# vllm-docs

**Repository Path**: omnihorizon/vllm-docs

## Basic Information

- **Project Name**: vllm-docs
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-07-13
- **Last Updated**: 2026-07-20

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# vLLM 特性文档

本目录收录 vLLM 关键特性的中文说明文档，便于本地查阅与维护。

## 文档索引

| 文档 | 说明 |
|------|------|
| [gpu-request-processing-flow.md](./gpu-request-processing-flow.md) | GPU 场景：HTTP 收请求 → 调度 → 前向/采样 → 回包；含 TP 多卡、async scheduling、KV offload hook 与函数调用链 |
| [gpu-dma-ddr-features.md](./gpu-dma-ddr-features.md) | 与 GPU DMA 读写 DDR 相关的特性、数据路径、配置与源码索引 |
| [kv-offload-batch-memcpy.md](./kv-offload-batch-memcpy.md) | KV offloading 批量 DMA API（`cuMemcpyBatchAsync` 等）与 `cudaMemcpyAsync` 的区别 |
| [capability-evaluation.md](./capability-evaluation.md) | 接下来可开展的能力评估计划（含 `cuMemcpyBatchAsync` 与 ANY 式预取） |
| [pcie-topology-evaluation.md](./pcie-topology-evaluation.md) | 单机四卡 L20：直连 RC vs PCIe Switch；五维评估（p2pBandwidthLatencyTest / nvbandwidth / H2D·D2H / NCCL / vLLM） |

## 上游官方文档

以下路径相对于 vLLM 源码仓库根目录（`/home/nathan/vllm-upstream`）：

- `docs/features/kv_offloading_usage.md` — KV Offloading 配置指南
- `docs/features/disagg_prefill.md` — Disaggregated Prefill 概览
- `docs/features/nixl_connector_usage.md` — NixlConnector 用法
- `docs/features/mooncake_connector_usage.md` — MooncakeConnector 用法
- `docs/features/mooncake_store_connector_usage.md` — MooncakeStore KV 池
- `docs/features/disagg_encoder.md` — Disaggregated Encoder
- `docs/serving/parallelism_scaling.md` — GPUDirect RDMA 配置
- `docs/models/extensions/instanttensor.md` — InstantTensor / GDS 权重加载