Varlink IPC Specifications

Hardware compute discovery, multi-accelerator lease arbitration, PSI telemetry, and streaming model inference.

interface io.syntrop.Inference1

type ComputePlane (
  id: string,
  name: string,
  kind: string,
  total_memory: int,
  available_memory: int,
  is_triage_reserved: bool
)

type LeaseInfo (
  id: string,
  plane_id: string,
  allocated_memory: int,
  priority: string,
  state: string,
  client_unit: ?string,
  client_pid: ?int
)

type PlaneSlice (
  plane_id: string,
  role: string,
  stage_index: ?int,
  memory_bytes: int
)

type CompositeLease (
  lease_id: string,
  gang_id: string,
  gang_policy: string,
  slices: []PlaneSlice
)

type ModelInfo (
  id: string,
  format: string,
  path: string,
  estimated_memory: int,
  placement: string
)

method GetTopology() -> (
  planes: []ComputePlane,
  total_ram: int,
  available_ram: int,
  cpu_cores: int
)

method GetPressure() -> (
  level: string,
  cpu_some: float,
  memory_some: float,
  io_some: float
)

method AcquireLease(
  priority: string,
  memory_bytes: int,
  plane: ?string,
  unit: ?string,
  pid: ?int
) -> (
  lease_id: string,
  plane_id: string,
  allocated_memory: int
)

method AcquireCompositeLease(
  gang_id: string,
  slices: []PlaneSlice,
  gang_policy: string,
  priority: string,
  client_unit: ?string
) -> (
  lease_id: string,
  allocated_slices: []PlaneSlice
)

method ReleaseLease(lease_id: string) -> ()
method Yield(lease_id: string) -> ()
method Freeze(lease_id: string) -> ()
method Thaw(lease_id: string) -> ()
method ListLeases() -> (leases: []LeaseInfo)
method ListCompositeLeases() -> (leases: []CompositeLease)
method ListModels() -> (models: []ModelInfo)

method RegisterModel(
  id: string,
  format: string,
  path: string,
  estimated_bytes: int
) -> ()

method StreamInference(
  model: string,
  prompt: string
) -> (
  chunk: string
)

error MethodNotFound (method: string)
error InvalidParameter (parameter: string)
error ResourceExhaustion (plane: string, requested: int, available: int)
error PlaneNotFound (plane: string)
error LeaseNotFound (lease_id: string)
error ModelNotFound (model: string)

Command-Line Invocation

varlinkctl call unix:/run/syntrop/io.syntrop.Inference1 io.syntrop.Inference1.GetTopology '{}'

Sample Reply Payload

{
  "planes": [
    {"available_memory": 15998242816, "id": "drm-renderD128", "is_triage_reserved": false, "kind": "drm_gpu", "name": "NVIDIA GeForce RTX 4090", "total_memory": 25769803776}
  ],
  "cpu_cores": 32,
  "total_ram": 67108864000,
  "available_ram": 48210948000
}