Files
FastDeploy/fastdeploy/cache_manager/v1/metadata.py
T
kevin 7707be8384 [Feature][KVCache] Implement Cache Manager V1 with GPU + CPU Cache Support (1/n) (#7097)
* [Feature][KVCache] Support cache manager v1 architecture

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* Update cache manager and related modules

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* chore: update cache_manager and related modules

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: add node to evictable set in complete_swap_to_device

When a node transitions from SWAP_TO_DEVICE to DEVICE via
complete_swap_to_device, it was not being added to the
_evictable_device set. This caused nodes with ref_count=0 to
become "orphaned" - not appearing in any evictable set despite
having cache_status=DEVICE.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: update cache manager v1 and related modules

- Add new cache_manager.py with cache management functionality
- Add radix_tree.py for prefix caching
- Update block_pool.py and metadata.py
- Update request.py and resource_manager_v1.py for scheduling
- Update gpu_model_runner.py for GPU model execution

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat(cache): add cache controller v1 implementation

- Add CacheController class for cache management
- Update config.py with cache related configurations
- Refactor gpu_model_runner.py for improved cache handling

* feat(cache_manager): update cache manager v1

* fix(cache_manager): 修复 swap_cache H2D/D2H 方向的 block_ids 逻辑并清理 ForwardMeta

## Motivation

修复 swap_cache_optimized.cu 中 H2D 方向时 src/dst block_ids 使用错误的问题,
并清理 ForwardMeta 中已废弃的 cache_controller 字段。

## Modifications

- fix: swap_cache_optimized.cu 中根据 D2H 模板参数正确选取 src/dst block_ids,
  修复 H2D 方向 src/dst 倒置 bug(同时修复 SwapCachePerLayerImpl 和 SwapCacheAllLayersBatchImpl)
- refactor: cache_manager/v1/__init__.py 将 LayerSwapTimeoutError 导入从
  cache_controller 改为 cache_utils(正确来源)
- refactor: ForwardMeta 移除废弃的 cache_controller 字段
- refactor: gpu_model_runner.py 移除对应的 cache_controller 赋值语句
- test: 新增 tests/cache_manager/v1/test_swap_cache_ops.py 单元测试

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(cache_manager): refactor cache manager v1 and optimize swap ops

## Motivation

对 cache manager v1 进行重构和优化,精简代码结构,提升可维护性。

## Modifications

- 重构 transfer_manager.py,大幅精简代码逻辑
- 优化 swap_cache_optimized.cu GPU 算子实现
- 调整 cache_manager.py、cache_controller.py 逻辑,修复 free_device_blocks 方法缺失问题
- 更新 block_pool.py、cache_utils.py、metadata.py、radix_tree.py
- 精简 gpu_model_runner.py、forward_meta.py、attention.py 中相关调用
- 更新对应单元测试(test_cache_controller、test_swap_cache_ops、test_transfer_manager)
- 调整 config.py 中相关配置项

* [KVCache][MTP] 支持 cache_manager_v1 下的 MTP KV Cache 初始化及多模态 hash

## Motivation

在 enable_cache_manager_v1 路径下,MTP(speculative decode)的 KV Cache 需要由
CacheController 统一管理,以复用 swap/transfer 能力,同时修复多模态场景下 block
hash 未携带 multimodal extra_keys 的问题。

## Modifications

- `cache_controller.py`
  - 新增 `initialize_mtp_kv_cache`:通过 CacheController 初始化 MTP KV Cache,
    并将其注册到 cache_kvs_map,使 transfer_manager 自动覆盖 MTP 层
  - `initialize_host_cache` 中的 num_layers 改为包含 MTP 额外 cache 层数,保证
    Host Cache 也为 MTP 分配足够空间
  - `_free_gpu_cache` 改名为 `free_gpu_cache`(对外可调用)

- `cache_utils.py`
  - 新增 `get_block_hash_extra_keys`:提取单个 block 内的多模态 hash 信息,
    对齐 PrefixCacheManager 的 multimodal extra_keys 逻辑
  - `get_request_block_hasher` 中在 hash_block_tokens 时携带 extra_keys,
    修复多模态场景 prefix cache 命中率不准的问题

- `spec_decode/mtp.py`
  - `update_mtp_block_num` 新增 `skip_cache_init` 参数,避免 v1 cache manager
    路径下重复初始化 MTP KV Cache

- `gpu_model_runner.py`
  - `initialize_kv_cache(v1)` 路径:在主模型 cache 初始化后,调用
    `cache_controller.initialize_mtp_kv_cache` 完成 MTP cache 创建
  - `clear_cache` / `wakeup` / `reset` 等路径:respect `enable_cache_manager_v1`
    标志,跳过重复的 proposer.initialize_kv_cache 调用

## Usage or Command

```bash
# 启动支持 MTP + cache_manager_v1 的推理服务(示例)
bash run.sh
```

* fix(cache_manager): multi-GPU fix, mm hash boundary fix, and remove batch ops

1. Fix CuPy stream/event creation for multi-GPU: wrap all stream operations
   with cp.cuda.Device(device_id) context to ensure streams/events are bound
   to the correct device, preventing cross-device errors in multi-GPU setups.

2. Remove cudaSetDevice from SwapCacheAllLayers (handled by cupy context now).

3. Remove swap_cache_all_layers_batch op: simplified the implementation by
   removing the batch upload variant; all-layer transfers now use the standard
   swap_cache_all_layers with cupy device context.

4. Fix mm hash boundary comparison in get_block_hash_extra_keys: change
   strict less-than (<) to less-than-or-equal (<=) so that multimodal items
   ending exactly at block start are correctly excluded.

5. Extract config fields to KVCacheBase: model_config, cache_config,
   quant_config, parallel_config are now set in the base class __init__ to
   avoid duplication in CacheController and CacheManager subclasses.

6. Translate metadata.py docstrings from Chinese to English for broader
   contributor accessibility.

7. Add test_cache_utils.py: comprehensive unit tests for
   get_block_hash_extra_keys covering all boundary and overlap scenarios.

8. Expand test suite: test_request.py cache fields tests, test_radix_tree.py
   backup candidate tests, test_transfer_manager.py and test_cache_manager.py
   multi-GPU and concurrent operation tests.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [BugFix][KVCache] fix List import and move write_policy normalization to CacheManager

## Motivation

修复两处问题:
1. `fastdeploy/engine/request.py` 中 `List` 未导入导致 pre-commit F821 报错
2. `write_policy` 归一化逻辑(`write_through` → `write_through_selective`)不应放在 `FDConfig`,移至 `CacheManager.__init__` 中,使其只影响 Cache Manager V1 的内部逻辑

## Modifications

- `fastdeploy/engine/request.py`: 在 `typing` 导入中补充 `List`,删除重复的 `CacheSwapMetadata` TYPE_CHECKING 导入,修复 F821/F811
- `fastdeploy/config.py`: 删除 `write_policy` 归一化逻辑
- `fastdeploy/cache_manager/v1/cache_manager.py`: 将归一化逻辑移入 `CacheManager.__init__`

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [BugFix][KVCache] fix pre-commit code style issues

## Motivation

修复 CI pre-commit 代码风格检查失败问题。

## Modifications

- `fastdeploy/engine/common_engine.py`: black 格式化
- `fastdeploy/worker/worker_process.py`: black 格式化 + isort 修复
- `fastdeploy/cache_manager/v1/storage/__init__.py`: isort 修复
- `fastdeploy/worker/gpu_worker.py`: isort 修复

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [Feature][KVCache] update cache_manager_v1 modules

## Motivation

更新 Cache Manager V1 相关模块,完善版权信息、改进模块结构与可维护性。

## Modifications

- `fastdeploy/cache_manager/v1/` 系列模块:补充版权 header,优化代码结构
- `fastdeploy/config.py`:配置项更新
- `fastdeploy/engine/sched/resource_manager_v1.py`:调度相关更新

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [Feature][KVCache] add BatchRequest.from_tasks and refactor worker task parsing

## Motivation

将 worker_process 中重复的 task 解析逻辑收敛到 BatchRequest,减少代码冗余,提升可维护性。

## Modifications

- `fastdeploy/engine/request.py`:新增 `BatchRequest.from_tasks()` 类方法,统一将 task_queue 任务分类为推理请求和控制请求
- `fastdeploy/worker/worker_process.py`:使用 `BatchRequest.from_tasks()` 替代内联解析逻辑,并修复重复的 control_reqs 处理块

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [Feature][KVCache] add NUMA affinity for host cache and skip swap cache tests

## Motivation

优化 Host cache 内存分配的 NUMA 亲和性,减少跨 NUMA 访问延迟;
同时跳过 swap cache ops 测试(当前环境不支持)。

## Modifications

- `fastdeploy/cache_manager/v1/cache_controller.py`:
  - 新增 `_get_numa_node_for_gpu()` 方法,通过 nvidia-smi 或 sysfs 获取 GPU 对应的 NUMA 节点
  - 新增 `_bind_to_closest_numa_node()` 方法,绑定当前线程到 GPU 最近的 NUMA 节点
  - 在 `initialize_host_cache()` 中调用 NUMA 绑定,优化 H2D 传输性能
- `tests/cache_manager/v1/test_swap_cache_ops.py`:跳过所有测试类(`TestSwapCacheAllLayersCorrectness`、`TestSwapCacheAllLayersPerformance`、`TestSwapCacheRandomBlockIndices`)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [BugFix][KVCache] fix unittest failures for cache_manager_v1

三个单测因接口变更或 Mock 方式问题导致失败,需修复。

- tests/distributed/chunked_moe.py:`setup_model_runner` 使用 `__new__` 跳过 `__init__`,补加 `enable_cache_manager_v1 = False`,修复 `AttributeError`
- tests/engine/test_resource_manager.py:`PrefixCacheManager` 为局部导入,`patch` 路径改为定义位置 `fastdeploy.cache_manager.prefix_cache_manager.PrefixCacheManager`
- tests/v1/test_resource_manager_v1.py:`_trigger_preempt` 第四参数已由 `list` 改为 `BatchRequest`,更新测试传参和断言

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [BugFix][KVCache] remove debug logging code

## Modifications

- fastdeploy/engine/request.py:删除调试用 logger 及 prompt_hashes 中的 debug 日志
- fastdeploy/worker/worker_process.py:删除 __main__ 中的调试 import 和 print 语句

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [BugFix][KVCache] fix cupy device id caching and pickle for _match_result

## Motivation

修复两个 bug:
1. `transfer_manager.py` 中每次调用 `cp.cuda.runtime.getDevice()` 存在隐患,应在初始化时缓存为实例变量,保证后续操作使用一致的设备 ID。
2. `request.py` 的 `__getstate__` 未跳过 `_match_result`,该字段包含 BlockNode 树的父子循环引用,pickle 时会触发 `RecursionError`;同时补充 `__setstate__` 确保 unpickle 后字段恢复为安全默认值。

## Modifications

- `transfer_manager.py`:初始化时调用 `cp.cuda.runtime.getDevice()` 并缓存到 `self._cupy_device_id`,后续 `with cp.cuda.Device(...)` 和日志均使用该缓存值。
- `request.py`:
  - `__getstate__` 中将 `_match_result` 加入跳过集合 `_SKIP_KEYS`,避免循环引用导致 pickle 失败。
  - 新增 `__setstate__`,unpickle 后将 `_block_hasher` 和 `_match_result` 恢复为 `None`。

## Usage or Command

* fix(test): fix unit test errors for _trigger_preempt and wakeup with MTP

- Use BatchRequest instead of list in test_trigger_preempt_records_tasks
- Add missing enable_cache_manager_v1 attr in TestSleepWakeupBehavior._make_runner

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [BugFix][KVCache] fix gpu_free_block_list returning wrong block IDs

## Motivation

`gpu_free_block_list` 的兼容 property 中误用了 `list(range(N))`,
将 `available_blocks()` 的返回值当作整数传给 `range()`,
导致返回 `[0, 1, ..., N-1]` 的假列表,而非真实的空闲 block ID。

## Modifications

- `cache_manager/v1/cache_manager.py`:将 `list(range(self._device_pool.available_blocks()))` 改为 `list(self._device_pool.available_blocks())`

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [BugFix][KVCache] 修复 gpu_free_block_list 返回 int 导致 TypeError

## Motivation

gpu_free_block_list 属性中调用 BlockPool.available_blocks(),
该方法返回 int(空闲块数量),用 list() 包装 int 会触发
TypeError: 'int' object is not iterable。

## Modifications

将 list(self._device_pool.available_blocks()) 改为
list(self._device_pool._free_blocks),直接返回空闲块索引列表。

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [KVCache][CacheManager] 适配 V1 CacheManager 的 pause/sleep/free_cache 操作

## Motivation

V1 CacheManager 引入了新的 reset_cache() 接口,pause 和 sleep 操作需要适配,
同时 free_cache 需要支持可选的 clear_storage 参数。

## Modifications

- cache_controller.py: free_cache 新增 clear_storage 参数(默认 False),
  仅当 clear_storage=True 时才调用 _clear_storage(),避免不必要的 storage 清空
- common_engine.py: pause 和 sleep 操作中,当 ENABLE_V1_KVCACHE_MANAGER 时
  使用 cache_manager.reset_cache() 替代旧的 reset() 和 pause_transfer 逻辑
- gpu_model_runner.py: sleep 时仅在非 V1 cache manager 下执行 MTP cache 清除

## Usage or Command

# 启动服务(V1 CacheManager)
python -m fastdeploy.entrypoints.openai.api_server \
  --enable-v1-kvcache-manager \
  ...

* [BugFix][KVCache] fix missing enable_cache_manager_v1 in test mocks and remove unused select_blocks_for_backup

- Remove unused `select_blocks_for_backup` method from radix_tree.py
- Fix `match_prefix` default param `skip_storage=True` and log order in cache_manager.py
- Sync test_gpu_model_runner.py with upstream/develop (add TestInsertTasksV1SplitwiseSuffix)
- Add `enable_cache_manager_v1=False` to all mock runners to fix AttributeError in CI

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [BugFix][KVCache] simplify _free_blocks in ResourceManagerV1 for non-v1 path

Remove redundant prefix_caching branch in else path; always call
recycle_gpu_blocks with full block_tables for non-cache-manager-v1 case.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* [KVCache][Optimization][BugFix] fix and optimize block_pool, cache_manager, transfer_manager, request

## Motivation

修复 cache_manager v1 中若干代码质量问题,提升性能并消除潜在的类型不一致 Bug。

## Modifications

1. **block_pool.py**:`BlockPool.allocate` 将逐个 pop 循环替换为切片 + 批量 set.update,消除 Python 循环开销,O(n) → O(k)(C 层批量操作)
2. **cache_manager.py**:`match_prefix` 在 prefix caching 关闭时提前 return 前写入空 `MatchResult()`,避免调用方解引用 `_match_result=None` 崩溃
3. **transfer_manager.py**:`_build_device_layer_indices` 在 `_cache_kvs_map` 为空时也重置四个层索引列表,防止残留旧 tensor 被 swap 算子使用
4. **request.py**:`BatchRequest.append_swap_metadata` / `append_evict_metadata` 构造 `CacheSwapMetadata` 时将 `src_type`/`dst_type` 从字符串改为 `CacheLevel` 枚举,与字段类型声明一致;补充 `CacheLevel` 导入;`match_result` 属性返回类型标注修正为 `Optional[MatchResult]`
5. **resource_manager_v1.py**:`_allocate_gpu_blocks` 日志从 `INFO` 降级为 `DEBUG`,消除高频调度路径的日志噪音
6. **tests/engine/test_request.py**:同步更新 `src_type`/`dst_type` 断言为 `CacheLevel` 枚举值,补充 `CacheLevel` 导入

## Usage or Command

单元测试:
```bash
source .venv/py310/bin/activate
cd baidu/FastDeploy
python -m pytest tests/cache_manager/v1/test_cache_manager.py -v
python -m pytest tests/cache_manager/v1/test_transfer_manager.py -v
python -m pytest tests/engine/test_request.py -v
```

* [BugFix][KVCache] Fix BlockPool.allocate returns all blocks when num_blocks=0

## Motivation

当 `allocate(num_blocks=0)` 被调用时,Python 负索引陷阱导致严重错误:
`-0 == 0`,所以 `self._free_blocks[-0:]` 等价于 `self._free_blocks[0:]`,
会返回并清空整个空闲块列表,而非返回空列表。

## Modifications

在 `BlockPool.allocate` 中增加对 `num_blocks == 0` 的提前判断,直接返回 `[]`,
避免触发 Python 负索引陷阱。

## Usage or Command

```bash
# 运行相关单元测试验证修复
python -m pytest tests/cache_manager/v1/test_cache_manager.py -vv -s
```

* [KVCache][Test] add unit tests for cache_manager v1 modules

## Motivation

补全 cache_manager/v1 各模块的单测覆盖,确保核心方法有完整的测试保障。

## Modifications

新增/补充以下测试文件,全部 326 个用例通过:

- tests/cache_manager/v1/test_block_pool.py(新建)
  覆盖 BlockPool.get_metadata/set_metadata/resize、DeviceBlockPool/HostBlockPool
- tests/cache_manager/v1/test_metadata.py(新建)
  覆盖 BlockNode、RadixTreeStats、MatchResult、CacheSwapMetadata、AsyncTaskHandler
- tests/cache_manager/v1/test_cache_utils.py(补充)
  新增 hash_block_tokens、get_request_block_hasher、LayerDoneCounter 时间追踪及内部辅助方法
- tests/cache_manager/v1/test_radix_tree.py(补充)
  新增 TestCompleteSwapToDevice 专项测试类(6 个用例)
- tests/cache_manager/v1/test_cache_manager.py(补充)
  新增 offload_to_host、load_from_host、pending backup 系列、prepare_prefetch_metadata
- tests/cache_manager/v1/test_transfer_manager.py(补充)
  新增 _swap_single_layer 校验路径、sync_input/output_stream、record_input_stream_event

## Usage or Command

```bash
# 运行所有新增单测
source .venv/py310/bin/activate
python -m pytest tests/cache_manager/v1/test_block_pool.py \
  tests/cache_manager/v1/test_metadata.py \
  tests/cache_manager/v1/test_cache_utils.py \
  tests/cache_manager/v1/test_radix_tree.py \
  tests/cache_manager/v1/test_cache_manager.py \
  tests/cache_manager/v1/test_transfer_manager.py -v
# 期望结果:326 passed
```

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Jiang-Jia-Jun <163579578+Jiang-Jia-Jun@users.noreply.github.com>
2026-04-21 14:39:00 +08:00

591 lines
19 KiB
Python

"""
# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
import time
import uuid
from dataclasses import dataclass, field
from enum import Enum, auto
from typing import Any, Dict, List, Optional
class TransferStatus(Enum):
"""Status of a transfer task."""
PENDING = "pending"
IN_PROGRESS = "in_progress"
COMPLETED = "completed"
FAILED = "failed"
CANCELLED = "cancelled"
class StorageType(Enum):
"""Supported storage backend types."""
MOONCAKE = "mooncake"
ATTNSTORE = "attnstore"
LOCAL = "local"
class TransferType(Enum):
"""Supported transfer mechanism types."""
RDMA = "rdma"
IPC = "ipc"
class CacheLevel(Enum):
"""Cache hierarchy levels for transfer operations."""
DEVICE = "device"
HOST = "host"
STORAGE = "storage"
class CacheStatus(Enum):
"""Cache status enum representing the current location and state of a BlockNode.
Attributes:
DEVICE: Block is in device (GPU) memory, ready for use. Can be matched.
HOST: Block is in host (CPU) memory, needs to be loaded to device. Can be matched.
SWAP_TO_HOST: Block is being evicted from device to host. Cannot be matched.
SWAP_TO_DEVICE: Block is being loaded from host to device.
LOADING_FROM_STORAGE: Block is being loaded from storage.
DELETING: Block is being deleted (removed from host or deleted when no host cache). Cannot be matched.
"""
DEVICE = auto()
HOST = auto()
SWAP_TO_HOST = auto()
SWAP_TO_DEVICE = auto()
DELETING = auto()
LOADING_FROM_STORAGE = auto()
@dataclass
class RadixTreeStats:
"""
Snapshot of RadixTree statistics.
Encapsulates all state counters for monitoring and statistics.
Returns as a snapshot to ensure consistent values across all fields.
Attributes:
node_count: Total number of nodes in the tree.
evictable_device_count: GPU nodes available for eviction (ref_count==0, status==DEVICE).
evictable_host_count: CPU nodes available for deletion (ref_count==0, status==HOST).
"""
node_count: int = 0
evictable_device_count: int = 0
evictable_host_count: int = 0
@property
def evictable_count(self) -> int:
"""Total evictable nodes count."""
return self.evictable_device_count + self.evictable_host_count
def to_dict(self) -> Dict[str, Any]:
"""Convert to dictionary for JSON serialization."""
return {
"node_count": self.node_count,
"evictable_device_count": self.evictable_device_count,
"evictable_host_count": self.evictable_host_count,
"evictable_count": self.evictable_count,
}
@dataclass
class CacheBlockMetadata:
"""
Metadata for a cache block.
Attributes:
block_id: Unique identifier for the block
device_id: GPU device ID where the block resides
block_size: Size of the block in bytes
ref_count: Reference count for the block
is_pinned: Whether the block is pinned in memory
layer_indices: List of layer indices stored in this block
token_count: Number of tokens in this block
hash_value: Hash value for the block content
last_access_time: Last access timestamp
"""
block_id: int
device_id: int
block_size: int
ref_count: int = 0
is_pinned: bool = False
layer_indices: List[int] = field(default_factory=list)
token_count: int = 0
hash_value: Optional[str] = None
last_access_time: float = 0.0
@dataclass
class TransferTask:
"""
Represents a cache transfer task.
Attributes:
task_id: Unique identifier for the task
src_location: Source location (device/host/storage/remote)
dst_location: Destination location
block_indices: List of block indices to transfer
layer_indices: List of layer indices to transfer
status: Current status of the task
priority: Task priority (lower is higher priority)
created_time: Task creation timestamp
started_time: Task start timestamp
completed_time: Task completion timestamp
error_message: Error message if task failed
metadata: Additional task metadata
"""
task_id: str
src_location: str
dst_location: str
block_indices: List[int] = field(default_factory=list)
layer_indices: List[int] = field(default_factory=list)
status: TransferStatus = TransferStatus.PENDING
priority: int = 0
created_time: float = 0.0
started_time: Optional[float] = None
completed_time: Optional[float] = None
error_message: Optional[str] = None
metadata: Dict[str, Any] = field(default_factory=dict)
@dataclass
class StorageConfig:
"""
Configuration for storage backend.
Attributes:
storage_type: Type of storage backend
storage_path: Base path for storage
max_size_bytes: Maximum storage size in bytes
enable_compression: Whether to enable compression
compression_algorithm: Compression algorithm to use
connection_timeout: Connection timeout in seconds
read_timeout: Read timeout in seconds
write_timeout: Write timeout in seconds
extra_config: Additional backend-specific configuration
"""
storage_type: StorageType = StorageType.MOONCAKE
storage_path: str = ""
max_size_bytes: int = 0
enable_compression: bool = False
compression_algorithm: str = "lz4"
connection_timeout: float = 30.0
read_timeout: float = 60.0
write_timeout: float = 60.0
extra_config: Dict[str, Any] = field(default_factory=dict)
@dataclass
class TransferConfig:
"""
Configuration for transfer mechanism.
Attributes:
transfer_type: Type of transfer mechanism
enable_async: Whether to enable async transfer
max_concurrent_transfers: Maximum concurrent transfer tasks
buffer_size: Buffer size for transfer in bytes
enable_checksum: Whether to enable checksum verification
retry_count: Number of retries on failure
retry_delay: Delay between retries in seconds
extra_config: Additional transfer-specific configuration
"""
transfer_type: TransferType = TransferType.RDMA
enable_async: bool = True
max_concurrent_transfers: int = 4
buffer_size: int = 1024 * 1024 # 1MB
enable_checksum: bool = True
retry_count: int = 3
retry_delay: float = 1.0
extra_config: Dict[str, Any] = field(default_factory=dict)
@dataclass
class BlockNode:
"""
Node in the block management tree.
Represents a node in the radix tree or block allocation structure,
tracking block relationships and reference counts.
Attributes:
node_id: Globally unique identifier for this node (UUID)
block_id: Block identifier (may be reused across device/host)
parent: Parent BlockNode reference (None for root)
children: Dict mapping hash values to child BlockNodes (for radix tree)
children_ids: List of child block IDs
ref_count: Number of references to this block (defaults to 1 on creation)
token_count: Number of tokens stored in this block
hash_value: Hash value for prefix matching
cache_status: Current cache status (DEVICE/HOST/SWAP_TO_HOST/SWAP_TO_DEVICE)
last_access_time: Last access timestamp (defaults to current time on creation)
backuped: Whether this block has a backup on host memory
host_block_id: Host block ID where the backup is stored (if backuped=True)
"""
node_id: str = field(default_factory=lambda: str(uuid.uuid4()))
block_id: int = 0
parent: Optional["BlockNode"] = None
children: Dict[str, "BlockNode"] = field(default_factory=dict)
children_ids: List[int] = field(default_factory=list)
ref_count: int = 0
token_count: int = 0
hash_value: Optional[str] = None
cache_status: CacheStatus = CacheStatus.DEVICE
last_access_time: float = field(default_factory=time.time)
# Backup-related fields
backuped: bool = False # Whether a backup exists on host memory
host_block_id: Optional[int] = None # Host block ID where the backup is stored
hit_count: int = 1 # triggers backup when reaching the threshold
def __post_init__(self):
"""Initialize instance with current time if last_access_time not set."""
if self.last_access_time == 0.0:
self.last_access_time = time.time()
def add_child(self, child_id: int) -> None:
"""Add a child block ID."""
if child_id not in self.children_ids:
self.children_ids.append(child_id)
def remove_child(self, child_id: int) -> bool:
"""Remove a child block ID. Returns True if removed."""
if child_id in self.children_ids:
self.children_ids.remove(child_id)
return True
return False
def increment_ref(self) -> int:
"""Increment reference count and return new count."""
self.ref_count += 1
return self.ref_count
def decrement_ref(self) -> int:
"""Decrement reference count and return new count."""
if self.ref_count > 0:
self.ref_count -= 1
return self.ref_count
def touch(self) -> None:
"""
Update last_access_time to current time.
This method should be called whenever the block is accessed
to track access recency for eviction policies.
"""
self.last_access_time = time.time()
def update_access(self, delta_ref: int = 0) -> None:
"""
Update reference count and last_access_time.
Args:
delta_ref: Change in reference count (positive to increment, negative to decrement)
"""
if delta_ref > 0:
self.ref_count += delta_ref
elif delta_ref < 0:
self.ref_count = max(0, self.ref_count + delta_ref)
self.touch()
def is_leaf(self) -> bool:
"""Check if this is a leaf node (no children)."""
return len(self.children_ids) == 0 and len(self.children) == 0
def is_root(self) -> bool:
"""Check if this is a root node (no parent)."""
return self.parent is None
def is_on_device(self) -> bool:
"""Check if block is on device (GPU) memory."""
return self.cache_status == CacheStatus.DEVICE
def is_on_host(self) -> bool:
"""Check if block is on host (CPU) memory."""
return self.cache_status == CacheStatus.HOST
def is_swapping(self) -> bool:
"""Check if block is currently being swapped or deleted."""
return self.cache_status in (
CacheStatus.SWAP_TO_HOST,
CacheStatus.SWAP_TO_DEVICE,
CacheStatus.DELETING,
)
@dataclass
class MatchResult:
"""
Three-level cache prefix match result.
Contains matched nodes from Device, Host, and Storage levels.
Attributes:
storage_nodes: List of matched BlockNodes in Storage.
device_nodes: List of matched BlockNodes in Device.
host_nodes: List of matched BlockNodes in Host.
"""
device_nodes: List["BlockNode"] = field(default_factory=list)
host_nodes: List["BlockNode"] = field(default_factory=list)
storage_nodes: List["BlockNode"] = field(default_factory=list)
uncached_block_ids: List[int] = field(default_factory=list)
@property
def device_block_ids(self) -> List[int]:
"""Get list of matched device block IDs."""
return [node.block_id for node in self.device_nodes]
@property
def total_matched_blocks(self) -> int:
"""Get total number of matched device blocks."""
return self.matched_device_nums + self.matched_host_nums + self.matched_storage_nums
@property
def matched_device_nums(self) -> int:
"""Get total number of matched device blocks."""
return len(self.device_nodes)
@property
def matched_host_nums(self) -> int:
"""Get total number of matched host blocks."""
return len(self.host_nodes)
@property
def matched_storage_nums(self) -> int:
"""Get total number of matched storage hashes."""
return len(self.storage_nodes)
@dataclass
class StorageMetadata:
"""
Base metadata for storage transfer operations.
Encapsulates all information for storage load/evict operations.
Different storage implementations can extend this class with additional fields.
Attributes:
hash_values: List of hash values to transfer.
block_ids: Target/source host block IDs (pre-allocated by Scheduler).
direction: Transfer direction ("load" from storage, "evict" to storage).
storage_type: Storage type ("mooncake", "attnstore", "rdma", etc.).
endpoint: Storage service endpoint address.
timeout: Operation timeout in seconds.
layer_num: Number of layers to transfer (for layer-by-layer transfer).
extra_params: Storage-specific extra parameters.
"""
hash_values: List[str] = field(default_factory=list)
block_ids: List[int] = field(default_factory=list)
direction: str = "load"
storage_type: str = "mooncake"
endpoint: Optional[str] = None
timeout: float = 30.0
layer_num: int = 0
extra_params: Dict[str, Any] = field(default_factory=dict)
@dataclass
class PDTransferMetadata:
"""
Base metadata for PD separation transfer operations.
Encapsulates all information for cross-node transfer in PD separation architecture.
Different transfer mechanisms (RDMA, IPC) can extend this class with additional fields.
Attributes:
source_node_id: Source node identifier (P node ID).
target_node_id: Target node identifier (D node ID).
block_ids: List of block IDs to transfer.
layer_num: Total number of model layers (for layer-by-layer transfer sync).
timeout: Operation timeout in seconds.
extra_params: Transfer-specific extra parameters.
"""
source_node_id: str = ""
target_node_id: str = ""
block_ids: List[int] = field(default_factory=list)
layer_num: int = 0
timeout: float = 30.0
extra_params: Dict[str, Any] = field(default_factory=dict)
@dataclass
class CacheSwapMetadata:
"""
Metadata for cache transfer operations.
Encapsulates the mapping between source and destination block IDs
for Host↔Device, Storage→Host, and other transfer operations.
Attributes:
src_block_ids: Source block IDs (transfer origin).
dst_block_ids: Destination block IDs (transfer target).
src_type: Source cache level (CacheLevel.DEVICE/HOST/STORAGE).
dst_type: Destination cache level (CacheLevel.DEVICE/HOST/STORAGE).
hash_values: Corresponding hash values (used for storage-related operations).
success: Whether the transfer succeeded.
error_message: Error message if transfer failed.
async_handler: Async task handler for tracking the swap task execution state.
"""
src_block_ids: List[int] = field(default_factory=list)
dst_block_ids: List[int] = field(default_factory=list)
src_type: Optional[CacheLevel] = None
dst_type: Optional[CacheLevel] = None
hash_values: List[str] = field(default_factory=list)
success: bool = False
error_message: Optional[str] = None
async_handler: Optional["AsyncTaskHandler"] = None
def is_success(self) -> bool:
"""Return whether the transfer succeeded."""
return self.success
@property
def mapping(self) -> Dict[int, int]:
"""Get the src -> dst block ID mapping dict."""
if not self.success:
return {}
return dict(zip(self.src_block_ids, self.dst_block_ids))
@dataclass
class TransferResult:
"""
Cache transfer operation result.
Encapsulates the mapping between source and destination block IDs
for Host↔Device, Storage→Host, and other transfer operations.
Attributes:
src_block_ids: Source block IDs (transfer origin).
dst_block_ids: Destination block IDs (transfer target).
src_type: Source cache level (CacheLevel.DEVICE/HOST/STORAGE).
dst_type: Destination cache level (CacheLevel.DEVICE/HOST/STORAGE).
success: Whether the transfer succeeded.
error_message: Error message if transfer failed.
"""
src_block_ids: List[int] = field(default_factory=list)
dst_block_ids: List[int] = field(default_factory=list)
src_type: Optional[CacheLevel] = None
dst_type: Optional[CacheLevel] = None
success: bool = True
error_message: Optional[str] = None
@dataclass
class AsyncTaskHandler:
"""
Async task handler.
Used for submitting and tracking the state of async tasks.
External callers use this handler to check whether a task has completed.
Attributes:
task_id: Unique task identifier.
is_completed: Whether the task has completed.
result: Task result (available after completion).
error: Task error message (if failed).
"""
task_id: str = field(default_factory=lambda: str(uuid.uuid4()))
is_completed: bool = False
result: Optional[Any] = None
error: Optional[str] = None
_event: Any = field(default=None, repr=False)
def __post_init__(self):
"""Initialize event for synchronization."""
import threading
object.__setattr__(self, "_event", threading.Event())
def wait(self, timeout: Optional[float] = None) -> bool:
"""
Wait for the task to complete.
Args:
timeout: Maximum wait time in seconds. None means wait indefinitely.
Returns:
True if completed, False if timed out.
"""
return self._event.wait(timeout=timeout)
def cancel(self) -> bool:
"""
Cancel the task.
Returns:
True if successfully cancelled, False otherwise.
"""
if self.is_completed:
return False
self.error = "Task cancelled"
self.is_completed = True
self._event.set()
return True
def get_result(self) -> Any:
"""
Get the task result (blocking).
Returns:
Task result.
Raises:
RuntimeError: If the task failed or was cancelled.
"""
self._event.wait()
if self.error:
raise RuntimeError(self.error)
return self.result
def set_result(self, result: Any) -> None:
"""
Set the task result and mark as completed.
Args:
result: Task result.
"""
self.result = result
self.is_completed = True
self._event.set()
def set_error(self, error: str) -> None:
"""
Set the error message and mark as completed.
Args:
error: Error message.
"""
self.error = error
self.is_completed = True
self._event.set()