qwen3next-all

缓存长度必须是 64 的整数倍。如果分叉位置是 300,则缓存位置会回退到 floor(300/64)×64=256floor(300 / 64) \times 64 = 256。

  • 如果 当前路径上从未发生分叉 → prefix = 最后一个 64 对齐点
  • 如果 中途发生过分叉 → prefix = 分叉点的最后一个 64 对齐点
  1. CONTENT = “1” * 256 + “2” * 256 # prefix 0
  2. CONTENT = “1” * 256 + “3” * 256 # prefix 0
  3. CONTENT = “1” * 256 + “4” * 256 # prefix 256
  4. CONTENT = “1” * 256 + “2” * 256 + “5” * 256 # prefix 512
  5. CONTENT = “1” * 256 + “4” * 256 + “5” * 256 # prefix 512
  6. CONTENT = “1” * 256 + “3” * 256 + “5” * 256 # prefix 256
  7. CONTENT = “1” * 256 + “3” * 256 # prefix 256
  8. CONTENT = “1” * 256 + “3” * 256 # prefix 512

R8: (1+3) -> Prefix 512

R7: (1+3) -> Prefix 256

R6: (1+3+5) -> Prefix 256

R5: (1+4+5) -> Prefix 512

R4: (1+2+5) -> Prefix 512

R3: (1+4) -> Prefix 256

R2: (1+3) -> Prefix 0

R1: (1+2) -> Prefix 0

Root

1 (256)

2 (256)

Root

1 (256)

2 (256)

3 (256)

Root

1 (256)

2 (256)

3 (256)

4 (256)

Root

1 (256)

2 (256)

5 (256)

3 (256)

4 (256)

Root

1 (256)

2 (256)

5 (256)

3 (256)

4 (256)

5 (256)

Root

1 (256)

2 (256)

5 (256)

3 (256)

5 (256)

4 (256)

5 (256)

Root

1 (256)

2 (256)

5 (256)

3 (256)

5 (256)

4 (256)

5 (256)

Root

1 (256)

...

3 (256)

...

...

mermaid
graph TD
    subgraph S1["R1: (1+2) -> Prefix 0"]
        r1((Root)) --- n1_1["1 (256)"] --- n1_2["2 (256)"]
        style n1_2 fill:#4CAF50,color:#fff
    end

    subgraph S2["R2: (1+3) -> Prefix 0"]
        r2((Root)) --- n2_1["1 (256)"]
        n2_1 --- n2_2["2 (256)"]
        n2_1 --- n2_3["3 (256)"]
        style n2_1 stroke-dasharray: 5 5
        style n2_2 fill:#4CAF50,color:#fff
        style n2_1 fill:#4CAF50,color:#fff
    end

    subgraph S3["R3: (1+4) -> Prefix 256"]
        r3((Root)) --- n3_1["1 (256)"]
        n3_1 --- n3_2["2 (256)"]
        n3_1 --- n3_3["3 (256)"]
        n3_1 --- n3_4["4 (256)"]
        style n3_1 fill:#4CAF50,color:#fff
        style n3_2 fill:#4CAF50,color:#fff
        style n3_4 fill:#4CAF50,color:#fff
    end

    subgraph S4["R4: (1+2+5) -> Prefix 512"]
        r4((Root)) --- n4_1["1 (256)"]
        n4_1 --- n4_2["2 (256)"]
        n4_2 --- n4_5["5 (256)"]
        n4_1 --- n4_3["3 (256)"]
        n4_1 --- n4_4["4 (256)"]
        style n4_1 fill:#4CAF50,color:#fff
        style n4_2 fill:#4CAF50,color:#fff
        style n4_4 fill:#4CAF50,color:#fff
        style n4_5 fill:#4CAF50,color:#fff
    end

    subgraph S5["R5: (1+4+5) -> Prefix 512"]
        r5((Root)) --- n5_1["1 (256)"]
        n5_1 --- n5_2["2 (256)"]
        n5_2 --- n5_2_1["5 (256)"]
        n5_1 --- n5_3["3 (256)"]
        n5_1 --- n5_4["4 (256)"]
        n5_4 --- n5_5["5 (256)"]
        style n5_1 fill:#4CAF50,color:#fff
        style n5_4 fill:#4CAF50,color:#fff
        style n5_2 fill:#4CAF50,color:#fff
        style n5_2_1 fill:#4CAF50,color:#fff
        style n5_5 fill:#4CAF50,color:#fff
    end

    subgraph S6["R6: (1+3+5) -> Prefix 256"]
        r6((Root)) --- n6_1["1 (256)"]
        n6_1 --- n6_2["2 (256)"]
        n6_2 --- n6_2_1["5 (256)"]
        n6_1 --- n6_3["3 (256)"]
        n6_3 --- n6_5["5 (256)"]
        n6_1 --- n6_4["4 (256)"]
        n6_4 --- n6_4_1["5 (256)"]
        style n6_2 fill:#4CAF50,color:#fff
        style n6_2_1 fill:#4CAF50,color:#fff
        style n6_1 fill:#4CAF50,color:#fff
        style n6_4 fill:#4CAF50,color:#fff
        style n6_4_1 fill:#4CAF50,color:#fff
        style n6_5 fill:#4CAF50,color:#fff
    end

    subgraph S7["R7: (1+3) -> Prefix 256"]
        r7((Root)) --- n7_1["1 (256)"]
        n7_1 --- n7_2["2 (256)"]
        n7_2 --- n7_2_1["5 (256)"]
        n7_1 --- n7_3["3 (256)"]
        n7_3 --- n7_5["5 (256)"]
        n7_1 --- n7_4["4 (256)"]
        n7_4 --- n7_4_1["5 (256)"]
        style n7_1 fill:#4CAF50,color:#fff
        style n7_2 fill:#4CAF50,color:#fff
        style n7_2_1 fill:#4CAF50,color:#fff
        style n7_3 fill:#4CAF50,color:#fff
        style n7_5 fill:#4CAF50,color:#fff
        style n7_4 fill:#4CAF50,color:#fff
        style n7_4_1 fill:#4CAF50,color:#fff
    end

    subgraph S8["R8: (1+3) -> Prefix 512"]
        r8((Root)) --- n8_1["1 (256)"]
        n8_1 --- n8_2["..."]
        n8_1["1 (256)"] --- n8_3["3 (256)"]
        n8_3["3 (256)"] --- n8_3_1["..."]
        n8_1 --- n8_4["..."]
        style n8_1 fill:#4CAF50,color:#fff
        style n8_2 fill:#4CAF50,color:#fff
        style n8_3 fill:#4CAF50,color:#fff
        style n8_4 fill:#4CAF50,color:#fff
        style n8_3_1 fill:#4CAF50,color:#fff
    end

    S1 --> S2 --> S3 --> S4 --> S5 --> S6 --> S7 --> S8

1. 总体架构

1.1 组件关系图

HybridReqToTokenPool (请求到Token的映射池)
  │
  ├─> ReqToTokenPool (基础类)
  │    └─> req_to_token: [size, max_context_len]
  │         # 映射请求索引到token位置索引
  │
  ├─> MambaPool (Mamba状态池)
  │    │
  │    ├─> State (正常状态)
  │    │    ├─> conv: List[torch.Tensor]  # 卷积状态
  │    │    │     Shape: [num_layers, size + 1, conv_dim, kernel_size]
  │    │    └─> temporal: torch.Tensor    # SSM时序状态
  │    │          Shape: [num_layers, size + 1, num_heads, head_dim, 2]
  │    │
  │    ├─> SpeculativeState (投机采样状态, 可选)
  │    │    ├─> intermediate_ssm: torch.Tensor
  │    │    └─> intermediate_conv_window: List[torch.Tensor]
  │    │
  │    └─> free_slots: torch.Tensor  # 空闲slot索引
  │
  ├─> req_index_to_mamba_index_mapping: torch.Tensor
  │    # Shape: [size]
  │    # 映射请求索引 → mamba slot索引
  │
  └─> req_index_to_mamba_ping_pong_track_buffer_mapping: torch.Tensor (extra_buffer模式,可选)
        # Shape: [size, 2]
        # 映射请求索引 → [ping_slot_idx, pong_slot_idx]

1.2 内存布局

GPU Memory Layout
├─────────────────────────────────────────────────────────┐
│ Model Weights                                          │
├─────────────────────────────────────────────────────────┤
│ KV Cache (MHATokenToKVPool / MLATokenToKVPool)         │
│ - k_buffer: [num_layers, size, num_heads, head_dim]    │
│ - v_buffer: [num_layers, size, num_heads, head_dim]    │
├─────────────────────────────────────────────────────────┤
│ Mamba Cache (MambaPool)                                │
│ - conv_state: List[torch.Tensor]                       │
│   [num_layers, size+1, conv_dims]                      │
│ - temporal_state: torch.Tensor                         │
│   [num_layers, size+1, temporal_dims]                  │
└─────────────────────────────────────────────────────────┘

1.3 关键参数

参数 类型 默认值 说明
size int - Mamba pool大小 (slot数量)
spec_state_size int - 投机采样状态池大小
enable_mamba_extra_buffer bool False 是否启用ping-pong buffer
mamba_ping_pong_track_buffer_size int 2 Ping-pong buffer大小
speculative_num_draft_tokens int None 投机采样draft token数量

2. MambaPool 类详解

文件位置: python/sglang/srt/mem_cache/memory_pool.py:128-330

2.1 类定义和状态数据结构

python
class MambaPool:
    """Mamba状态内存池

    功能:
    1. 管理线性注意力层的SSM状态 (卷积状态 + 时序状态)
    2. 提供slot分配/释放机制
    3. 支持状态复制 (copy_from) 和写时复制 (fork_from)
    4. 可选的投机采样状态缓存
    """

    @dataclass(frozen=True, kw_only=True)
    class State:
        """基础Mamba状态数据结构

        Attributes:
            conv: 卷积状态列表,每个元素是一个tensor
                  每个卷积状态可能有多维 (如不同层的卷积维度不同)
            temporal: SSM递归状态
        """
        conv: List[torch.Tensor]
        temporal: torch.Tensor

        def at_layer_idx(self, layer: int) -> "State":
            """提取指定层的状态

            Args:
                layer: 层索引

            Returns:
                只包含指定层的新State对象

            示例:
                >>> # 获取第5层的状态
                >>> layer_5_state = mamba_cache.at_layer_idx(5)
                >>> # layer_5_state.conv 是 [conv[5]]
                >>> # layer_5_state.temporal 是 temporal[5]
            """
            kwargs = {}
            for k, v in vars(self).items():
                if k == "conv" or k == "intermediate_conv_window":
                    # conv是列表,每个元素对应一层
                    kwargs[k] = [conv[layer] for conv in v]
                else:
                    # temporal是tensor,第一个维度是层
                    kwargs[k] = v[layer]
            return type(self)(**kwargs)

        def mem_usage_bytes(self) -> int:
            """计算当前状态的内存占用 (字节)

            Returns:
                所有tensor的总字节数
            """
            return sum(
                get_tensor_size_bytes(getattr(self, f.name))
                for f in dataclasses.fields(self)
            )

2.2 SpeculativeState (投机采样状态)

python
    @dataclass(frozen=True, kw_only=True)
    class SpeculativeState(State):
        """投机采样时的扩展状态

        在投机采样模式下,需要缓存中间状态用于验证:
        - intermediate_ssm: 每个draft token的SSM状态
        - intermediate_conv_window: 每个draft token的卷积窗口

        Attributes:
            intermediate_ssm: 中间SSM状态
                Shape: [num_layers, size+1, num_draft_tokens, num_heads, head_dim, 2]
            intermediate_conv_window: 中间卷积窗口
                每个conv shape: [num_layers, size+1, num_draft_tokens, dim, kernel_size]
        """
        intermediate_ssm: torch.Tensor
        intermediate_conv_window: List[torch.Tensor]

2.3 初始化方法

python
    def __init__(
        self,
        *,
        size: int,                              # 池大小 (slot数量)
        spec_state_size: int,                   # 投机采样状态池大小
        cache_params: BaseLinearStateParams,    # 缓存参数 (包含shape和dtype)
        device: str,                            # 设备 ("cuda" 或 "cpu")
        enable_memory_saver: bool = False,      # 是否启用内存节省模式
        speculative_num_draft_tokens: Optional[int] = None,  # Draft token数量
    ):
        """初始化Mamba内存池

        Args:
            size: 最大支持的并发请求数
            spec_state_size: 投机采样状态池大小
            cache_params: 包含以下属性:
                - shape.conv: List[tuple] - 卷积状态shape列表
                - shape.temporal: tuple - 时序状态shape
                - dtype.conv: torch.dtype - 卷积状态数据类型
                - dtype.temporal: torch.dtype - 时序状态数据类型
                - layers: List[int] - Mamba层ID列表
            device: 设备字符串
            enable_memory_saver: 是否使用TorchMemorySaverAdapter
            speculative_num_draft_tokens: 投机采样draft token数量
        """
        # ===== 1. 提取配置参数 =====
        conv_state_shape = cache_params.shape.conv      # 卷积状态shape列表
        temporal_state_shape = cache_params.shape.temporal  # 时序状态shape
        conv_dtype = cache_params.dtype.conv            # 卷积状态数据类型 (float32)
        ssm_dtype = cache_params.dtype.temporal         # SSM状态数据类型 (float32)

        # 内存节省适配器 (用于CPU卸载)
        self.memory_saver_adapter = TorchMemorySaverAdapter.create(
            enable=enable_memory_saver
        )

        num_mamba_layers = len(cache_params.layers)     # Mamba层数量

        self.size = size                                # 记录池大小
        self.device = device                            # 记录设备

        # ===== 2. NVLink disaggregation 支持 =====
        # 用于多GPU分布式场景,使用自定义内存池
        self.enable_custom_mem_pool, self.custom_mem_pool, _ = (
            maybe_init_custom_mem_pool(device=self.device)
        )

        # ===== 3. 分配内存 =====
        # 使用内存池上下文管理器 (如果启用custom_mem_pool)
        with self.memory_saver_adapter.region(GPU_MEMORY_TYPE_KV_CACHE), (
            torch.cuda.use_mem_pool(self.custom_mem_pool)
            if self.enable_custom_mem_pool
            else nullcontext()
        ):
            # ===== 3.1 分配卷积状态 =====
            # conv_state 是一个列表,因为不同层的卷积维度可能不同
            # 每个 tensor shape: [num_layers, size + 1, conv_dim, conv_kernel_size]
            # 注意: size + 1 是因为 slot 0 用作 padding (避免边界检查)
            conv_state = [
                torch.zeros(
                    size=(num_mamba_layers, size + 1) + conv_shape,
                    dtype=conv_dtype,
                    device=device,
                )
                for conv_shape in conv_state_shape
            ]

            # ===== 3.2 分配时序状态 =====
            # temporal_state shape: [num_layers, size + 1, num_heads, head_dim, 2]
            # 最后的维度2存储:
            #   - index 0: h (主状态)
            #   - index 1: 辅助状态 (用于门控等)
            temporal_state = torch.zeros(
                size=(num_mamba_layers, size + 1) + temporal_state_shape,
                dtype=ssm_dtype,
                device=device,
            )

            # ===== 3.3 分配投机采样状态 (可选) =====
            if speculative_num_draft_tokens is not None:
                # Cache中间SSM状态: 每个draft token一个状态
                # Shape: [num_layers, size + 1, speculative_num_draft_tokens, HV, K, V]
                # 说明:
                #   - num_layers: Mamba层数量
                #   - size + 1: 池大小 (slot 0用于padding)
                #   - speculative_num_draft_tokens: draft token数量
                #   - HV: value头数量 (linear_num_value_heads)
                #   - K: key头维度 (linear_key_head_dim)
                #   - V: value头维度 (linear_value_head_dim)
                intermediate_ssm_state_cache = torch.zeros(
                    size=(
                        num_mamba_layers,
                        spec_state_size + 1,
                        speculative_num_draft_tokens,
                        temporal_state_shape[0],
                        temporal_state_shape[1],
                        temporal_state_shape[2],
                    ),
                    dtype=ssm_dtype,
                    device="cuda",
                )

                # Cache中间卷积窗口: 每个draft token一个窗口
                # 存储最近K-1个输入 (K = conv_kernel_size)
                # Shape: [num_layers, size + 1, speculative_num_draft_tokens, dim, K-1]
                intermediate_conv_window_cache = [
                    torch.zeros(
                        size=(
                            num_mamba_layers,
                            spec_state_size + 1,
                            speculative_num_draft_tokens,
                            conv_shape[0],
                            conv_shape[1],
                        ),
                        dtype=conv_dtype,
                        device="cuda",
                    )
                    for conv_shape in conv_state_shape
                ]

                # 创建SpeculativeState对象
                self.mamba_cache = self.SpeculativeState(
                    conv=conv_state,
                    temporal=temporal_state,
                    intermediate_ssm=intermediate_ssm_state_cache,
                    intermediate_conv_window=intermediate_conv_window_cache,
                )

                # 记录内存使用情况
                logger.info(
                    f"Mamba Cache is allocated. "
                    f"max_mamba_cache_size: {size}, "
                    f"conv_state size: {get_tensor_size_bytes(conv_state) / GB:.2f}GB, "
                    f"ssm_state size: {get_tensor_size_bytes(temporal_state) / GB:.2f}GB "
                    f"intermediate_ssm_state_cache size: {get_tensor_size_bytes(intermediate_ssm_state_cache) / GB:.2f}GB "
                    f"intermediate_conv_window_cache size: {get_tensor_size_bytes(intermediate_conv_window_cache) / GB:.2f}GB "
                )
            else:
                # 非投机采样模式: 创建基础State对象
                self.mamba_cache = self.State(
                    conv=conv_state,
                    temporal=temporal_state
                )

                # 记录内存使用情况
                logger.info(
                    f"Mamba Cache is allocated. "
                    f"max_mamba_cache_size: {size}, "
                    f"conv_state size: {get_tensor_size_bytes(conv_state) / GB:.2f}GB, "
                    f"ssm_state size: {get_tensor_size_bytes(temporal_state) / GB:.2f}GB "
                )

            # ===== 3.4 初始化空闲slot列表 =====
            # free_slots: 可用的slot索引列表
            # 初始时所有slot都可用: [0, 1, 2, ..., size-1]
            # 注意: slot 0 实际被用作padding,分配时会跳过
            self.free_slots = torch.arange(
                self.size, dtype=torch.int64, device=self.device
            )

            # 记录总内存使用
            self.mem_usage = self.mamba_cache.mem_usage_bytes() / GB
            self.num_mamba_layers = num_mamba_layers

2.4 内存分配和释放

python
    def available_size(self) -> int:
        """获取可用slot数量

        Returns:
            当前空闲的slot数量
        """
        return len(self.free_slots)

    def alloc(self, need_size: int) -> Optional[torch.Tensor]:
        """分配mamba cache slot

        Args:
            need_size: 需要分配的slot数量 (通常为1)

        Returns:
            分配的slot索引tensor,如果空间不足返回None

        注意:
            - 分配时会清零对应的slot
            - Slot 0 作为padding,实际分配从slot 1开始
        """
        # 检查是否有足够的空闲slot
        if need_size > len(self.free_slots):
            return None

        # 从free_slots前面取need_size个slot
        select_index = self.free_slots[:need_size]
        # 更新free_slots,移除已分配的slot
        self.free_slots = self.free_slots[need_size:]

        # 清零分配的slot (避免使用旧数据)
        for i in range(len(self.mamba_cache.conv)):
            self.mamba_cache.conv[i][:, select_index] = 0
        self.mamba_cache.temporal[:, select_index] = 0

        # 再次清零 (确保完全清零,可能是冗余操作)
        for i in range(len(self.mamba_cache.conv)):
            self.mamba_cache.conv[i][:, select_index] = 0
        self.mamba_cache.temporal[:, select_index] = 0

        return select_index

    def free(self, free_index: torch.Tensor):
        """释放mamba cache slot

        Args:
            free_index: 要释放的slot索引tensor

        注意:
            - 释放时不会清零数据,只是将索引加回free_slots
            - 下次分配时会清零
        """
        if free_index.numel() == 0:
            return
        # 将释放的索引追加到free_slots
        self.free_slots = torch.cat((self.free_slots, free_index))

    def clear(self):
        """清空所有slot (重置池状态)

        用途:
            - 测试时重置
            - 发生错误时恢复
        """
        self.free_slots = torch.arange(
            self.size, dtype=torch.int64, device=self.device
        )

2.5 状态复制和写时复制

python
    def copy_from(self, src_index: torch.Tensor, dst_index: torch.Tensor):
        """复制mamba状态: src → dst

        Args:
            src_index: 源slot索引
            dst_index: 目标slot索引

        用途:
            - Radix cache中的copy-on-write
            - Prefix cache命中时的状态复制

        示例:
            >>> # 从slot 5复制到slot 10
            >>> pool.copy_from(torch.tensor([5]), torch.tensor([10]))
        """
        # 复制所有卷积状态
        for i in range(len(self.mamba_cache.conv)):
            self.mamba_cache.conv[i][:, dst_index] = self.mamba_cache.conv[i][
                :, src_index
            ]
        # 复制时序状态
        self.mamba_cache.temporal[:, dst_index] = self.mamba_cache.temporal[
            :, src_index
        ]
        return

    def fork_from(self, src_index: torch.Tensor) -> Optional[torch.Tensor]:
        """写时复制: 分配新slot并复制源状态

        Args:
            src_index: 源slot索引

        Returns:
            新分配的slot索引,如果分配失败返回None

        用途:
            - Radix cache的copy-on-write机制
            - 多个请求共享相同prefix时fork状态

        流程:
            1. 分配新slot
            2. 从src复制状态到新slot
            3. 返回新slot索引

        示例:
            >>> # Fork slot 5到新slot
            >>> new_slot = pool.fork_from(torch.tensor([5]))
            >>> # new_slot 现在包含 slot 5 的状态副本
        """
        dst_index = self.alloc(1)
        if dst_index == None:
            return None
        self.copy_from(src_index, dst_index)
        return dst_index

2.6 辅助方法

python
    def get_speculative_mamba2_params_all_layers(self) -> SpeculativeState:
        """获取所有层的投机采样状态

        Returns:
            SpeculativeState对象

        断言:
            当前mamba_cache必须是SpeculativeState类型
        """
        assert isinstance(self.mamba_cache, self.SpeculativeState)
        return self.mamba_cache

    def mamba2_layer_cache(self, layer_id: int) -> State:
        """获取指定层的mamba cache

        Args:
            layer_id: 层ID (全局层索引)

        Returns:
            只包含指定层的State对象

        示例:
            >>> layer_cache = pool.mamba2_layer_cache(5)
            >>> # layer_cache.conv 是 [conv[5]]
            >>> # layer_cache.temporal 是 temporal[5]
        """
        return self.mamba_cache.at_layer_idx(layer_id)

    def get_contiguous_buf_infos(self) -> Tuple[List[int], List[int], List[int]]:
        """获取所有状态tensor的内存信息

        Returns:
            data_ptrs: 每层每个tensor的数据指针列表
            data_lens: 每层每个tensor的总字节数
            item_lens: 每层每个tensor的单个元素字节数

        用途:
            - NVLink disaggregation
            - 跨GPU内存共享
        """
        state_tensors = []
        # 将mamba_cache中的所有tensor展平成列表
        for field in vars(self.mamba_cache):
            value = getattr(self.mamba_cache, field)
            if isinstance(value, list):
                state_tensors.extend(value)
            else:
                state_tensors.append(value)

        data_ptrs, data_lens, item_lens = [], [], []

        # 遍历所有层,收集内存信息
        for _, state_tensor in enumerate(state_tensors):
            data_ptrs += [
                state_tensor[i].data_ptr() for i in range(self.num_mamba_layers)
            ]
            data_lens += [state_tensor[i].nbytes for i in range(self.num_mamba_layers)]
            item_lens += [
                state_tensor[i][0].nbytes for i in range(self.num_mamba_layers)
            ]

        return data_ptrs, data_lens, item_lens

3. HybridReqToTokenPool 类详解

文件位置: python/sglang/srt/mem_cache/memory_pool.py:332-510

3.1 类定义

python
class HybridReqToTokenPool(ReqToTokenPool):
    """混合请求到Token的映射池

    继承自 ReqToTokenPool,添加了Mamba cache支持

    功能:
    1. 映射请求索引 → token位置 (基础功能)
    2. 映射请求索引 → mamba slot索引 (新增)
    3. 管理ping-pong buffer分配 (extra_buffer模式)

    内存布局:
        req_to_token: [size, max_context_len]
        req_index_to_mamba_index_mapping: [size]
        req_index_to_mamba_ping_pong_track_buffer_mapping: [size, 2] (可选)
    """

    def __init__(
        self,
        *,
        size: int,                              # 请求池大小
        mamba_size: int,                        # Mamba pool大小
        mamba_spec_state_size: int,             # Mamba投机采样状态池大小
        max_context_len: int,                   # 最大上下文长度
        device: str,                            # 设备
        enable_memory_saver: bool,              # 是否启用内存节省
        cache_params: BaseLinearStateParams,    # Mamba cache参数
        enable_mamba_extra_buffer: bool,        # 是否启用ping-pong buffer
        speculative_num_draft_tokens: int = None,  # 投机采样draft token数
    ):
        # ===== 1. 初始化基类 ReqToTokenPool =====
        # 分配 req_to_token: [size, max_context_len]
        super().__init__(
            size=size,
            max_context_len=max_context_len,
            device=device,
            enable_memory_saver=enable_memory_saver,
        )

        # ===== 2. 设置ping-pong buffer大小 =====
        # 正常模式: 2个slot (ping和pong)
        # 投机采样模式: 1个slot (因为已有intermediate状态)
        self.mamba_ping_pong_track_buffer_size = (
            2 if speculative_num_draft_tokens is None else 1
        )
        self.enable_mamba_extra_buffer = enable_mamba_extra_buffer
        self.enable_memory_saver = enable_memory_saver

        # ===== 3. 初始化MambaPool =====
        self._init_mamba_pool(
            size=mamba_size,
            mamba_spec_state_size=mamba_spec_state_size,
            cache_params=cache_params,
            device=device,
            enable_mamba_extra_buffer=enable_mamba_extra_buffer,
            speculative_num_draft_tokens=speculative_num_draft_tokens,
        )

3.2 MambaPool初始化

python
    def _init_mamba_pool(
        self,
        size: int,                              # Mamba pool大小
        mamba_spec_state_size: int,             # 投机采样状态池大小
        cache_params: BaseLinearStateParams,    # 缓存参数
        device: str,                            # 设备
        enable_mamba_extra_buffer: bool,        # 是否启用ping-pong buffer
        speculative_num_draft_tokens: int = None,  # Draft token数量
    ):
        """初始化Mamba内存池和映射表

        功能:
            1. 创建MambaPool实例
            2. 创建层ID映射表
            3. 创建请求到mamba slot的映射表
            4. 创建请求到ping-pong buffer的映射表 (extra_buffer模式)
        """
        # ===== 1. 创建MambaPool =====
        self.mamba_pool = MambaPool(
            size=size,
            spec_state_size=mamba_spec_state_size,
            cache_params=cache_params,
            device=device,
            enable_memory_saver=self.enable_memory_saver,
            speculative_num_draft_tokens=speculative_num_draft_tokens,
        )

        # ===== 2. 创建层ID映射 =====
        # mamba_map: 全局层ID → 连续层索引
        # 示例: {5: 0, 10: 1, 15: 2}
        # 说明: 只有Mamba层才会在map中
        self.mamba_map = {
            layer_id: i
            for i, layer_id in enumerate(cache_params.layers)
        }

        self.device = device

        # ===== 3. 创建请求索引 → mamba slot索引映射 =====
        # Shape: [size]
        # 值: 每个请求对应的mamba slot索引
        # 初始值: 0 (表示未分配)
        self.req_index_to_mamba_index_mapping: torch.Tensor = torch.zeros(
            size, dtype=torch.int32, device=self.device
        )

        # ===== 4. 创建请求索引 → ping-pong buffer映射 (可选) =====
        # Shape: [size, mamba_ping_pong_track_buffer_size]
        #   - 正常模式: [size, 2]
        #   - 投机采样: [size, 1]
        # 值: 每个请求对应的ping-pong slot索引
        # 初始值: 0 (表示未分配)
        if enable_mamba_extra_buffer:
            self.req_index_to_mamba_ping_pong_track_buffer_mapping: torch.Tensor = (
                torch.zeros(
                    (size, self.mamba_ping_pong_track_buffer_size),
                    dtype=torch.int32,
                    device=self.device,
                )
            )

3.3 分配方法

python
    def alloc(
        self,
        need_size: int,                      # 需要分配的请求数量
        reqs: Optional[List["Req"]]          # 请求对象列表
    ) -> Optional[List[int]]:
        """分配请求slot和对应的mamba cache

        Args:
            need_size: 需要分配的请求数
            reqs: 请求对象列表,每个请求包含:
                - req.mamba_pool_idx: 已分配的mamba slot (可能来自radix cache)
                - req.mamba_ping_pong_track_buffer: ping-pong buffer (extra_buffer模式)

        Returns:
            分配的请求索引列表,分配失败返回None

        分配流程:
            1. 分配请求slot (基类)
            2. 为每个请求分配mamba slot
            3. 如果启用extra_buffer,分配ping-pong buffer
            4. 更新映射表

        注意:
            - 如果请求已有mamba_pool_idx (来自radix cache),复用之
            - Chunked prefill时不重新分配mamba cache
        """
        assert reqs is not None

        # ===== 1. 分配请求slot (基类) =====
        select_index = super().alloc(need_size)
        if select_index == None:
            return None

        # ===== 2. 为每个请求分配mamba slot =====
        mamba_index = []
        mamba_ping_pong_track_buffer_list = []

        for req in reqs:
            mid = None

            # 检查是否已有mamba slot (来自radix cache)
            if req.mamba_pool_idx is not None:
                # Radix cache命中,复用已有slot
                mid = req.mamba_pool_idx
            else:
                # 分配新的mamba slot
                mid = self.mamba_pool.alloc(1)
                assert (
                    mid is not None
                ), (
                    f"Not enough space for mamba cache, "
                    f"try to increase --mamba-full-memory-ratio or --max-mamba-cache-size. "
                    f"{mid=}, {self.mamba_pool.size=}, "
                    f"{self.mamba_pool.available_size()=}, {len(reqs)=}"
                )
                mid = mid[0]  # 从tensor转为int
                req.mamba_pool_idx = mid  # 记录到请求对象

            mamba_index.append(mid)

            # ===== 3. 分配ping-pong buffer (extra_buffer模式) =====
            if self.enable_mamba_extra_buffer:
                if req.mamba_ping_pong_track_buffer is None:
                    # 首次分配ping-pong buffer
                    req.mamba_ping_pong_track_buffer = self.mamba_pool.alloc(
                        self.mamba_ping_pong_track_buffer_size  # 2或1
                    )
                    assert (
                        req.mamba_ping_pong_track_buffer is not None
                    ), "Not enough space for mamba ping pong idx, try to increase --mamba-full-memory-ratio."
                    req.mamba_next_track_idx = 0  # 初始化ping-pong索引

                # 记录ping-pong buffer索引
                mamba_ping_pong_track_buffer_list.append(
                    req.mamba_ping_pong_track_buffer.tolist()
                )

        # ===== 4. 验证分配成功 =====
        assert len(select_index) == len(
            mamba_index
        ), f"Not enough space for mamba cache, try to increase --mamba-full-memory-ratio or --max-mamba-cache-size."

        if self.enable_mamba_extra_buffer:
            assert len(select_index) == len(
                mamba_ping_pong_track_buffer_list
            ), f"Not enough space for mamba ping pong idx, try to increase --mamba-full-memory-ratio."

        # ===== 5. 更新映射表 =====
        # req_index_to_mamba_index_mapping[request_idx] = mamba_slot_idx
        self.req_index_to_mamba_index_mapping[select_index] = torch.tensor(
            mamba_index, dtype=torch.int32, device=self.device
        )

        # req_index_to_mamba_ping_pong_track_buffer_mapping[request_idx] = [slot1, slot2]
        if self.enable_mamba_extra_buffer:
            self.req_index_to_mamba_ping_pong_track_buffer_mapping[select_index] = (
                torch.tensor(
                    mamba_ping_pong_track_buffer_list,
                    dtype=torch.int32,
                    device=self.device,
                )
            )

        return select_index

3.4 释放方法

python
    def free(
        self,
        free_index: Union[int, List[int]],     # 要释放的请求索引
        free_mamba_cache: bool = True,         # 是否释放mamba cache
        mamba_ping_pong_track_buffer_to_keep: Optional[int] = None,  # 保留哪个ping-pong slot
    ):
        """释放请求slot和对应的mamba cache

        Args:
            free_index: 请求索引 (单个或列表)
            free_mamba_cache: 是否释放mamba cache (chunked prefill时为False)
            mamba_ping_pong_track_buffer_to_keep: 保留的ping-pong slot索引 (0或1)
                - 用于prefix cache保存时保留一个ping-pong slot
                - None表示释放所有ping-pong slot

        释放策略:
            - 基础请求slot: 总是释放
            - Mamba工作slot: 可选释放 (chunked prefill时保留)
            - Ping-pong buffer: 可选保留一个slot

        示例:
            >>> # 释放所有cache
            >>> pool.free(req_idx, free_mamba_cache=True)
            >>>
            >>> # Chunked prefill: 保留mamba cache
            >>> pool.free(req_idx, free_mamba_cache=False)
            >>>
            >>> # Prefix cache: 保留ping-pong slot 0
            >>> pool.free(req_idx, free_mamba_cache=True, mamba_ping_pong_track_buffer_to_keep=0)
        """
        if isinstance(free_index, (int,)):
            free_index = [free_index]

        # ===== 1. 释放基础请求slot =====
        super().free(free_index)

        # ===== 2. 释放mamba cache (可选) =====
        if free_mamba_cache:
            # 获取要释放的mamba slot索引
            mamba_index = self.req_index_to_mamba_index_mapping[free_index]
            self.mamba_pool.free(mamba_index)

            # ===== 3. 释放ping-pong buffer (extra_buffer模式) =====
            if self.enable_mamba_extra_buffer:
                # 获取ping-pong buffer索引
                mamba_ping_pong_track_buffer_to_free = (
                    self.req_index_to_mamba_ping_pong_track_buffer_mapping[
                        free_index
                    ].squeeze(0)
                )

                # 如果指定要保留某个slot
                if mamba_ping_pong_track_buffer_to_keep is not None:
                    assert mamba_ping_pong_track_buffer_to_keep in [0, 1], (
                        f"mamba_ping_pong_track_buffer_to_keep must be 0 or 1, "
                        f"{mamba_ping_pong_track_buffer_to_keep=}"
                    )

                    # 创建所有slot的索引
                    idx_to_free = list(range(self.mamba_ping_pong_track_buffer_size))
                    # 移除要保留的slot
                    idx_to_free.remove(mamba_ping_pong_track_buffer_to_keep)
                    # 获取要释放的slot (另一个)
                    mamba_ping_pong_track_buffer_to_free = (
                        mamba_ping_pong_track_buffer_to_free[idx_to_free]
                    )

                # 释放ping-pong buffer (可能只释放一个)
                self.mamba_pool.free(mamba_ping_pong_track_buffer_to_free)

    def clear(self):
        """清空所有slot (重置池状态)"""
        logger.info("Reset HybridReqToTokenPool")

        # 清空基类
        super().clear()

        # 清空mamba pool
        self.mamba_pool.clear()

        # 清零映射表
        self.req_index_to_mamba_index_mapping.zero_()

        if self.enable_mamba_extra_buffer:
            self.req_index_to_mamba_ping_pong_track_buffer_mapping.zero_()

3.5 辅助方法

python
    def get_mamba_indices(self, req_indices: torch.Tensor) -> torch.Tensor:
        """获取请求索引对应的mamba slot索引

        Args:
            req_indices: 请求索引tensor

        Returns:
            mamba slot索引tensor

        示例:
            >>> req_indices = torch.tensor([5, 10, 15])
            >>> mamba_indices = pool.get_mamba_indices(req_indices)
            >>> # mamba_indices: [3, 7, 12]
        """
        return self.req_index_to_mamba_index_mapping[req_indices]

    def mamba2_layer_cache(self, layer_id: int):
        """获取指定层的mamba cache

        Args:
            layer_id: 层ID (全局层索引)

        Returns:
            指定层的State对象

        断言:
            layer_id必须在mamba_map中 (即必须是Mamba层)
        """
        assert layer_id in self.mamba_map
        # 将全局层ID转换为连续索引
        return self.mamba_pool.mamba2_layer_cache(self.mamba_map[layer_id])

    def get_speculative_mamba2_params_all_layers(self) -> MambaPool.SpeculativeState:
        """获取所有层的投机采样mamba状态

        Returns:
            SpeculativeState对象
        """
        return self.mamba_pool.get_speculative_mamba2_params_all_layers()

    def get_mamba_ping_pong_other_idx(self, mamba_next_track_idx: int) -> int:
        """获取ping-pong buffer的另一个索引

        Args:
            mamba_next_track_idx: 当前ping-pong索引 (0或1)

        Returns:
            另一个索引 (1或0)

        示例:
            >>> pool.get_mamba_ping_pong_other_idx(0)  # 返回 1
            >>> pool.get_mamba_ping_pong_other_idx(1)  # 返回 0

        注意:
            - 投机采样模式下 (buffer_size=1),返回自身
        """
        if self.mamba_ping_pong_track_buffer_size == 2:
            return 1 - mamba_next_track_idx
        else:
            return mamba_next_track_idx

4. 内存分配流程

4.1 整体初始化流程

1. 计算内存大小
   └─> model_runner_kv_cache_mixin.py:179-192
       计算公式: mamba_memory = total * ratio / (1 + ratio)

2. 创建HybridReqToTokenPool
   └─> new HybridReqToTokenPool(
       size=max_total_tokens,
       mamba_size=max_mamba_cache_size,
       mamba_spec_state_size=speculative_num_draft_tokens,
       ...
   )

3. 初始化MambaPool
   └─> _init_mamba_pool()
       ├─> 创建MambaPool实例
       ├─> 分配conv_state
       ├─> 分配temporal_state
       └─> 分配intermediate状态 (投机采样)

4. 初始化映射表
   ├─> req_index_to_mamba_index_mapping
   └─> req_index_to_mamba_ping_pong_track_buffer_mapping (extra_buffer)

4.2 请求分配流程

请求到达
    │
    ├─> 检查是否有radix cache
    │    └─> 有: req.mamba_pool_idx 已设置
    │    └─> 无: req.mamba_pool_idx = None
    │
    ├─> 调用 pool.alloc(need_size, reqs)
    │
    ├─> 分配请求slot (基类)
    │    └─> select_index = super().alloc(need_size)
    │
    ├─> 为每个请求分配mamba slot
    │    ├─> 如果 req.mamba_pool_idx is None
    │    │    └─> mid = mamba_pool.alloc(1)
    │    │         └─> 从free_slots取出一个slot
    │    │         └─> 清零slot
    │    │    └─> req.mamba_pool_idx = mid
    │    └─> 否则复用已有slot
    │
    ├─> 如果启用extra_buffer
    │    └─> 如果 req.mamba_ping_pong_track_buffer is None
    │         └─> 分配2个slot: [slot_ping, slot_pong]
    │         └─> req.mamba_next_track_idx = 0
    │
    └─> 更新映射表
         ├─> req_index_to_mamba_index_mapping[req_idx] = mamba_slot
         └─> req_index_to_mamba_ping_pong_...[req_idx] = [slot1, slot2]

4.3 请求释放流程

请求完成
    │
    ├─> 调用 pool.free(req_idx, free_mamba_cache, keep_ping_pong)
    │
    ├─> 释放请求slot (基类)
    │    └─> super().free(req_idx)
    │
    ├─> 如果 free_mamba_cache = True
    │    ├─> 获取mamba slot索引
    │    └─> mamba_pool.free(mamba_slot)
    │         └─> 将slot加回free_slots
    │
    └─> 如果启用extra_buffer
         ├─> 如果 keep_ping_pong is None
         │    └─> 释放所有ping-pong slot
         └─> 否则
              └─> 只释放另一个slot,保留keep_ping_pong指定的slot
Join the discussion Powered by GitHub Discussions