Skip to contents

Serialize a cuML model

Source: R/model.R
cuda_ml_serialize.Rd

Given a cuML model, serialize its state into a connection.

Usage

cuda_ml_serialize(model, connection = NULL, ...)

Arguments

model

The model object.

connection

An open connection or NULL. If NULL, then the model state is serialized to a raw vector. Default: NULL.

...

Additional arguments to base::serialize().

Value

NULL unless connection is NULL, in which case the serialized model state is returned as a raw vector.

Persistence contract

cuda.ml schema 1 model states contain a schema number, the cuda.ml package version that created the state, backend provenance, a model ABI identifier, and a payload. The package version is provenance only: a difference from the installed cuda.ml version does not prevent restoration.

Compatibility is determined before the payload is restored:

  • The schema must be the integer 1. Unknown and unversioned schemas are rejected.

  • The state class and model ABI must identify a restoration method supported by the installed package. A change to a model's payload layout requires a new model ABI.

  • Linear-model and logistic-regression states have portable R payloads and do not require matching backend identity fields.

  • PCA, SVC, one-vs-rest SVC, SVR, and UMAP states require an exact rapids_version match because their payloads reconstruct RAPIDS native state.

  • Random-forest and nvForest states require an exact treelite_version match because their payloads contain serialized Treelite model bytes.

The remaining recorded backend fields—cuda_version, nvforest_version, and platform—are provenance for schema 1, not compatibility gates. A missing payload, unsupported ABI, or missing or unequal required backend field is rejected. cuda.ml does not implicitly migrate a state or fall back to serializing native pointers.

Current nvForest and random-forest states store device-neutral Treelite model bytes. They retain model semantics and prediction precision, but not the inference device, device identifier, tree layout, chunk size, or memory alignment. Select those settings when restoring with cuda_ml_unserialize(); GPU is the default. Legacy v1 nvForest and random-forest states remain supported and restore with the inference settings recorded in their payloads.

Saving a state to a file connection and restoring it in another R process uses this same contract. The target process must have a compatible cuda.ml installation and must prepare the corresponding backend before prediction: cuda_ml_install() for GPU operation or cuda_ml_install(device = "cpu") for CPU-only nvForest inference. bundle::bundle() stores the same explicit state, so saving a bundle with saveRDS() and restoring it with readRDS() and bundle::unbundle() has the same compatibility requirements. For an nvForest-backed model, the bundle also stores its chosen deployment device separately from the device-neutral state. A bundle is not required for deployment; cuda_ml_serialize() returns the complete state artifact directly.

See also

serialize