CPUExecutionProvider
    -55 Add
    -55 LayerNormalization axis=-1 epsilon=9.999999974752427e-07 stash_type=1
    +55 SkipLayerNormalization epsilon=9.999999974752427e-07
    379 nodes, weights 7bf9172d6053
    FuseSkipLayerNorm x55 06e2e771
    _ConstantifyReshapeTarget x1 f786528a
    _ScalarGatherToSlice x1 60360a10
    unstamped x322 22ddc949

CUDAExecutionProvider
    -55 LayerNormalization axis=-1 epsilon=9.999999974752427e-07 stash_type=1
    +55 SkipLayerNormalization epsilon=9.999999974752427e-07
    -54 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)
    380 nodes, weights e6b33b388ef1
    FuseSkipLayerNorm x55 74bb412a
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 132b0235
    _ScalarGatherToSlice x1 60360a10
    unstamped x319 08063b59

CoreMLExecutionProvider
    +168 MatMul
    +140 Slice
    +112 Add
    +112 Reshape
    +112 Transpose perm=(0, 2, 1, 3)
    +29 Mul
    -28 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +28 Softmax axis=-1
    +28 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)
    1135 nodes, weights 574fd00f8bc9
    DecomposeAttention x364 6347034a
    DecomposeReduceL2 x3 4dfbb073
    SplitLargeReduction x392 6940d8e3
    _ConstantifyReshapeTarget x1 f786528a
    _ScalarGatherToSlice x1 60360a10
    unstamped x374 56512432

MIGraphXExecutionProvider
    +112 Reshape
    +112 Transpose perm=(0, 2, 1, 3)
    +56 MatMul
    -28 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +28 Mul
    +28 Softmax axis=-1
    +28 Transpose perm=(0, 1, 3, 2)
    770 nodes, weights 7cda44383b2f
    DecomposeAttention x364 6347034a
    _ConstantifyReshapeTarget x1 f786528a
    _ScalarGatherToSlice x1 60360a10
    unstamped x404 c3ba758c

NvTensorRTRTXExecutionProvider
    +114 Reshape
    +113 Transpose perm=(0, 2, 1, 3)
    +57 MatMul
    -28 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +28 Mul
    +28 Softmax axis=-1
    +28 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)
    771 nodes, weights 79f938563e5b
    DecomposeAttention x364 6347034a
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 132b0235
    _ScalarGatherToSlice x1 60360a10
    unstamped x401 6624dc0b

OpenVINOExecutionProvider
    +114 Reshape
    +113 Transpose perm=(0, 2, 1, 3)
    +57 MatMul
    -28 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +28 Mul
    +28 Softmax axis=-1
    +28 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)
    771 nodes, weights 79f938563e5b
    DecomposeAttention x364 6347034a
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 132b0235
    _ScalarGatherToSlice x1 60360a10
    unstamped x401 6624dc0b

RKNPU static
    +504 Slice
    +449 MatMul
    +421 Add
    +114 Reshape
    +113 Transpose perm=(0, 2, 1, 3)
    +83 Mul
    -28 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +28 Div
    +28 Erf
    -28 Gelu approximate=tanh
    +28 Softmax axis=-1
    +28 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 Sub
    -1 Transpose perm=(0, 2, 1)
    -1 Transpose perm=(0, 3, 1, 2)
    2170 nodes, weights a51ed328e3b0
    contract {"dims":[{}]}
    opset 19
    DecomposeAttention x364 6347034a
    DecomposeGelu x140 3c6aabc5
    OpaqueZeroMul x1 3037fff5
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 37ecaa30
    SplitLargeReduction x1400 d0fabd8b
    _ScalarGatherToSlice x1 60360a10
    unstamped x259 b7a40c98

TensorrtExecutionProvider
    +114 Reshape
    +113 Transpose perm=(0, 2, 1, 3)
    +57 MatMul
    -28 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +28 Mul
    +28 Softmax axis=-1
    +28 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)
    771 nodes, weights 79f938563e5b
    DecomposeAttention x364 6347034a
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 132b0235
    _ScalarGatherToSlice x1 60360a10
    unstamped x401 6624dc0b
