Supported LiteRT operators

When you import a model, each LiteRT (formerly TensorFlow Lite) operator in the model graph becomes one or more Sentis layers. For more information, refer to How Sentis optimizes a model.

Supported LiteRT operators

The following table shows the LiteRT operators that Sentis supports. It also outlines the data types that Sentis supports for each backend type.

Name	Sentis operators	Supported data types with BackendType.CPU	Supported data types with BackendType.GPUCompute	Supported data types with BackendType.GPUPixel	Notes
abs	Abs	float, int	float, int	float, int
add	Add	float, int	float, int	float, int
add_n	Add	float, int	float, int	float, int
arg_max	ArgMax	float, int	float, int	float, int	Dynamic dim input tensor is not supported
arg_min	ArgMin	float, int	float, int	float, int	Dynamic dim input tensor is not supported
atan2	Atan2	float	float	float
average_pool_2d	AveragePool	float	float	float
batch_matmul	MatMul	float	float*	float
bitwise_xor	BitwiseXor	int	int	int
broadcast_args	BroadcastArgs	float, int	float, int	float, int
broadcast_to	Expand	float, int	float, int	float, int
cast	Cast	float, int	float, int	float, int
ceil	Ceil	float	float	float
concatenation	Concat	float, int	float, int	float, int
conv_2d	Conv	float	float*	float
conv_3d	Conv	float	float*	float
conv_3d_transpose	ConvTranspose	float	float	float	Dynamic output_shape input tensor is not supported.
cos	Cos	float	float	float
cumsum	CumSum	float, int	float, int	float, int
depth_to_space	DepthToSpace	float, int	float, int	float, int
depthwise_conv_2d	Conv	float	float*	float
dequantize	Identity	float	float	float	int type is not supported
div	Div	float, int	float, int	float, int
elu	Elu	float	float	float
embedding_lookup	Gather	float, int	float, int	float, int
equal	Equal	float, int	float, int	float, int
exp	Exp	float	float	float
expand_dims	Unsqueeze	float, int	float, int	float, int
fill	Expand	float, int	float, int	float, int
floor	Floor	float	float	float
floor_div	FloorDiv	float, int	float, int	float, int
floor_mod	Mod	float	float	float
fully_connected	MatMul	float	float*	float
gather	Gather	float, int	float, int	float, int
gather_nd	GatherND	float, int	float, int	float, int
gelu	Gelu	float	float	float
greater	Greater	float, int	float, int	float, int
greater_equal	GreaterOrEqual	float, int	float, int	float, int
hard_swish	HardSwish	float	float	float
l2_normalization	ReduceSumSquare + Sqrt + Div	float	float*	float
l2_pool_2d	Square + AveragePool + Sqrt	float	float	float
leaky_relu	LeakyRelu	float	float	float
less	Less	float, int	float, int	float, int
less_equal	LessOrEqual	float, int	float, int	float, int
local_response_normalization	LRN	float	Not supported	Not supported
log	Log	float	float	float
log_softmax	LogSoftmax	float	float	float
logical_and	LogicalAnd	int	int	int
logical_not	LogicalNot	int	int	int
logical_or	LogicalOr	int	int	int
logistic	Sigmoid	float	float	float
max_pool_2d	MaxPool	float	float	float
maximum	Max	float, int	float, int	float, int
mean	ReduceMean	float	float	float	Reduction over an axis of length 0 is not supported. int type is not supported.
minimum	Min	float, int	float, int	float, int
mirror_pad	Pad	float, int	float, int	float, int
mul	Mul	float, int	float, int	float, int
multinomial	Multinomial	float	float	float	Dynamic num_samples input tensor is not supported
neg	Neg	float, int	float, int	float, int
not_equal	NotEqual	float, int	float, int	float, int
one_hot	OneHot	float, int	float, int	float, int
pack	Unsqueeze + Concat	float, int	float, int	float, int
pad	Pad	float, int	float, int	float, int
padv2	Pad	float, int	float, int	float, int
pow	Pow	float, int	float, int	float, int
prelu	PRelu	float	float	float
random_standard_normal	RandomNormal	float	float	float
random_uniform	RandomUniform	float	float	float
range	Range	float, int	float, int	float, int
rank	Shape + Size	-	-	-	The operator returns a CPU tensor without downloading the input tensor
reduce_all	ReduceMin	float, int	float*, int*	float, int
reduce_any	ReduceMax	float, int	float*, int*	float, int
reduce_max	ReduceMax	float, int	float*, int*	float, int
reduce_min	ReduceMin	float, int	float*, int*	float, int
reduce_prod	ReduceProd	float, int	float*, int*	float, int
relu	Relu	float	float	float
relu6	Relu6	float	float	float
relu_0_to_1	Clip	float	float	float
relu_n1_to_1	Clip	float	float	float
reshape	Reshape	float, int	float, int	float, int
resize_bilinear	Resize	float	float	float
resize_nearest_neighbor	Resize	float	float	float
reverse_v2	Slice	float, int	float, int	float, int
round	Round	float	float	float
rsqrt	Rsqrt	float	float	float
scatter_nd	ScatterND	float, int	float, int	float, int
select	Where	float, int	float, int	float, int
select_v2	Where	float, int	float, int	float, int
shape	Shape	-	-	-	The operator returns a CPU tensor without downloading the input tensor
sign	Sign	float, int	float, int	float, int
sin	Sin	float	float	float
slice	Slice	float, int	float, int	float, int	Dynamic begin and size input tensors are not supported
softmax	Softmax	float	float	float
space_to_depth	SpaceToDepth	float, int	float, int	float, int
sparse_to_dense	Expand + ScatterND	float, int	float, int	float, int	indices input tensor must have a rank greater than 0
split	Split	float, int	float, int	float, int	Dynamic split_dim input tensor is not supported
split_v	Split	float, int	float, int	float, int	Dynamic size_split and split_dim input tensors are not supported
sqrt	Sqrt	float	float	float
square	Square	float, int	float, int	float, int
squared_difference	Sub + Square	float, int	float, int	float, int
squeeze	Squeeze	float, int	float, int	float, int
strided_slice	Slice	float, int	float, int	float, int	The offset and ellipsis_mask attributes are not supported. Dynamic begin, end, and strides input tensors are not supported for masked slices.
sub	Sub	float, int	float, int	float, int
sum	ReduceSum	float, int	float, int	float, int
tanh	Tanh	float	float	float
tile	Tile	float, int	float, int	float, int
topk_v2	TopK	float, int	float, int	Not supported
transpose	Transpose	float, int	float, int	float, int	Dynamic perm input tensor is not supported
transpose_conv	ConvTranspose	float	float*	float	Supports 1D, 2D or 3D convolutions. Dynamic output_shape input tensor is not supported.
unpack	Select	float, int	float, int	float, int
where	NonZero + Transpose	float, int	float, int	float, int
zeros_like	Shape + ConstantOfShape	float, int	float, int	float, int

* Sentis uses DirectML to accelerate these operators on supported hardware.

Additional layers

LiteRT uses NHWC (batch size, height, width, channels) layout for image operations, such as convolution and pooling. Sentis uses NCHW (batch size, channels, height, width) internally, and might insert Transpose operators where needed in the model graph. The Sentis importer tries to minimize the number of transposes used for intermediate tensor operations. Consequently, intermediate tensors in the Sentis model may use a different layout than in the original LiteRT model. Input and output tensor shapes remain unchanged.

Sentis might create additional layers when it optimizes the model. A full list of Sentis-only layers is available here.

Unsupported features

Sentis currently does not support the following LiteRT features:

• Multiple subgraphs: If the model contains multiple subgraphs, only the first one is used. All others are ignored.
• SIGN_BIT activation: Operators using the fused_activation_function attribute with the value SIGN_BIT are not supported.
• Custom operators: Custom LiteRT operators are not supported.
• Sparse tensors: Sparse tensor formats are not supported.

Unsupported tensor types

The following tensor data types are not supported in any context:

• string
• complex64
• complex128
• resource
• variant

In addition, the following types are not supported for constant tensors:

• int4
• bfloat16

Unsupported operators

The following LiteRT operators are not supported in the current version of Sentis.

• assign_variable
• basic_lstm
• batch_to_space_nd
• bidirectional_sequence_lstm
• bitcast
• bucketize
• call_once
• complex_abs
• control_node
• custom
• custom_tf
• densify
• dilate
• dynamic_update_slice
• external_const
• fake_quant
• hashtable
• hashtable_find
• hashtable_import
• hashtable_size
• if
• imag
• lstm
• matrix_diag
• matrix_set_diag
• no_value
• non_max_suppression_v4
• non_max_suppression_v5
• NumericVerify
• poly_call
• pseudo_const
• pseudo_qconst
• pseudo_sparse_const
• pseudo_sparse_qconst
• quantize
• read_variable
• real
• reverse_sequence
• rfft2d
• right_shift
• segment_sum
• space_to_batch_nd
• StableHLO - All operators
• svdf
• unidirectional_sequence_lstm
• unidirectional_sequence_rnn
• unique
• unsorted_segment_max
• unsorted_segment_min
• unsorted_segment_prod
• unsorted_segment_sum
• var_handle
• while
• yield

Additional resources

• 'tfl' Dialect
• Profile a model
• Supported functional methods
• Supported ONNX operators