CPUExecutionProvider
    unchanged

CUDAExecutionProvider
    -2 HardSigmoid alpha=0.1666666716337204 beta=0.5
    +2 HardSwish
    -2 Mul
    -1 ArgMax axis=2 keepdims=0
    -1 ReduceMax keepdims=1
    +1 Squeeze
    +1 TopK axis=2 largest=1 sorted=1
    89 nodes, weights 6307b4e60881
    FuseBatchNormIntoConv x2 a6b06e3d
    FuseBatchNormIntoConv, _HoistBatchNormOverSqueeze x2 7e7cb29a
    FuseGreedyCtcTopK x2 cf2f64e5
    FuseHardSwish x2 f424200c
    _FoldBiasAdd x33 b2294c37
    _FuseErfGelu x10 094e651a
    _HoistBatchNormOverSqueeze x2 34d81872
    unstamped x36 1a25fd6c

CoreMLExecutionProvider
    -1 Transpose perm=(0, 3, 1, 2)
    90 nodes, weights 4dc0e0a39884
    FuseBatchNormIntoConv x2 a6b06e3d
    FuseBatchNormIntoConv, _HoistBatchNormOverSqueeze x2 7e7cb29a
    _FoldBiasAdd x33 b2294c37
    _FuseErfGelu x10 094e651a
    _HoistBatchNormOverSqueeze x2 34d81872
    unstamped x41 c18248a1

MIGraphXExecutionProvider
    unchanged

NvTensorRTRTXExecutionProvider
    -1 ArgMax axis=2 keepdims=0
    -1 ReduceMax keepdims=1
    +1 Squeeze
    +1 TopK axis=2 largest=1 sorted=1
    91 nodes, weights 6307b4e60881
    FuseBatchNormIntoConv x2 a6b06e3d
    FuseBatchNormIntoConv, _HoistBatchNormOverSqueeze x2 7e7cb29a
    FuseGreedyCtcTopK x2 cf2f64e5
    _FoldBiasAdd x33 b2294c37
    _FuseErfGelu x10 094e651a
    _HoistBatchNormOverSqueeze x2 34d81872
    unstamped x40 492c5c16

OpenVINOExecutionProvider
    -2 HardSigmoid alpha=0.1666666716337204 beta=0.5
    +2 HardSwish
    -2 Mul
    -1 ArgMax axis=2 keepdims=0
    -1 Cast to=6
    -1 Exp
    -1 Reciprocal
    -1 ReduceMax keepdims=1
    -1 ReduceSum keepdims=0
    +1 Softmax axis=2
    -1 Sub
    83 nodes, weights 4dc0e0a39884
    FuseBatchNormIntoConv x2 a6b06e3d
    FuseBatchNormIntoConv, _HoistBatchNormOverSqueeze x2 7e7cb29a
    FuseHardSwish x2 f424200c
    _FoldBiasAdd x33 b2294c37
    _FuseErfGelu x10 094e651a
    _HoistBatchNormOverSqueeze x2 34d81872
    unstamped x32 80fa2071

RKNPU width=2048
    +20 Mul
    +10 Add
    +10 Div
    +10 Erf
    -10 Gelu approximate=none
    +4 Transpose perm=(0, 3, 2, 1)
    -2 Sub
    -1 ArgMax axis=2 keepdims=0
    -1 Cast to=1
    -1 Cast to=6
    +1 Concat axis=3
    -1 Exp
    -1 Reciprocal
    -1 ReduceMax keepdims=1
    -1 ReduceSum keepdims=0
    +1 Split axis=1
    +1 Transpose perm=(0, 2, 1)
    -1 Transpose perm=(0, 3, 1, 2)
    +1 Unsqueeze
    129 nodes, weights fac4c916c227
    config {'mean_values': [[127.5, 127.5, 127.5]], 'std_values': [[1.0, 1.0, 1.0]]}
    contract {"dims":[{"width":2048},{"width":224},{"width":320},{"width":448},{"width":640},{"width":1280}]}
    opset 19
    DecomposeGelu, _FuseErfGelu x50 c8c55aac
    FuseBatchNormIntoConv x2 b8290aa0
    FuseBatchNormIntoConv, _HoistBatchNormOverSqueeze x2 7e7cb29a
    _FoldBiasAdd x33 e15d452c
    _HoistBatchNormOverSqueeze x2 34d81872
    unstamped x40 cd9c5258

TensorrtExecutionProvider
    -1 ArgMax axis=2 keepdims=0
    -1 ReduceMax keepdims=1
    +1 Squeeze
    +1 TopK axis=2 largest=1 sorted=1
    91 nodes, weights 6307b4e60881
    FuseBatchNormIntoConv x2 a6b06e3d
    FuseBatchNormIntoConv, _HoistBatchNormOverSqueeze x2 7e7cb29a
    FuseGreedyCtcTopK x2 cf2f64e5
    _FoldBiasAdd x33 b2294c37
    _FuseErfGelu x10 094e651a
    _HoistBatchNormOverSqueeze x2 34d81872
    unstamped x40 492c5c16
