Struct GpuSpec 

pub struct GpuSpec {
    pub count: u32,
    pub vendor: String,
    pub mode: Option<String>,
    pub model: Option<String>,
    pub scheduling: Option<SchedulingPolicy>,
    pub distributed: Option<DistributedConfig>,
    pub sharing: Option<GpuSharingMode>,
    pub mps_pipe_dir: Option<String>,
    pub mps_log_dir: Option<String>,
    pub time_slice_index: Option<u32>,
    pub time_slicing_config_path: Option<String>,
}

GPU resource specification

Supported vendors:

• nvidia - NVIDIA GPUs via NVIDIA Container Toolkit (default)
• amd - AMD GPUs via ROCm (/dev/kfd + /dev/dri/renderD*)
• intel - Intel GPUs via VAAPI/i915 (/dev/dri/renderD*)
• apple - Apple Silicon GPUs via Metal/MPS (macOS only)

Unknown vendors fall back to DRI render node passthrough.

GPU mode (macOS only)

When vendor is "apple", the mode field controls how GPU access is provided:

• "native" – Seatbelt sandbox with direct Metal/MPS access (lowest overhead)
• "vm" – libkrun micro-VM with GPU forwarding (stronger isolation)
• None (default) – Auto-select based on platform and vendor

On Linux, mode is ignored; GPU passthrough always uses device node binding.

Fields

count: u32

Number of GPUs to request

vendor: String

GPU vendor (nvidia, amd, intel, apple) - defaults to nvidia

mode: Option<String>

GPU access mode (macOS only): "native", "vm", or None for auto-select

model: Option<String>

Pin to a specific GPU model (e.g. “A100”, “H100”). Substring match against detected GPU model names.

scheduling: Option<SchedulingPolicy>

Scheduling policy for GPU workloads.

• best-effort (default): place what fits
• gang: all-or-nothing for distributed jobs
• spread: distribute across nodes

distributed: Option<DistributedConfig>

Distributed GPU job coordination. When set, injects MASTER_ADDR, WORLD_SIZE, RANK, LOCAL_RANK env vars.

sharing: Option<GpuSharingMode>

GPU sharing mode: exclusive (default), mps, or time-slice.

mps_pipe_dir: Option<String>

Host directory for the NVIDIA MPS control pipe.

Only consulted when sharing == Mps. Defaults to /tmp/nvidia-mps when unset. The directory MUST exist on the host (created by the nvidia-cuda-mps-control daemon). It is bind-mounted into the container at the same path and exported as CUDA_MPS_PIPE_DIRECTORY.

mps_log_dir: Option<String>

Host directory for NVIDIA MPS log output.

Only consulted when sharing == Mps. Defaults to /tmp/nvidia-log when unset. The directory MUST exist on the host. It is bind-mounted into the container and exported as CUDA_MPS_LOG_DIRECTORY.

time_slice_index: Option<u32>

CUDA device index this replica should see when sharing == TimeSlice.

Emitted as CUDA_VISIBLE_DEVICES=<slice_index>, overriding the default 0..count visibility list. Use this together with a host-side NVIDIA time-slicing config to advertise a single physical GPU as multiple virtual slices.

time_slicing_config_path: Option<String>

Optional host path to a NVIDIA time-slicing config YAML.

When set, the file is bind-mounted read-only at /etc/nvidia/gpu-time-slicing.yaml inside the container so tools that inspect the slicing topology (e.g. monitoring sidecars) can read it. The file is not interpreted by ZLayer — it’s purely informational for the workload.

Trait Implementations

impl Clone for GpuSpec

fn clone(&self) -> GpuSpec

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for GpuSpec

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for GpuSpec

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Eq for GpuSpec

impl PartialEq for GpuSpec

fn eq(&self, other: &GpuSpec) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Serialize for GpuSpec

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for GpuSpec

impl Validate for GpuSpec

fn validate(&self) -> Result<(), ValidationErrors>

impl<'v_a> ValidateArgs<'v_a> for GpuSpec

type Args = ()

fn validate_with_args(&self, args: Self::Args) -> Result<(), ValidationErrors>