aws-cdk-neuronx-patterns/lib/base/neuronx-compiler/neuronx-cross-compiler.d.ts

> [!WARNING] > This library is experimental module.

import * as batch from "aws-cdk-lib/aws-batch";
import * as ec2 from "aws-cdk-lib/aws-ec2";
import { Construct } from "constructs";
import { NeuronxCompilerBase, NeuronxCompilerBaseProps, ComputeEnvironmentResult } from "./neuronx-compiler-base";
/**
 * Props of NeuronxCrossCompiler.
 */
export interface NeuronxCrossCompilerProps extends NeuronxCompilerBaseProps {
    /**
     * The EC2 instance type to use for cross-compilation.
     * This should be a non-Neuron instance type with sufficient memory and CPU
     * for model compilation.
     *
     * @default ec2.InstanceType.of(ec2.InstanceClass.C7I, ec2.InstanceSize.XLARGE4)
     */
    readonly compileInstanceType?: ec2.InstanceType;
}
/**
 * Neuronx cross-compiler construct.
 * Compile the model on a non-Neuron instance and upload the artifacts to an S3 bucket.
 * This avoids the need for expensive Neuron instances during the compilation phase.
 *
 * The compilation uses `vllm serve` which performs model tracing and neuronx-cc compilation
 * entirely on CPU. The resulting artifacts are compatible with Neuron instances for inference.
 */
export declare class NeuronxCrossCompiler extends NeuronxCompilerBase {
    constructor(scope: Construct, id: string, props: NeuronxCrossCompilerProps);
    protected createComputeEnvironment(props: NeuronxCompilerBaseProps): ComputeEnvironmentResult;
    protected createJobDefinition(props: NeuronxCompilerBaseProps): batch.IJobDefinition;
}