CPU Operators Supported by VART ML Runtime#
When a model is compiled with CPU partition support, NPU-incompatible operators are routed to CPU implementations provided by the AMD Vitis™ AI compiler. The following operators are supported by VART ML Runtime for CPU execution. Each listed operator is supported on AArch64 (ARM).
If a model contains NPU-incompatible operators that are not in this list, it cannot be compiled with CPU partition support and must use standard compilation with ONNX Runtime instead.
Operator |
AArch64 (ARM) |
|---|---|
Add |
Yes |
And |
Yes |
ArgMax |
Yes |
Atan |
Yes |
Cast |
Yes |
Clip |
Yes |
Concat |
Yes |
Conv |
Yes |
Cos |
Yes |
CumSum |
Yes |
DequantizeLinear |
Yes |
Div |
Yes |
Einsum |
Yes |
Equal |
Yes |
Erf |
Yes |
Expand |
Yes |
Flatten |
Yes |
Gather |
Yes |
GatherBlockQuantized |
Yes |
GatherElements |
Yes |
GatherND |
Yes |
Gemm |
Yes |
Gelu |
Yes |
Greater |
Yes |
GreaterOrEqual |
Yes |
GridSample |
Yes |
GroupQueryAttention |
Yes |
InstanceNormalization |
Yes |
Inverse |
Yes |
IsNaN |
Yes |
LayerNormalization |
Yes |
Less |
Yes |
Log |
Yes |
MatMul |
Yes |
MatMulNBits |
Yes |
Max |
Yes |
MaxPool |
Yes |
Mod |
Yes |
Mul |
Yes |
NonMaxSuppression |
Yes |
Neg |
Yes |
Not |
Yes |
OneHot |
Yes |
Pow |
Yes |
QuantizeLinear |
Yes |
ReduceMax |
Yes |
ReduceMean |
Yes |
ReduceL2 |
Yes |
Relu |
Yes |
Reshape |
Yes |
Resize |
Yes |
RMSNormalization |
Yes |
ScatterElements |
Yes |
ScatterND |
Yes |
Shape |
Yes |
Sigmoid |
Yes |
Sign |
Yes |
SimplifiedLayerNormalization |
Yes |
Sin |
Yes |
SkipSimplifiedLayerNormalization |
Yes |
Slice |
Yes |
Softmax |
Yes |
Split |
Yes |
Sqrt |
Yes |
Squeeze |
Yes |
Sub |
Yes |
Tile |
Yes |
TopK |
Yes |
Transpose |
Yes |
Unsqueeze |
Yes |
Where |
Yes |
Inverse uses the ONNX contrib schema com.microsoft.Inverse.
