Skip to content

vllm.v1.simple_kv_offload

Modules:

  • copy_backend –

    DMA copy backend for GPU<->CPU block transfers.

  • cuda_mem_ops –

    Low-level CUDA/HIP memory helpers: pinning and batch DMA transfers.

  • disk_backend –

    Disk I/O backend for GPU<->NVMe block transfers via pinned staging buffers.

  • manager –

    Scheduler-side manager for SimpleCPUOffloadConnector.

  • metadata –

    Metadata for SimpleCPUOffloadConnector.

  • metrics –

    Prometheus metrics for SimpleCPUOffloadConnector.

  • worker –

    Worker-side handler for SimpleCPUOffloadConnector.