Operator: aten._log_softmax.default
cnt: 1, ((T([511, 50257], f16), 1, False), {})
Operator: aten._log_softmax_backward_data.default
cnt: 1, ((T([511, 50257], f16), T([511, 50257], f16), 1, f16), {})
Operator: aten._softmax.default
cnt: 6, ((T([1, 12, 512, 512], f16), -1, False), {})
Operator: aten._softmax_backward_data.default
cnt: 6, ((T([1, 12, 512, 512], f16), T([1, 12, 512, 512], f16), -1, f16), {})
Operator: aten._to_copy.default
cnt: 6, ((T([1, 1, 512, 512], u8, stride=(1048576, 1048576, 1024, 1)),), {'dtype': torch.bool})
cnt: 6, ((T([], f16),), {'dtype': f16, 'layout': torch.strided, 'device': 'cuda'})
Operator: aten._unsafe_view.default
cnt: 6, ((T([12, 512, 512], f16), [1, 12, 512, 512]), {})
cnt: 6, ((T([12, 512, 64], f16), [1, 12, 512, 64]), {})
cnt: 1, ((T([512, 50257], f16), [1, 512, 50257]), {})
cnt: 12, ((T([1, 512, 12, 64], f16), [1, 512, 768]), {})
Operator: aten.add.Tensor
cnt: 25, ((T([1, 512, 768], f16), T([1, 512, 768], f16)), {})
cnt: 18, ((T([1, 512, 3072], f16), T([1, 512, 3072], f16)), {})
cnt: 6, ((T([1, 512, 3072], f16), 1.0), {})
cnt: 1, ((T([50257, 768], f16), T([50257, 768], f16)), {})
Operator: aten.addmm.default
cnt: 6, ((T([2304], f16), T([512, 768], f16), T([768, 2304], f16)), {})
cnt: 6, ((T([768], f16), T([512, 768], f16), T([768, 768], f16)), {})
cnt: 6, ((T([3072], f16), T([512, 768], f16), T([768, 3072], f16)), {})
cnt: 6, ((T([768], f16), T([512, 3072], f16), T([3072, 768], f16)), {})
Operator: aten.bmm.default
cnt: 6, ((T([12, 512, 64], f16, stride=(64, 2304, 1)), T([12, 64, 512], f16, stride=(64, 1, 2304))), {})
cnt: 12, ((T([12, 512, 512], f16), T([12, 512, 64], f16, stride=(64, 2304, 1))), {})
cnt: 6, ((T([12, 512, 512], f16, stride=(262144, 1, 512)), T([12, 512, 64], f16, stride=(64, 768, 1))), {})
cnt: 6, ((T([12, 512, 64], f16, stride=(64, 768, 1)), T([12, 64, 512], f16, stride=(64, 1, 2304))), {})
cnt: 6, ((T([12, 64, 512], f16, stride=(64, 1, 2304)), T([12, 512, 512], f16)), {})
Operator: aten.cat.default
cnt: 6, (([T([1, 512, 768], f16), T([1, 512, 768], f16, stride=(512, 1, 512)), T([1, 512, 768], f16)], 2), {})
Operator: aten.clone.default
cnt: 2, ((T([1, 512], i64),), {})
Operator: aten.copy_.default
cnt: 2, ((T([1, 512], i64), T([1, 512], i64)), {})
Operator: aten.div.Tensor
cnt: 12, ((T([1, 12, 512, 512], f16), T([], f16)), {})
Operator: aten.embedding.default
cnt: 1, ((T([50257, 768], f16), T([1, 512], i64)), {})
cnt: 1, ((T([1024, 768], f16), T([1, 512], i64)), {})
Operator: aten.embedding_dense_backward.default
cnt: 1, ((T([1, 512, 768], f16), T([1, 512], i64), 1024, -1, False), {})
cnt: 1, ((T([1, 512, 768], f16), T([1, 512], i64), 50257, -1, False), {})
Operator: aten.mm.default
cnt: 1, ((T([512, 768], f16), T([768, 50257], f16, stride=(1, 768))), {})
cnt: 1, ((T([50257, 512], f16, stride=(1, 50257)), T([512, 768], f16)), {})
cnt: 1, ((T([512, 50257], f16), T([50257, 768], f16)), {})
cnt: 6, ((T([512, 768], f16), T([768, 3072], f16, stride=(1, 768))), {})
cnt: 6, ((T([3072, 512], f16, stride=(1, 3072)), T([512, 768], f16)), {})
cnt: 6, ((T([512, 3072], f16), T([3072, 768], f16, stride=(1, 3072))), {})
cnt: 6, ((T([768, 512], f16, stride=(1, 768)), T([512, 3072], f16)), {})
cnt: 6, ((T([512, 768], f16), T([768, 768], f16, stride=(1, 768))), {})
cnt: 6, ((T([768, 512], f16, stride=(1, 768)), T([512, 768], f16)), {})
cnt: 6, ((T([512, 2304], f16), T([2304, 768], f16, stride=(1, 2304))), {})
cnt: 6, ((T([768, 512], f16, stride=(1, 768)), T([512, 2304], f16)), {})
Operator: aten.mul.Scalar
cnt: 6, ((T([1, 512, 3072], f16), 3.0), {})
Operator: aten.mul.Tensor
cnt: 12, ((T([1, 512, 3072], f16), 0.5), {})
cnt: 12, ((T([1, 512, 3072], f16), 0.044715), {})
cnt: 12, ((T([1, 512, 3072], f16), 0.7978845608028654), {})
cnt: 24, ((T([1, 512, 3072], f16), T([1, 512, 3072], f16)), {})
Operator: aten.native_layer_norm.default
cnt: 13, ((T([1, 512, 768], f16), [768], T([768], f16), T([768], f16), 1e-05), {})
Operator: aten.native_layer_norm_backward.default
cnt: 13, ((T([1, 512, 768], f16), T([1, 512, 768], f16), [768], T([1, 512, 1], f32), T([1, 512, 1], f32), T([768], f16), T([768], f16), [True, True, True]), {})
Operator: aten.nll_loss_backward.default
cnt: 1, ((T([], f16), T([511, 50257], f16), T([511], i64), None, 1, -100, T([], f16)), {})
Operator: aten.nll_loss_forward.default
cnt: 1, ((T([511, 50257], f16), T([511], i64), None, 1, -100), {})
Operator: aten.pow.Tensor_Scalar
cnt: 6, ((T([1, 512, 3072], f16), 3.0), {})
cnt: 6, ((T([1, 512, 3072], f16), 2.0), {})
Operator: aten.slice_backward.default
cnt: 1, ((T([1, 511, 50257], f16), [1, 511, 50257], 2, 0, 9223372036854775807, 1), {})
cnt: 1, ((T([1, 511, 50257], f16), [1, 512, 50257], 1, 0, -1, 1), {})
Operator: aten.split.Tensor
cnt: 6, ((T([1, 512, 2304], f16), 768, 2), {})
Operator: aten.sum.SymInt
cnt: 12, ((T([512, 768], f16), [0], True), {})
cnt: 6, ((T([512, 3072], f16), [0], True), {})
cnt: 6, ((T([512, 2304], f16), [0], True), {})
Operator: aten.tanh.default
cnt: 6, ((T([1, 512, 3072], f16),), {})
Operator: aten.tanh_backward.default
cnt: 6, ((T([1, 512, 3072], f16), T([1, 512, 3072], f16)), {})
Operator: aten.where.self
cnt: 12, ((T([1, 1, 512, 512], b8), T([1, 12, 512, 512], f16), T([], f16)), {})
