{
  "id": "okf_xydong0223_infer_forge_harness_openwiki__4474e617",
  "kind": "okf",
  "category": "dados",
  "name": "vLLM 硬件后端接入体系",
  "tagline": "Bundle OKF 0.2 · 0 conceitos · xyDong0223/infer-forge-harness",
  "body": "# vLLM 硬件后端接入体系\n\n> [返回 OpenWiki 根目录](../index.md) · [查看 vLLM-Kunlun 的下游实现](../vllm-kunlun/index.md)\n\n本 wiki 聚焦 vLLM 的 **Platform / 硬件后端接入体系**：一个非 CUDA 硬件（ROCm、XPU、CPU、TPU、昆仑芯 XPU 等）要把 vLLM 跑起来，需要挂进哪些抽象、实现哪些类、注册哪些 entry point。\n\n**证据基线**：vLLM main 分支 commit `94848eda600a07c28675f5753a11b2c212c146ed`（2026-09-09）。\n注意：本版本 vLLM 只有 **V1 引擎**，且相对老版本有大幅重构——`vllm/attention/`、`PlatformRegistry`、`get_worker_class`、`VLLM_ATTENTION_BACKEND` 均已不存在，读旧文档时需对照本 wiki 的勘误。\n\n## 总体架构\n\n```mermaid\nflowchart LR\nCLI[\"Engine / CLI 参数解析\"] --> CFG[\"VllmConfig\"]\nCFG --> P[\"current_platform (Platform 子类)\"]\nP -->|\"check_and_update_config: 设置 worker_cls\"| EXE[\"Executor.get_class\"]\nEXE -->|\"collective_rpc\"| W[\"Worker (WorkerBase 子类)\"]\nW --> MR[\"ModelRunner\"]\nMR --> ATTN[\"AttentionBackend 三件套<br/>(Backend / Builder / Impl)\"]\nMR --> OPS[\"CustomOp / torch.ops 分发\"]\nMR --> Q[\"量化层 (QuantizationConfig)\"]\nW --> DIST[\"GroupCoordinator → DeviceCommunicator\"]\n```\n\n一个硬件后端的\"落点\"只有五个：**平台类**（[platform-abstraction.md](platform/platform",
  "url": "https://github.com/xyDong0223/infer-forge-harness/blob/HEAD/openwiki/vllm-core/index.md",
  "status": "live",
  "origin": "okf_github",
  "origin_id": "xyDong0223/infer-forge-harness:openwiki/vllm-core/index.md",
  "install": "https://raw.githubusercontent.com/xyDong0223/infer-forge-harness/HEAD/openwiki/vllm-core/index.md",
  "source": "https://github.com/xyDong0223/infer-forge-harness",
  "transporte": "",
  "ns": "xydong0223",
  "versao": "0.2",
  "oficial_status": "",
  "repo_host": "github.com",
  "topico": "devtools",
  "linguagem": "Python",
  "licenca": "",
  "repo_topics": "",
  "repo_slug": "xydong0223/infer-forge-harness",
  "stars": 0,
  "forks": 0,
  "prs_abertos": 0,
  "pushed_at": "2026-09-11T14:28:42Z",
  "repo_estado": "ativo",
  "likes": 0,
  "comments": 0,
  "visits": 0,
  "readme_bytes": 15689,
  "readme_api": "https://agentalog.com/api/listings/okf_xydong0223_infer_forge_harness_openwiki__4474e617/readme",
  "readme_fonte": "repo",
  "created_at": "2026-09-11 13:05:33",
  "updated_at": "2026-09-11 13:05:33",
  "mine": false,
  "api": "https://agentalog.com/api/listings/okf_xydong0223_infer_forge_harness_openwiki__4474e617",
  "go": "https://agentalog.com/api/go/okf_xydong0223_infer_forge_harness_openwiki__4474e617",
  "comments_api": "https://agentalog.com/api/listings/okf_xydong0223_infer_forge_harness_openwiki__4474e617/comments",
  "readme_sha": "204a1467e7cfd03bcdad5f5a38d0c51379c6b8fd1f831637bf1aad4df9ee2a3c",
  "readme_em": "2026-09-14 00:36:21",
  "payload": {
    "name": "xyDong0223/infer-forge-harness:openwiki/vllm-core/index.md",
    "title": "vLLM 硬件后端接入体系",
    "version": "0.2",
    "okf_bundle_dir": "openwiki/vllm-core",
    "okf_index_path": "openwiki/vllm-core/index.md",
    "okf_conceitos": 0,
    "okf_ref": "HEAD",
    "okf_type": "index"
  },
  "volatil_em": "2026-09-13 20:02:29"
}