This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3

NCC_EHCA005

Error message: The compiler encountered a custom call instruction with a target name that is not recognized.

The Neuron compiler currently recognizes the following custom call targets:

• AwsNeuronErf

• AwsNeuronGelu

• AwsNeuronGeluApprxTanh

• AwsNeuronGeluBackward

• AwsNeuronSilu

• AwsNeuronSiluBackward

• AwsNeuronRmsNorm

• AwsNeuronSoftmax

• AwsNeuronSoftmaxBackward

• AwsNeuronCollectiveMatmul

• AwsNeuronIntMatmult

• AwsNeuronArgMax

• AwsNeuronArgMin

• AwsNeuronTopK

• AwsNeuronDropoutMaskV1

• AwsNeuronCustomNativeKernel

• AwsNeuronCustomOp

• AwsNeuronDevicePrint

• ResizeNearest

• ResizeBilinear

• ResizeNearestGrad

• AwsNeuronLNCShardingConstraint

• AwsNeuronTransferWithStaticRing

• AwsNeuronModuleMarkerStart-Forward

• AwsNeuronModuleMarkerStart-Backward

• AwsNeuronModuleMarkerEnd-Forward

• AwsNeuronModuleMarkerEnd-Backward

• NeuronBoundaryMarker-Start

• NeuronBoundaryMarker-End

Erroneous code example:

def lowering(ctx, x_val):
    result_type = ir.RankedTensorType(x_val.type)
    # This target name will not be recognized by HandleCustomCall
    return hlo.CustomCallOp(
        [result_type],
        [x_val],
        call_target_name="UNRECOGNIZED_TARGET",
        has_side_effect=ir.BoolAttr.get(False),
    ).results

Use a supported custom call target:

def lowering(ctx, x_val):
    result_type = ir.RankedTensorType(x_val.type)
    return hlo.CustomCallOp(
        [result_type],
        [x_val],
        call_target_name="AwsNeuronSilu",
        has_side_effect=ir.BoolAttr.get(False),
        backend_config=ir.StringAttr.get(""),
        api_version=ir.IntegerAttr.get(ir.IntegerType.get_signless(32), 2),
    ).results

This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3