Skip to content

vllm.model_executor.model_loader.weight_cache.ipc_backends

Local weight-sharing backends, independent of D2D copy movers.

Modules:

  • base –

    The local weight-sharing backend contract.

  • serial_utils –

    Pickle framing and descriptor attachments for trusted local IPC sockets.

  • torch_ipc –

    The original PyTorch CUDA/ROCm IPC weight-sharing backend.

  • vmm_ipc –

    CUDA/ROCm VMM storage shared by POSIX descriptors, with owned mappings.