混合缓存执行流程图

Create a clean vertical technical workflow infographic on a light gray background, using a minimalist modern product-diagram style with rounded white cards, thin colored outlines, simple vector line icons, dark navy text, and navy connector arrows. The composition is a single centered top-to-bottom flowchart with 7 numbered main steps, plus 2 parallel cache-group panels branching from step 4 into step 5, and a thick dark return arrow on the far left looping from the bottom back to the top. Use crisp sans-serif typography, generous spacing, subtle pastel accent colors, no gradients, no shadows, and presentation-slide clarity. At the top center place step card 1 with a blue outline and a code/chat icon on the left. Title text: "1. {argument name="step 1 title" default="chat completions request"}". Subtitle beneath: "conversation_id + cache_salt + new suffix messages". Below it place step card 2 with a blue outline and a document/list icon. Title: "2. Frontend conversation ledger". Subtitle: "lease same id + track committed messages". Below it place step card 3 with a cyan outline and a database-with-magnifier icon. Title: "3. Exact conversation cache lookup". Subtitle: "conversation_id  committed turn state". Below it place step card 4 with a purple outline and a branching scheduler icon. Title: "4. Scheduler cache attachment". Subtitle: "set num_computed_tokens + attach committed state". From step 4, branch downward into 2 side-by-side group panels. Left group panel: a pale green rounded container titled "Full-attention KV cache group". Inside it, stack 2 inner cards. First inner card has a green block-grid icon, title "Committed block refs", subtitle "share aligned full KV blocks". Second inner card below has a green layered-sheets icon, title "Tail COW copy", subtitle "copy unaligned KV tail". At the bottom of the green panel add small footer text: "paged K/V tensors for transformer layers". Right group panel: a pale purple rounded container titled "Mamba terminal-state cache group". Inside it, stack 2 inner cards. First inner card has a purple database/network icon, title "Committed terminal state", subtitle "exact state at committed length". Second inner card below has a purple wavy-lines icon, title "Request-owned terminal copy", subtitle "copy SSM + conv state". At the bottom of the purple panel add small footer text: "align-mode terminal state placement". Merge both group-panel outputs into a centered step card 5 with a blue outline and a microchip icon. Title: "5. Hybrid model execution". Subtitle: "run only the uncached suffix". Inside the bottom area of this card, include 2 pill-shaped labels side by side: "Transformer layers" and "Mamba layers". Below it place step card 6 with a blue outline and a sparkle icon. Title: "6. Decode assistant tokens". Subtitle: "stream response token by token". Below it place step card 7 with a warm yellow-orange outline and a database-with-check icon. Title: "7. Commit completed turn". Subtitle: "publish pending state or discard on failure". Add a thick dark navy loop arrow running down the far left side, entering step 1 near the top from the left and returning from step 7 at the bottom back upward. Along this left loop, near the lower half, place stacked annotation text: "next request reuses committed conversation head". Add 2 dashed publish arrows rising upward from step 7 toward the cache-group panels: one green dashed arrow on the left pointing to the green cache panel, labeled "publish new state"; one purple dashed arrow on the right pointing to the purple cache panel, also labeled "publish new state". Keep the exact total count of 7 numbered main cards, 2 cache group panels, 4 inner cache cards, and 2 pill labels. Preserve a portrait aspect ratio similar to a conference-slide architecture diagram.

创建一个背景为浅灰色的简洁垂直技术工作流信息图，采用极简现代产品图表风格，使用圆角白色卡片、细彩色轮廓、简单的矢量线条图标、深蓝色文本和深蓝色连接箭头。整体构图为居中的自上而下流程图，包含 7 个带编号的主要步骤，以及从第 4 步分支到第 5 步的 2 个并行缓存组面板，最左侧有一个从底部循环回顶部的深色粗返回箭头。使用清晰的无衬线字体，保持充裕的间距，使用柔和的淡色调，无渐变，无阴影，确保演示 Slides 的清晰度。 在顶部中心放置第 1 步卡片，带有蓝色轮廓和左侧的代码/聊天图标。标题文本：“1. {argument name="step 1 title" default="聊天补全请求"}”。副标题：“conversation_id + cache_salt + 新后缀消息”。 下方放置第 2 步卡片，带有蓝色轮廓和文档/列表图标。标题：“2. 前端对话账本”。副标题：“租用相同 ID + 追踪已提交消息”。 下方放置第 3 步卡片，带有青色轮廓和带放大镜的数据库图标。标题：“3. 精确对话缓存查找”。副标题：“conversation_id  已提交轮次状态”。 下方放置第 4 步卡片，带有紫色轮廓和分支调度器图标。标题：“4. 调度器缓存附加”。副标题：“设置 num_computed_tokens + 附加已提交状态”。 从第 4 步向下分支为 2 个并排的组面板。 左侧组面板：一个淡绿色圆角容器，标题为“Full-attention KV 缓存组”。内部堆叠 2 张内部卡片。第一张内部卡片带有绿色块网格图标，标题“已提交块引用”，副标题“共享对齐的完整 KV 块”。下方第二张内部卡片带有绿色分层纸张图标，标题“尾部 COW 拷贝”，副标题“拷贝未对齐的 KV 尾部”。在绿色面板底部添加小字页脚：“用于 Transformer 层的分页 K/V 张量”。 右侧组面板：一个淡紫色圆角容器，标题为“Mamba 终端状态缓存组”。内部堆叠 2 张内部卡片。第一张内部卡片带有紫色数据库/网络图标，标题“已提交终端状态”，副标题“已提交长度下的精确状态”。下方第二张内部卡片带有紫色波浪线图标，标题“请求拥有的终端拷贝”，副标题“拷贝 SSM + 对话状态”。在紫色面板底部添加小字页脚：“对齐模式终端状态放置”。 将两个组面板的输出合并到居中的第 5 步卡片中，带有蓝色轮廓和微芯片图标。标题：“5. 混合模型执行”。副标题：“仅运行未缓存的后缀”。在此卡片底部区域内，并排包含 2 个胶囊状标签：“Transformer 层”和“Mamba 层”。 下方放置第 6 步卡片，带有蓝色轮廓和闪光图标。标题：“6. 解码助手 Token”。副标题：“逐个 Token 流式传输响应”。 下方放置第 7 步卡片，带有暖黄橙色轮廓和带勾选的数据库图标。标题：“7. 提交完成的轮次”。副标题：“发布挂起状态或在失败时丢弃”。 在最左侧添加一个深蓝色粗循环箭头，从左侧上方进入第 1 步，并从底部的第 7 步返回上方。沿此左侧循环，在下半部分附近放置堆叠的注释文本：“下一次请求复用已提交的对话头部”。 从第 7 步向缓存组面板方向添加 2 个向上的虚线发布箭头：左侧一个绿色虚线箭头指向绿色缓存面板，标注“发布新状态”；右侧一个紫色虚线箭头指向紫色缓存面板，同样标注“发布新状态”。 保持 7 张带编号的主卡片、2 个缓存组面板、4 张内部缓存卡片和 2 个胶囊标签的总数不变。保持类似于会议 Slides 架构图的纵向比例。