_block_size is not None cpu_block_size_factor = cpu_block_size // kernel_block_size gpu_block_size_factor = gpu_block_size // kernel_block_size num_cpu_kernel_blocks = num_cpu_blocks * cpu_block_size_factor # allocate cpu tensors pin_memory = is_pin_memory_available() logger.info("Allocating %d CPU tensors...", len(parsed_gpu_tensors)) gpu_tensors: list[torch.Tensor] = [] cpu_tensors: list[torch.Tensor] = [] for gpu_tensor, split_k_and_v in parsed_gpu_tensors: cpu_shape = list(gpu_tensor.shape) cpu_shape[1 if split_k_and_v else 0] = num_cpu_kernel_blocks logger.debug("Allocating CPU tensor of shape %r", cpu_shape) cpu_tensor = torch.zeros( cpu_shape, dtype=gpu_tensor.dtype, device="cpu", pin_memory=pin_memory, ) gpu_tensors.extend(gpu_tensor.unbind(0) if split_k_and_v else [gpu_tensor]) cpu_tensors.extend(cpu_tensor.unbind(0) if split_k_and_v else [cpu_tensor]) self.gpu_to_cpu_handler = SingleDirectionOffloadingHandler( src_tensors=gpu_tensors, dst_tensors=cpu_tensors, src_block_size_factor=gpu_block_size_factor, dst_block_size_factor=cpu_block_size_factor, ) self.cpu_to_gpu_handler = SingleDirectionOffloadingHandler( src_tensors=cpu_tensors, dst_tensors=gpu_tensors, src_block_size_factor=cpu_block_size_factor, dst_block_size_factor=gpu_block_size_factor, )