Struct SessionBuilder 

pub struct SessionBuilder { /* private fields */ }

Builder for advanced session configuration (§20.6).

Implementations

impl SessionBuilder

pub fn new() -> Self

pub fn model(self, path: impl AsRef<Path>) -> Self

pub fn model_bytes(self, bytes: &[u8]) -> Self

pub fn device(self, pref: DevicePreference) -> Self

pub fn execution_provider( self, execution_provider: Arc<dyn ExecutionProvider>, ) -> Self

Use an explicitly constructed execution provider instead of device auto-selection.

pub fn memory_limit(self, bytes: usize) -> Self

pub fn profiling(self, enable: bool) -> Self

pub fn warmup(self, shapes: Vec<WarmupShape>) -> Self

pub fn decode_precision(self, precision: DecodePrecision) -> Self

Select the decoder-wide numeric precision (see DecodePrecision). Defaults to DecodePrecision::Model (graph as authored); selecting DecodePrecision::Fp16 opts into the fp32→fp16 decode rewrite, which only takes effect on a GPU fp32-activation quantized decoder.

pub fn option(self, key: &str, value: &str) -> Self

Set a namespaced option. Unknown keys — and unknown values for a known key — are rejected at Self::build.

Recognized options

Key	Values	Default	Effect
"optimization"	"none", "basic", "all"	"none"	Graph optimization level (see OptimizationLevel).
"ep.context_enable"	"0"/"1"/"false"/"true"	"0"	Dump a *_ctx.onnx EPContext model after compile (§21.4 / §55.4).
"ep.context_file_path"	any path	<orig>_ctx.onnx	Output path for the generated context model.
"ep.context_embed_mode"	"0" (external) / "1" (embed)	"1"	How the compiled blob is stored in each EPContext node.

"optimization" = "none" (the default) leaves the loaded graph untouched, so behavior is byte-identical to a runtime with no optimizer. "basic" runs constant folding + dead-node elimination; "all" adds operator fusion. When any pass runs, the session re-runs shape inference on the rewritten graph before compiling so fused/introduced nodes get inferred shapes.

pub fn build(self) -> Result<InferenceSession, SessionError>

Build the session: load → detect device → optimize → compile → allocate.

The optimize stage is driven by the "optimization" session option and defaults to OptimizationLevel::None (a no-op), so the default path is byte-identical to loading straight into the executor. When optimization is enabled the pipeline is:

load (+ loader shape inference)
  → run optimizer passes (constant-fold / DCE / fusion)
  → re-run shape inference on the rewritten graph
  → compile (kernel per node) → allocate

The re-inference step is essential: fusion can replace a multi-op decomposition (e.g. the 9-op LayerNorm) with a single fused node whose output has no inferred shape yet, and the compile/execute stages require every value to carry a resolved shape.

Device selection keeps CPU as the default and selects CUDA only when explicitly requested in a CUDA-enabled build. “Compile” resolves a kernel per node into the shape-keyed cache.

Trait Implementations

impl Default for SessionBuilder

fn default() -> SessionBuilder

Returns the “default value” for a type.