CPUExecutionProvider
    -81 Add
    -81 LayerNormalization axis=-1 epsilon=9.999999974752427e-07 stash_type=1
    +81 SkipLayerNormalization epsilon=9.999999974752427e-07
    548 nodes, weights 45fc1a0e9ea1
    FuseSkipLayerNorm x81 bf2cb848
    _ConstantifyReshapeTarget x1 f786528a
    _ScalarGatherToSlice x1 60360a10
    unstamped x465 4b156575

CUDAExecutionProvider
    -81 LayerNormalization axis=-1 epsilon=9.999999974752427e-07 stash_type=1
    +81 SkipLayerNormalization epsilon=9.999999974752427e-07
    -80 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)
    549 nodes, weights 19a8f72f2866
    FuseSkipLayerNorm x81 cceefd9f
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 132b0235
    _ScalarGatherToSlice x1 60360a10
    unstamped x462 ce0687bc

CoreMLExecutionProvider
    +287 MatMul
    +246 Slice
    +205 Add
    +164 Reshape
    +164 Transpose perm=(0, 2, 1, 3)
    +42 Mul
    -41 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +41 Softmax axis=-1
    +41 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)
    1778 nodes, weights 40ed9b5eff6d
    DecomposeAttention x533 ef1f6e6a
    DecomposeReduceL2 x3 8d9dc122
    SplitLargeReduction x697 dde528bc
    _ConstantifyReshapeTarget x1 f786528a
    _ScalarGatherToSlice x1 60360a10
    unstamped x543 3b523238

MIGraphXExecutionProvider
    +164 Reshape
    +164 Transpose perm=(0, 2, 1, 3)
    +82 MatMul
    -41 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +41 Mul
    +41 Softmax axis=-1
    +41 Transpose perm=(0, 1, 3, 2)
    1121 nodes, weights dc716552e93d
    DecomposeAttention x533 ef1f6e6a
    _ConstantifyReshapeTarget x1 f786528a
    _ScalarGatherToSlice x1 60360a10
    unstamped x586 34e18bd1

NvTensorRTRTXExecutionProvider
    +166 Reshape
    +165 Transpose perm=(0, 2, 1, 3)
    +83 MatMul
    -41 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +41 Mul
    +41 Softmax axis=-1
    +41 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)
    1122 nodes, weights f259183f6d02
    DecomposeAttention x533 ef1f6e6a
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 132b0235
    _ScalarGatherToSlice x1 60360a10
    unstamped x583 e0d1db39

OpenVINOExecutionProvider
    +166 Reshape
    +165 Transpose perm=(0, 2, 1, 3)
    +83 MatMul
    -41 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +41 Mul
    +41 Softmax axis=-1
    +41 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)
    1122 nodes, weights f259183f6d02
    DecomposeAttention x533 ef1f6e6a
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 132b0235
    _ScalarGatherToSlice x1 60360a10
    unstamped x583 e0d1db39

RKNPU static
    +861 Slice
    +780 MatMul
    +739 Add
    +166 Reshape
    +165 Transpose perm=(0, 2, 1, 3)
    +122 Mul
    -41 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +41 Div
    +41 Erf
    -41 Gelu approximate=tanh
    +41 Softmax axis=-1
    +41 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)
    3540 nodes, weights aab91472e726
    contract {"dims":[{}]}
    opset 19
    DecomposeAttention x533 ef1f6e6a
    DecomposeGelu x205 de5f9100
    OpaqueZeroMul x1 3037fff5
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 37ecaa30
    SplitLargeReduction x2419 cc665ae1
    _ScalarGatherToSlice x1 60360a10
    unstamped x376 46676bcd

TensorrtExecutionProvider
    +166 Reshape
    +165 Transpose perm=(0, 2, 1, 3)
    +83 MatMul
    -41 Attention is_causal=0 kv_num_heads=16 q_num_heads=16 qk_matmul_output_mode=0 softcap=0.0
    +41 Mul
    +41 Softmax axis=-1
    +41 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)
    1122 nodes, weights f259183f6d02
    DecomposeAttention x533 ef1f6e6a
    PatchEmbedToMatMul, _ConstantifyReshapeTarget x5 132b0235
    _ScalarGatherToSlice x1 60360a10
    unstamped x583 e0d1db39
