Class CoreMlProviderOptions

CoreML execution-provider options, as ORT's string-keyed provider dictionary spells them.

public sealed class CoreMlProviderOptions

Inherited Members: object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.ReferenceEquals(object, object)

object.ToString()

Properties

EnableModelCache: Emits ModelCacheDirectory. Unset, CoreML recompiles the captured subgraph on every session creation and leaks the artefact into the iOS tmp directory. The cache key is made content-addressed by the path itself: <OrtCache>/<modelId>/<sha16>.

MLComputeUnits: Which hardware CoreML may use. Default CPUAndNeuralEngine.

ModelFormat: MLProgram, not ORT's NeuralNetwork default: NeuralNetwork's supported-op table lacks LayerNormalization, Gelu and Erf, so a BERT encoder fragments into CPU-fallback partitions and pays a CPU-to-ANE round trip per block.

ProfileComputePlan: Diagnostics only: logs the hardware each operator was dispatched to.

RequireStaticInputShapes: Emits RequireStaticInputShapes=1. Default false, and that is load-bearing. CoreML partitions the graph at SESSION-CREATION time from the graph's DECLARED shapes, not from the shapes fed at Run time. All four presets declare input_ids/attention_mask/token_type_ids as [batch_size, sequence_length] - symbolic. Setting this to 1 against symbolic dims means CoreML takes few or no nodes and the whole graph silently falls back to CPU, which Accepted cannot detect. Turning it on is therefore only correct together with FreeDimensionOverrides (or OnnxEmbeddingOptions.PinnedSequenceLength, which sets both), and the session factory refuses the combination without them.