Enum P2pMsg 

pub enum P2pMsg {
    EnableAll {
        reply: Sender<Result<P2pGraph, GpuError>>,
    },
    CanAccess {
        from: u32,
        to: u32,
        reply: Sender<bool>,
    },
    CopyF32 {
        src: GpuRef<f32>,
        src_device: u32,
        dst: GpuRef<f32>,
        dst_device: u32,
        reply: Sender<Result<(), GpuError>>,
    },
    Topology {
        reply: Sender<P2pGraph>,
    },
    RefreshDevice {
        device_idx: u32,
        new_generation: u64,
    },
}

Variants

EnableAll

Fields

reply: Sender<Result<P2pGraph, GpuError>>

CanAccess

Fields

from: u32

to: u32

reply: Sender<bool>

CopyF32

Async peer copy from src (on src_device) to dst (on dst_device). Both GpuRefs must be valid; copy size is min(src.len, dst.len) × sizeof(f32).

Fields

src: GpuRef<f32>

src_device: u32

dst: GpuRef<f32>

dst_device: u32

reply: Sender<Result<(), GpuError>>

Topology

Fields

reply: Sender<P2pGraph>

RefreshDevice

Internal: a device’s generation_watch fired, meaning its ContextActor rebuilt the underlying CudaContext. The cached Arc<CudaContext> in self.contexts is now stale — re-snapshot it. Peer-access mappings persist across rebuilds only if the new context lives in the same primary slot, so the topology marks itself disabled and a follow-up EnableAll is required to refresh graph + re-issue cuCtxEnablePeerAccess.

Fields

device_idx: u32

new_generation: u64