CPUExecutionProvider
    -24 Add
    -24 LayerNormalization axis=-1 epsilon=9.999999747378752e-06 stash_type=1
    +24 SkipLayerNormalization epsilon=9.999999747378752e-06
    196 nodes, weights 35882ff153a1
    FuseSkipLayerNorm x23 a19fcdfa
    FuseSkipLayerNorm, _FuseClassTokenPrepend x1 ed3bb91a
    RemoveOptionalBiasFromConv x1 b03923d8
    _ConstantifyReshapeTarget x1 f786528a
    _FuseClassTokenPrepend x1 408ab1ab
    _ScalarGatherToSlice, _SelectBeforeLayerNorm x1 cde77a8c
    _SelectBeforeLayerNorm x1 ab62bee7
    unstamped x167 821d82d1

CUDAExecutionProvider
    -24 LayerNormalization axis=-1 epsilon=9.999999747378752e-06 stash_type=1
    +24 SkipLayerNormalization epsilon=9.999999747378752e-06
    -23 Add
    +2 Reshape
    -1 Conv auto_pad=NOTSET dilations=(1, 1) group=1 pads=(0, 0, 0, 0) strides=(16, 16)
    +1 MatMul
    -1 Reshape allowzero=1
    -1 Transpose perm=(0, 2, 1)
    +1 Transpose perm=(0, 2, 1, 3)
    -1 Transpose perm=(0, 3, 1, 2)
    197 nodes, weights afe570d12b24
    FuseSkipLayerNorm x23 a19fcdfa
    FuseSkipLayerNorm, _FuseClassTokenPrepend x1 ed3bb91a
    PatchEmbedToMatMul, _ConstantifyReshapeTarget, RemoveOptionalBiasFromConv x5 132b0235
    _FuseClassTokenPrepend x1 1c1de38b
    _ScalarGatherToSlice, _SelectBeforeLayerNorm x1 cde77a8c
    _SelectBeforeLayerNorm x1 ab62bee7
    unstamped x165 b3634aef

CoreMLExecutionProvider
    +48 Reshape
    +48 Transpose perm=(0, 2, 1, 3)
    +24 MatMul
    +13 Mul
    -12 Attention is_causal=0 kv_num_heads=12 q_num_heads=12 qk_matmul_output_mode=0 softcap=0.0
    +12 Softmax axis=-1
    +12 Transpose perm=(0, 1, 3, 2)
    -1 ReduceL2 keepdims=1 noop_with_empty_axes=0
    +1 ReduceSum keepdims=1
    +1 Sqrt
    -1 Transpose perm=(0, 3, 1, 2)
    365 nodes, weights c296c34b4325
    DecomposeAttention x156 04519fbd
    DecomposeReduceL2 x3 44b2cf3c
    RemoveOptionalBiasFromConv x1 464cce20
    _ConstantifyReshapeTarget x1 f786528a
    _FuseClassTokenPrepend x2 64b83b02
    _ScalarGatherToSlice, _SelectBeforeLayerNorm x1 cde77a8c
    _SelectBeforeLayerNorm x1 ab62bee7
    unstamped x200 201dc6d4

MIGraphXExecutionProvider
    +48 Reshape
    +48 Transpose perm=(0, 2, 1, 3)
    +24 MatMul
    -12 Attention is_causal=0 kv_num_heads=12 q_num_heads=12 qk_matmul_output_mode=0 softcap=0.0
    +12 Mul
    +12 Softmax axis=-1
    +12 Transpose perm=(0, 1, 3, 2)
    364 nodes, weights c296c34b4325
    DecomposeAttention x156 04519fbd
    RemoveOptionalBiasFromConv x1 b03923d8
    _ConstantifyReshapeTarget x1 f786528a
    _FuseClassTokenPrepend x2 64b83b02
    _ScalarGatherToSlice, _SelectBeforeLayerNorm x1 cde77a8c
    _SelectBeforeLayerNorm x1 ab62bee7
    unstamped x202 83047069

NvTensorRTRTXExecutionProvider
    +50 Reshape
    +49 Transpose perm=(0, 2, 1, 3)
    +25 MatMul
    -12 Attention is_causal=0 kv_num_heads=12 q_num_heads=12 qk_matmul_output_mode=0 softcap=0.0
    +12 Mul
    +12 Softmax axis=-1
    +12 Transpose perm=(0, 1, 3, 2)
    +1 Add
    -1 Conv auto_pad=NOTSET dilations=(1, 1) group=1 pads=(0, 0, 0, 0) strides=(16, 16)
    -1 Reshape allowzero=1
    -1 Transpose perm=(0, 2, 1)
    -1 Transpose perm=(0, 3, 1, 2)
    365 nodes, weights 1083d8dcce1b
    DecomposeAttention x156 04519fbd
    PatchEmbedToMatMul, _ConstantifyReshapeTarget, RemoveOptionalBiasFromConv x5 132b0235
    _FuseClassTokenPrepend x2 db84f216
    _ScalarGatherToSlice, _SelectBeforeLayerNorm x1 cde77a8c
    _SelectBeforeLayerNorm x1 ab62bee7
    unstamped x200 8b17c953

OpenVINOExecutionProvider
    +50 Reshape
    +49 Transpose perm=(0, 2, 1, 3)
    +25 MatMul
    -12 Attention is_causal=0 kv_num_heads=12 q_num_heads=12 qk_matmul_output_mode=0 softcap=0.0
    +12 Mul
    +12 Softmax axis=-1
    +12 Transpose perm=(0, 1, 3, 2)
    +1 Add
    -1 Conv auto_pad=NOTSET dilations=(1, 1) group=1 pads=(0, 0, 0, 0) strides=(16, 16)
    -1 Reshape allowzero=1
    -1 Transpose perm=(0, 2, 1)
    -1 Transpose perm=(0, 3, 1, 2)
    365 nodes, weights 1083d8dcce1b
    DecomposeAttention x156 04519fbd
    PatchEmbedToMatMul, _ConstantifyReshapeTarget, RemoveOptionalBiasFromConv x5 132b0235
    _FuseClassTokenPrepend x2 db84f216
    _ScalarGatherToSlice, _SelectBeforeLayerNorm x1 cde77a8c
    _SelectBeforeLayerNorm x1 ab62bee7
    unstamped x200 8b17c953

RKNPU static
    +85 MatMul
    +72 Slice
    +61 Add
    +50 Reshape
    +49 Transpose perm=(0, 2, 1, 3)
    -12 Attention is_causal=0 kv_num_heads=12 q_num_heads=12 qk_matmul_output_mode=0 softcap=0.0
    +12 Mul
    +12 Softmax axis=-1
    +12 Transpose perm=(0, 1, 3, 2)
    -1 Cast to=1
    -1 Conv auto_pad=NOTSET dilations=(1, 1) group=1 pads=(0, 0, 0, 0) strides=(16, 16)
    -1 Reshape allowzero=1
    -1 Transpose perm=(0, 2, 1)
    -1 Transpose perm=(0, 3, 1, 2)
    556 nodes, weights a086caacb39f
    contract {"dims":[{}]}
    opset 19
    DecomposeAttention x156 04519fbd
    PatchEmbedToMatMul, _ConstantifyReshapeTarget, RemoveOptionalBiasFromConv x5 37ecaa30
    SplitLargeReduction x204 61274ad0
    _FuseClassTokenPrepend x2 db84f216
    _ScalarGatherToSlice, _SelectBeforeLayerNorm x1 cde77a8c
    _SelectBeforeLayerNorm x1 ab62bee7
    unstamped x187 534d702a

TensorrtExecutionProvider
    +50 Reshape
    +49 Transpose perm=(0, 2, 1, 3)
    +25 MatMul
    -12 Attention is_causal=0 kv_num_heads=12 q_num_heads=12 qk_matmul_output_mode=0 softcap=0.0
    +12 Mul
    +12 Softmax axis=-1
    +12 Transpose perm=(0, 1, 3, 2)
    +1 Add
    -1 Conv auto_pad=NOTSET dilations=(1, 1) group=1 pads=(0, 0, 0, 0) strides=(16, 16)
    -1 Reshape allowzero=1
    -1 Transpose perm=(0, 2, 1)
    -1 Transpose perm=(0, 3, 1, 2)
    365 nodes, weights 1083d8dcce1b
    DecomposeAttention x156 04519fbd
    PatchEmbedToMatMul, _ConstantifyReshapeTarget, RemoveOptionalBiasFromConv x5 132b0235
    _FuseClassTokenPrepend x2 db84f216
    _ScalarGatherToSlice, _SelectBeforeLayerNorm x1 cde77a8c
    _SelectBeforeLayerNorm x1 ab62bee7
    unstamped x200 8b17c953
