콘텐츠로 이동

Mobilint IR 연산 형식

아래 표는 Mobilint IR에 나타날 수 있는 연산의 형식을 설명합니다. 목록에 있다는 사실만으로 NPU 또는 CPU 오프로딩 지원이 보장되지는 않습니다. 지원 범위는 NPU 연산 지원 범위와 CPU 오프로딩을 확인하세요.

연산 비고
Abs num inputs: 1
num outputs: 1
AddPartial num inputs: 2
num outputs: 1
attributes:
- is_input0_larger: bool
AddingConstant num inputs: 1
num outputs: 1
attributes:
- constant: Weight
- is_const_first: bool (default=false)
- shape: int32[] (default=[])
Adding num inputs: 2
num outputs: 1
Batchnorm num inputs: 1
num outputs: 1
attributes:
- gamma: Weight
- beta: Weight
- moving_mean: Weight
- moving_var: Weight
- epsilon: Weight
- channel_last: bool (default=true)
Cast num inputs: 1
num outputs: 1
attributes:
- dtype: str
Ceil num inputs: 1
num outputs: 1
Celu num inputs: 1
num outputs: 1
attributes:
- alpha: float32
Clip num inputs: 1
num outputs: 1
attributes:
- clip_max_value: float32 (optional)
- clip_min_value: float32 (optional)
Concatenate num inputs: variable
num outputs: 1
attributes:
- original_key: str (optional)
- axis: int32 (optional)
- original_axis: int32 (optional)
- original_inputs: int32[] (optional)
- input_index: int32[] (optional)
- input_names: str[] (optional)

description:
input_names: input activation names in original order. input_index: input index in computation order.
Convolution num inputs: 1
num outputs: 1
attributes:
- in_channels: int32
- out_channels: int32
- h_kernel: int32
- w_kernel: int32
- h_stride: int32
- w_stride: int32
- h_dilation: int32
- w_dilation: int32
- west_padding: int32
- east_padding: int32
- north_padding: int32
- south_padding: int32
- padding_list: int32[] (default=[])
- group_number: int32
- weight: Weight
- bias: Weight (optional)
- channel_last: bool (default=true)
Convolution3d num inputs: 1
num outputs: 1
attributes:
- in_channels: int32
- out_channels: int32
- h_kernel: int32
- w_kernel: int32
- d_kernel: int32
- h_stride: int32
- w_stride: int32
- d_stride: int32
- h_dilation: int32
- w_dilation: int32
- d_dilation: int32
- west_padding: int32
- east_padding: int32
- north_padding: int32
- south_padding: int32
- front_padding: int32
- back_padding: int32
- padding_list: int32[] (default=[])
- group_number: int32
- weight: Weight
- bias: Weight (optional)
- channel_last: bool (default=true)
Crop num inputs: 1
num outputs: 1
attributes:
- w_offset: int32 (default=0)
- h_offset: int32 (default=0)
- crop_width: int32
- crop_height: int32
DepthwiseConvolution num inputs: 1
num outputs: 1
attributes:
- in_channels: int32
- out_channels: int32
- h_kernel: int32
- w_kernel: int32
- h_stride: int32
- w_stride: int32
- h_dilation: int32
- w_dilation: int32
- west_padding: int32
- east_padding: int32
- north_padding: int32
- south_padding: int32
- padding_list: int32[] (default=[])
- group_number: int32
- weight: Weight
- bias: Weight (optional)
- channel_last: bool (default=true)
DeviceBridge num inputs: 1
num outputs: 1
attributes:
- squeeze_axes: int32[]
- unsqueeze_axes: int32[]
- permute_axes: int64[]
- device: str (optional)
DivConstant num inputs: 1
num outputs: 1
attributes:
- constant: Weight
- is_const_first: bool
- shape: int32[] (default=[])
Div num inputs: 2
num outputs: 1
Elu num inputs: 1
num outputs: 1
attributes:
- leaky_alpha: float32
Erf num inputs: 1
num outputs: 1
Exp num inputs: 1
num outputs: 1
Expand num inputs: 1
num outputs: 1
attributes:
- constant: Weight (optional)
- is_const_first: bool
Gelu num inputs: 1
num outputs: 1
attributes:
- approximate: bool (default=false)
GLU num inputs: 1
num outputs: 1
GridSample num inputs: 2
num outputs: 1
attributes:
- align_corners: int32 (default=0)
- mode: str (default="linear")
- padding_mode: str (default="zeros")

description:
inputs: X, grid
X is in channel-last format.
X: (n, h, w, c)
grid: (n, h_out, w_out, num_spatial_dims). in this case, num_spatial_dims=2.
GroupConvolution num inputs: 1
num outputs: 1
attributes:
- in_channels: int32
- out_channels: int32
- h_kernel: int32
- w_kernel: int32
- h_stride: int32
- w_stride: int32
- h_dilation: int32
- w_dilation: int32
- west_padding: int32
- east_padding: int32
- north_padding: int32
- south_padding: int32
- padding_list: int32[] (default=[])
- group_number: int32
- weight: Weight
- bias: Weight (optional)
- channel_last: bool (default=true)
GroupNormalization num inputs: 1
num outputs: 1
attributes:
- num_groups: int32
- scale: Weight
- bias: Weight (optional)
- epsilon: Weight
- channel_last: bool (default=true)
HardSigmoid num inputs: 1
num outputs: 1
attributes:
- alpha: float32 (optional)
- beta: float32 (optional)
Hardtanh num inputs: 1
num outputs: 1
attributes:
- min_val: float32 (default=-1.0)
- max_val: float32 (default=1.0)
HardSwish num inputs: 1
num outputs: 1
Hardshrink num inputs: 1
num outputs: 1
attributes:
- lambd: float32 (default=0.5)
Softshrink num inputs: 1
num outputs: 1
attributes:
- lambd: float32 (default=0.5)
HeaderView num inputs: 1
num outputs: 1
attributes:
- view_shape: int32[] (default=[])
- num_heads: int32 (default=-1)
- batch_order: int32 (default=0)

description:
HeaderView simplifies the following sequence of operations.
If batch_order == 0
(n,h,w,c)-[reshape]->(1, h, w, num_heads c/num_heads)
-[transpose01324]->(1, h, num_heads, w, c/num_heads)
-[reshape]->(1, hnum_heads, w, c/num_heads)
If batch_order == 1
(n,h,w,c)-[reshape]->(1, h, w, num_heads, c/num_heads)
-[transpose03124]->(1, num_heads, h, w, c/num_heads)
-[reshape]->(1, num_heads
h, w, c/num_heads)
HeaderViewRevert num inputs: 1
num outputs: 1
attributes:
- revert_shape: int32[] (default=[])
- num_heads: int32 (default=-1)
- batch_order: int32 (default=0)

description:
HeaderViewRevert simplifies the following sequence of operations.
If batch_order == 0
(n,h,w,c)-[reshape]->(1, h/num_heads, num_heads, w, c)
-[transpose01324]->(1, h/num_heads, w, num_heads, c)
-[reshape]->(1, h/num_heads, w, num_heads * c)
If batch_order == 1
(n,h,w,c)-[reshape]->(1, num_heads, h/num_heads, w, c)
-[transpose02314]->(1, h/num_heads, w, num_heads, c)
-[reshape]->(1, h/num_heads, w, num_heads*c)
Identity num inputs: 1
num outputs: 1
InputConstant num inputs: 0
num outputs: 1
attributes:
- constant: Weight
- dtype: str (default="")
- shape: int32[] (default=[])
Input num inputs: 1
num outputs: 1
attributes:
- orig_name: str (default="")
InstanceNormalization num inputs: 1
num outputs: 1
attributes:
- scale: Weight
- bias: Weight (optional)
- epsilon: Weight
- channel_last: bool (default=true)
- num_groups: int32 (default=-1)
- group_order: int32 (default=-1)

description:
If num_groups and order are specified, the channel dimension is grouped and rearranged before normalization.
For example, the following sequence:
(n, 1, w, c2)-[reshape]->(n, w, c, 2)
-[transpose0312]->(n, 2, w, c)
-[InstanceNorm, num_features=c]->(n, 2, w, c)
-[transpose0231]->(n, w, c, 2)
-[reshape]->(n, 1, w, c
2)
is equivalent to a single operation:
(n, 1, w, c2)-[InstanceNormalization, num_features=c, num_groups=2, group_order=1]->(n, 1, w, c2)
The follosing sequence:
(n, 1, w, 2c)-[reshape]->(n,w,2,c)
-[InstanceNorm, num_features=c]->(n, w, 2, c)
-[reshape]v(n, 1, w, c
2)
is equivalent to a single operation:
(n, 1, w, 2c)-[InstanceNormalization, num_features=c, num_groups=2, group_order=0]->(n, 1, w, 2c)
L1Normalization num inputs: 1
num outputs: 1
attributes:
- norm_axis: int32[]
L2Normalization num inputs: 1
num outputs: 1
attributes:
- norm_axis: int32[]
- epsilon: Weight (optional)
SumNormalization num inputs: 1
num outputs: 1
attributes:
- norm_axis: int32[]
- epsilon: float32 (optional)
LSTM num inputs: variable
num outputs: 1
attributes:
- W: Weight
- R: Weight
- B: Weight (optional)
- lstm_group: str
- input_mask: str
- output_mask: str
- P: Weight (optional)
- direction: str (default="forward")
- batch_first: bool (default=false)
- hidden_size: int32

description:
inputs:[input, sequence_lens, initial_h, initial_c]
outputs:[output] or [hidden_out] or [cell_out]
input_mask: "ishc" or "i", etc,...
output_mask: "o" or "h" or "c"
LayerNormalization num inputs: 1
num outputs: 1
attributes:
- normalized_shape: int64[] (default=[])
- scale: Weight
- bias: Weight (optional)
- epsilon: Weight
Logit num inputs: 1
num outputs: 1
attributes:
- eps: float32 (optional)

description:
Compute y = ln(z / (1 - z)). If eps is given, we clamp input z = clamp(z, min=eps, max=1-eps)
RNN num inputs: variable
num outputs: 1
attributes:
- W: Weight
- R: Weight
- B: Weight (optional)
- rnn_group: str
- input_mask: str
- output_mask: str
- direction: str (default="forward")
- batch_first: bool (default=false)
- hidden_size: int32

description:
inputs:[input, sequence_lens, initial_h]
outputs:[output] or [hidden_out]
input_mask: "ish" or "i", etc,...
output_mask: "o" or "h"
GRU num inputs: variable
num outputs: 1
attributes:
- W: Weight
- R: Weight
- B: Weight (optional)
- gru_group: str
- input_mask: str
- output_mask: str
- direction: str (default="forward")
- batch_first: bool (default=false)
- linear_before_reset: bool (default=false)
- hidden_size: int32

description:
inputs:[input, sequence_lens, initial_h]
outputs:[output] or [hidden_out]
input_mask: "ish" or "i", etc,...
output_mask: "o" or "h"
LeakyRelu num inputs: 1
num outputs: 1
attributes:
- leaky_alpha: float32
Log num inputs: 1
num outputs: 1
LogSigmoid num inputs: 1
num outputs: 1
MaskSoftmax num inputs: 1
num outputs: 1
attributes:
- beta: float32
- axis: int32
MatMul num inputs: variable
num outputs: 1
attributes:
- constant: Weight (optional)
- is_const_first: bool (default=false)
- repeat: int32[2] (default=[-1, -1])

description:
If repeat=(0, 4), then input0 is repeated 4 times along the h-direction.
For example, you can perform a matrix multiplication with input0 of shape (1, 8, 52, 64) and
input1 of shape (1, 32, 52, 64) by concatenating input0 4 times, resulting in a input0 shape of (1, 32, 52, 64).
Mish num inputs: 1
num outputs: 1
MultiplyConstant num inputs: 1
num outputs: 1
attributes:
- constant: Weight
- is_const_first: bool (default=false)
- shape: int32[] (default=[])
Multiply num inputs: 2
num outputs: 1
NonMaxSuppression num inputs: 2
num outputs: variable
attributes:
- max_output_size: int64
- iou_threshold: float32
- score_threshold: float32
- soft_nms_sigma: float32
- pad_to_max_output_size: bool
- center_point_box: int32 (default=0)
Output num inputs: 1
num outputs: 1
attributes:
- orig_name: str (default="")
PRelu num inputs: 1
num outputs: 1
attributes:
- leaky_alpha: Weight
Pad num inputs: 1
num outputs: 1
attributes:
- west_padding: int32 (default=-1)
- east_padding: int32 (default=-1)
- south_padding: int32 (default=-1)
- north_padding: int32 (default=-1)
- pad_mode: str
- constant: float32 (default=0.0)
- padding_list: int32[] (optional)
- channel_last: bool (default=true)

description:
padding_list format should be [x1_begin, x2_begin, …, x1_end, x2_end,…],
where xi_begin is the number of pad values added at the beginning of axis
Pooling num inputs: 1
num outputs: 1
attributes:
- h_kernel: int32
- w_kernel: int32
- h_stride: int32
- w_stride: int32
- h_dilation: int32 (default=1)
- w_dilation: int32 (default=1)
- west_padding: int32
- east_padding: int32
- north_padding: int32
- south_padding: int32
- padding_list: int32[] (default=[])
- pool_type: str
- pad_mode: str
- is_ceil_mode: bool (default=false)
- is_include_pad: bool (default=false)
- channel_last: bool (default=true)
GlobalAveragePooling num inputs: 1
num outputs: 1
attributes:
- channel_last: bool (default=true)
GlobalResponseNormalization num inputs: 1
num outputs: 1
attributes:
- scale: Weight
- bias: Weight
- epsilon: float32
- channel_last: bool (default=true)
Pow num inputs: 1
num outputs: 1
attributes:
- exponent: float32
QuickGelu num inputs: 1
num outputs: 1
ReduceL2 num inputs: 1
num outputs: 1
attributes:
- keep_dims: bool
- reduce_axes: int32[]
- noop_with_empty_axes: int32
ReduceMax num inputs: 1
num outputs: 1
attributes:
- keep_dims: bool
- reduce_axes: int32[]
- noop_with_empty_axes: int32
ReduceMin num inputs: 1
num outputs: 1
attributes:
- keep_dims: bool
- reduce_axes: int32[]
- noop_with_empty_axes: int32
ReduceMean num inputs: 1
num outputs: 1
attributes:
- keep_dims: bool
- reduce_axes: int32[]
- noop_with_empty_axes: int32
ReduceProd num inputs: 1
num outputs: 1
attributes:
- keep_dims: bool
- reduce_axes: int32[]
- noop_with_empty_axes: int32
ReduceSum num inputs: 1
num outputs: 1
attributes:
- keep_dims: bool
- reduce_axes: int32[]
- noop_with_empty_axes: int32
ReduceRMS num inputs: 1
num outputs: 1
attributes:
- keep_dims: bool
- reduce_axes: int32[]
- epsilon: float32 (optional)

description:
reduce rms with epsilon if exists. i.e.,sqrt(mean**2 + eps)
we let eps to be optional so that it has more general fusing opportunities.
MeanCentering num inputs: 1
num outputs: 1
attributes:
- axes: int32[]

description:
compute x - mean(x, axes)
Relu num inputs: 1
num outputs: 1
Reshape num inputs: 1
num outputs: 1
attributes:
- new_shape: int32[] (optional)
- allowzero: int32 (default=0)
- is_const_first: bool (default=false)
ReshapeV2 num inputs: 2
num outputs: 1
attributes:
- allowzero: int32 (default=0)
- is_const_first: bool (default=false)

description:
Reshape with variable shape argument. inputs[1] will be used as a shape.
Resize num inputs: 1
num outputs: 1
attributes:
- resize_type: str (optional)
- resize_mode: str (optional)
- nearest_mode: str (optional)
- transform_mode: str (optional)
- roi: int32[] (optional)
- scales: float32[] (optional)
- sizes: int32[] (optional)
- cubic_coeff_a: float32 (optional)
- exclude_outside: int32 (optional)
- extrapolation_value: float32 (optional)
- antialias: int32 (optional)
- axes: int32[] (optional)
- keep_aspect_ratio_policy: str (optional)
RmsNormalization num inputs: 1
num outputs: 1
attributes:
- normalized_shape: int64[] (default=[])
- epsilon: Weight
- scale: Weight
- rms_center: Weight (optional)
RotaryEmbedding num inputs: 1
num outputs: 2
attributes:
- inv_freq: Weight
- attention_scaling: float32

description:
following the LlamaRotaryEmbedding of the Huggingface Transformer library.
inputs => (position_ids,), outputs => (cos, sin)
Rsqrt num inputs: 1
num outputs: 1
Rstd num inputs: 1
num outputs: 1
attributes:
- reduce_axes: int32[]
- epsilon: Weight
- num_groups: int32 (default=1)

description:
computes 1/std(x). if gruops > 1, then activation is grouped along channel axis.
Gather num inputs: 2
num outputs: 1
attributes:
- axis: int32 (default=0)
GatherElements num inputs: 2
num outputs: 1
attributes:
- axis: int32 (default=0)
ScatterND num inputs: variable
num outputs: 1
attributes:
- constant_data: Weight (optional)
- constant_indices: Weight (optional)
Selu num inputs: 1
num outputs: 1
Shape num inputs: 1
num outputs: 1
attributes:
- start: int32 (optional)
- end: int32 (optional)
Sigmoid num inputs: 1
num outputs: 1
attributes:
- alpha: float32 (default=1.0)

description:
alpha is added to deal with more general case like 1 / (1 + exp(-alpha * x))
InverseSigmoid num inputs: 1
num outputs: 1
attributes:
- eps: float32 (default=1e-05)

description:
x = x.clamp(x, min=0, max=1)
x1 = x.clamp(min=eps)
x2 = (1 - x).clamp(min=eps)
out = log(x1 / x2)
Slice num inputs: 1
num outputs: 1
attributes:
- axis: int32[]
- start: int64[]
- end: int64[]
- step: int64[] (optional)
Softmax num inputs: 1
num outputs: 1
attributes:
- beta: float32 (default=1.0)
- axis: int32
LogSoftmax num inputs: 1
num outputs: 1
attributes:
- axis: int32
Softplus num inputs: 1
num outputs: 1
attributes:
- beta: float32 (default=1.0)
- threshold: float32 (default=20.0)
Softsign num inputs: 1
num outputs: 1
Sqrt num inputs: 1
num outputs: 1
StaggeredPadding num inputs: 1
num outputs: 1
attributes:
- axis: int32 (default=3)
- reverse: bool (default=false)
- window_size: int32 (default=-1)

description:
StaggeredPadding does the following:
[[[[1,1,1],
[2,2,2],
[3,3,3],
[4,4,4]]
[5,5,5]]]]
=> [[[[1,1,1,0,0,0,0],
[0,2,2,2,0,0,0],
[0,0,3,3,3,0,0],
[0,0,0,4,4,4,0],
[0,0,0,0,5,5,5]]]]
That is, each original row is shifted to the right so that its non-zero block lies on a “diagonal,” and all other elements are zero-padded.
Then slice last dim by (2window_size + 1). For example, if window_size=1 in the above, 5=21+1 elements are sliced.
=> [[[[1,1,0,0,0],
[2,2,2,0,0],
[0,3,3,3,0],
[0,0,4,4,4],
[0,0,0,5,5]]]]
If reverse
[[[[1,1,1],
[2,2,2],
[3,3,3],
[4,4,4]]
[5,5,5]]]]
=> [[[[0,0,0,0,1,1,1],
[0,0,0,2,2,2,0],
[0,0,3,3,3,0,0],
[0,4,4,4,0,0,0],
[5,5,5,0,0,0,0]]]]
StatefulAttentionMaskWrapper num inputs: 1 or 2
num outputs: 1
attributes:
- past_seqlen: int32 (default=0)
- is_sliding: bool (default=false)
- sliding_window: int32 (default=1024)
- mask_type: str (default="causal_mask")

description:
mask_type: "causal_mask", "custom_mask" — "custom_mask" takes the mask as a second input
StatefulKVCacheWrapper num inputs: 1
num outputs: 1
attributes:
- cache_name: str
- past_seqlen: int32 (default=0)
- is_sliding: bool (default=false)
- sliding_window: int32 (default=1024)
- max_cache_len: int32 (default=8192)
StatefulCachedStack num inputs: 1
num outputs: 1
attributes:
- axis: int32 (default=2)
- cache_len: int32 (default=0)
- cache_name: str
StatefulLSTMWrapper num inputs: variable
num outputs: 3
attributes:
- W: Weight
- R: Weight
- B: Weight (optional)
- input_mask: str
- output_mask: str
- P: Weight (optional)
- direction: str (default="forward")
- batch_first: bool (default=false)
- hidden_size: int32
- hidden_state_name: str
- cell_state_name: str
- subgraph: str
- activation_alpha: float32[] (optional)
- activation_beta: float32[] (optional)
- activations: str[] (optional)
- clip: float32 (optional)
- input_forget: bool (default=false)

description:
When input_mask is ishc, then inputs are [input, sequence_lens, initial_h, initial_c],
and if input_mask is ihc, then inputs are [input, initial_h, initial_c].
The output_mask is h means that among [output, hidden_out, cell_out], only the hidden_out is being used.
StatefulRNNWrapper num inputs: variable
num outputs: 2
attributes:
- W: Weight
- R: Weight
- B: Weight (optional)
- input_mask: str
- output_mask: str
- direction: str (default="forward")
- batch_first: bool (default=false)
- hidden_size: int32
- hidden_state_name: str
- subgraph: str
- activation_alpha: float32[] (optional)
- activation_beta: float32[] (optional)
- activations: str[] (optional)
- clip: float32 (optional)

description:
When input_mask is ish, then inputs are [input, sequence_lens, initial_h],
and if input_mask is ih, then inputs are [input, initial_h].
The output_mask is h means that among [output, hidden_out], only the hidden_out is being used.
StatefulGRUWrapper num inputs: variable
num outputs: 2
attributes:
- W: Weight
- R: Weight
- B: Weight (optional)
- input_mask: str
- output_mask: str
- direction: str (default="forward")
- batch_first: bool (default=false)
- hidden_size: int32
- hidden_state_name: str
- subgraph: str
- activation_alpha: float32[] (optional)
- activation_beta: float32[] (optional)
- activations: str[] (optional)
- clip: float32 (optional)
- linear_before_reset: bool (default=false)

description:
When input_mask is ish, then inputs are [input, sequence_lens, initial_h],
and if input_mask is ih, then inputs are [input, initial_h].
The output_mask is h means that among [output, hidden_out], only the hidden_out is being used.
StatefulRoPEWrapper num inputs: 3
num outputs: 1
attributes:
- rope_type: int32 (default=0)
- seqlen: int32 (default=0)
- rotary_emb_dim: int32 (default=0)

description:
StatefulRoPEWrapper
inputs: (q_or_k, con, sin)
outputs: (q_or_k_embed)
rope_type=> 0: llama, 1: cohere2(aya-vision)
SubConstant num inputs: 1
num outputs: 1
attributes:
- constant: Weight
- is_const_first: bool
- shape: int32[] (default=[])
Sub num inputs: 2
num outputs: 1
Swish num inputs: 1
num outputs: 1
Tanh num inputs: 1
num outputs: 1
Sin num inputs: 1
num outputs: 1
Cos num inputs: 1
num outputs: 1
TransposeConvolution num inputs: 1
num outputs: 1
attributes:
- in_channels: int32
- out_channels: int32
- h_kernel: int32
- w_kernel: int32
- h_stride: int32
- w_stride: int32
- h_dilation: int32
- w_dilation: int32
- west_padding: int32
- east_padding: int32
- north_padding: int32
- south_padding: int32
- padding_list: int32[] (default=[])
- og_padding_list: int32[] (default=[])
- og_out_padding_list: int32[] (default=[])
- group_number: int32
- weight: Weight
- bias: Weight (optional)
- channel_last: bool (default=true)
Transpose num inputs: 1
num outputs: 1
attributes:
- transpose_axes: int64[]
Unsqueeze num inputs: 1
num outputs: 1
attributes:
- axes: int32[]
OrderedPatchify num inputs: 1
num outputs: 1
attributes:
- h_kernel: int32
- w_kernel: int32
- order: int32 (default=0)

description:
This operation splits a 4D input tensor into non-overlapping patches and rearranges them into a sequence
according to the specified order. Supports row-major (order=0) and column-major (order=1) layouts.
If order=0 (row-major), the equivalent is: 256 patches, each 12x12, (h_kernel=w_kernel=12)
(1,192,192,64)-[reshape]->(1,16,12,16,12,64)-[transpose 013245]->(1,16,16,12,12,64)-[reshape]->(1,256,144,64)
Note that order0 is its own inverse.
If order=1 (column-major), the equivalent is: 144 patches, each 16x16, (h_kernel=w_kernel=12)
(1,192,192,64)-[reshape]->(1,16,12,16,12,64)-[transpose 024135]->(1,12,12,16,16,64)-[reshape]->(1,144,256,64)
if order=2 (reversing order=1), h_kernel=12, w_kernel=16
(1,144,256,64)-[reshape]->(1,12,12,16,16,64)-[transpose 031425]->(1,16,12,16,12,64)-[reshape]->(1,192,192,64)
DynamicSliceInputConst num inputs: 1
num outputs: 1
attributes:
- mod: int32
- start: int64[]
- end: int64[]
- axis: int64[]
- step: int64[] (optional)
RelShift num inputs: 1
num outputs: 1
MeloTTSPiecewiseRationalQuadraticTransform num inputs: 2
num outputs: 1
attributes:
- num_bins: int32
- filter_channels: int32
- tail_bound: float32
- reverse: bool

description:
helper function for parsing melotts model.
Neg num inputs: 1
num outputs: 1
Einsum num inputs: variable
num outputs: 1
attributes:
- equation: str
- constant: Weight (optional)
- is_const_first: bool (default=false)
Where num inputs: 3
num outputs: 1


description:
condition,X,Y, if cond == True: else y
TensorCmp num inputs: 2
num outputs: 1
attributes:
- kind: str (default="eq")

description:
kind: eq, neq, lt, le, gt, ge
Logical num inputs: variable
num outputs: 1
attributes:
- kind: str (default="and")

description:
this function represents the logical operation of two inputs.
single input if kind is "not" else two inputs
kind: and, or, xor, not
MaskedFill num inputs: 2
num outputs: 1
attributes:
- value: float32

description:
inputs: (x, mask)
SubgraphCall num inputs: variable
num outputs: variable
attributes:
- subgraph: str
- scope: str
- kind: str
MoeDispatcher num inputs: 2
num outputs: 2
attributes:
- top_k: int32
- norm_top_k_prob: bool (default=true)

description:
inputs: (hidden_states, router_logits)
outputs: (hidden_states, top_k_logits)
if norm_top_k_prob is true
topk_vals, topk_idx = torch.topk(
routing_weights, self.top_k, dim=-1
) # (18,8), (18,8)
weight_new = torch.zeros_like(routing_weights)
weight_new.scatter_(1, topk_idx, topk_vals)
routing_weights = weight_new
routing_weights /= routing_weights.sum(dim=1, keepdim=True)
MoeAggregator num inputs: variable
num outputs: 1


description:
inputs: (router_weights, experts_output0, expert_output1, ...)
TopKValues num inputs: 1
num outputs: 1
attributes:
- axis: int32
- k: int32 (default=1)
- largest: bool (default=true)
- sorted: bool (default=true)
TopKIndices num inputs: 1
num outputs: 1
attributes:
- axis: int32
- k: int32 (default=1)
- largest: bool (default=true)
- sorted: bool (default=true)