CPUExecutionProvider
    unchanged

CUDAExecutionProvider
    unchanged

CoreMLExecutionProvider
    +2 MatMul
    +2 Mul
    -1 Attention is_causal=0 qk_matmul_output_mode=0 softcap=0.0
    -1 ReduceL2 keepdims=1 noop_with_empty_axes=0
    +1 ReduceSum keepdims=1
    +1 Softmax axis=-1
    +1 Sqrt
    +1 Transpose perm=(0, 1, 3, 2)
    -1 Transpose perm=(0, 3, 1, 2)
    169 nodes, weights 5988167b1977
    DecomposeAttention x5 b00b8abe
    DecomposeReduceL2 x3 aff166da
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x55 2040c61b
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 6e430d0a
    unstamped x96 4edb0915

MIGraphXExecutionProvider
    +2 MatMul
    -1 Attention is_causal=0 qk_matmul_output_mode=0 softcap=0.0
    +1 Mul
    +1 Softmax axis=-1
    +1 Transpose perm=(0, 1, 3, 2)
    168 nodes, weights 5988167b1977
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x55 18e54fd8
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 6e430d0a
    unstamped x98 a5d8a72d

NvTensorRTRTXExecutionProvider
    +2 MatMul
    -1 Attention is_causal=0 qk_matmul_output_mode=0 softcap=0.0
    +1 Mul
    +1 Softmax axis=-1
    +1 Transpose perm=(0, 1, 3, 2)
    168 nodes, weights 5988167b1977
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x55 18e54fd8
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 6e430d0a
    unstamped x98 a5d8a72d

OpenVINOExecutionProvider
    +2 MatMul
    -1 Attention is_causal=0 qk_matmul_output_mode=0 softcap=0.0
    +1 Mul
    +1 Softmax axis=-1
    +1 Transpose perm=(0, 1, 3, 2)
    168 nodes, weights 5988167b1977
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x55 18e54fd8
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 6e430d0a
    unstamped x98 a5d8a72d

RKNPU static
    +15 Slice
    +11 Add
    +11 MatMul
    +2 Conv auto_pad=NOTSET dilations=(1, 1) group=1 pads=(1, 1, 1, 1) strides=(1, 1)
    -1 Attention is_causal=0 qk_matmul_output_mode=0 softcap=0.0
    -1 Cast to=1
    +1 Mul
    +1 Softmax axis=-1
    +1 Transpose perm=(0, 1, 3, 2)
    204 nodes, weights 307a9ce37455
    contract {"dims":[{}]}
    opset 19
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x54 8f65283c
    SplitLargeConvReduction, FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x8 679b95aa
    SplitLargeReduction x33 3cabc813
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 6e430d0a
    unstamped x94 9d87f2e4

TensorrtExecutionProvider
    +2 MatMul
    -1 Attention is_causal=0 qk_matmul_output_mode=0 softcap=0.0
    +1 Mul
    +1 Softmax axis=-1
    +1 Transpose perm=(0, 1, 3, 2)
    168 nodes, weights 5988167b1977
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x55 18e54fd8
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 6e430d0a
    unstamped x98 a5d8a72d
