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)
    288 nodes, weights cbc2ec5c479a
    DecomposeAttention x5 b00b8abe
    DecomposeReduceL2 x3 80ee310c
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x106 16bb8366
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 68716d7d
    unstamped x164 84646c72

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)
    287 nodes, weights cbc2ec5c479a
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x106 e66e5f08
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 68716d7d
    unstamped x166 91df0027

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)
    287 nodes, weights cbc2ec5c479a
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x106 e66e5f08
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 68716d7d
    unstamped x166 91df0027

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)
    287 nodes, weights cbc2ec5c479a
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x106 e66e5f08
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 68716d7d
    unstamped x166 91df0027

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)
    323 nodes, weights 094e4023914c
    contract {"dims":[{}]}
    opset 19
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x105 95c1ee4e
    SplitLargeConvReduction, FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x8 c9a3f0f3
    SplitLargeReduction x33 463c69a3
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 68716d7d
    unstamped x162 e8bc9e8b

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)
    287 nodes, weights cbc2ec5c479a
    DecomposeAttention x5 b00b8abe
    FuseBatchNormIntoConv, RemoveOptionalBiasFromConv x106 e66e5f08
    _ConstantifyReshapeTarget x9 8c4eb3c1
    _ScalarGatherToSlice x1 68716d7d
    unstamped x166 91df0027
