{
"model": "Qwen/Qwen3-32B",
"scenario": "decode",
"total_ops": 2392,
"unique_ops": 30,
"mapped_count": 26,
"unmapped_count": 4,
"unmapped_ops": [
"aten.bitwise_and.Tensor",
"aten.ge.Scalar",
"aten.lt.Scalar",
"aten.sub.Tensor"
],
"mapped_ops": [
"aten.add.Tensor",
"aten.alias.default",
"aten.copy_.default",
"aten.embedding.default",
"aten.full.default",
"aten.index.Tensor",
"aten.mm.default",
"aten.mul.Tensor",
"aten.permute.default",
"aten.select.int",
"aten.slice.Tensor",
"aten.split.Tensor",
"aten.split_with_sizes.default",
"aten.unsqueeze.default",
"aten.view.default",
"aten.where.self",
"prims.convert_element_type.default",
"tensor_cast.add_rms_norm2.default",
"tensor_cast.all_gather.default",
"tensor_cast.all_reduce.default",
"tensor_cast.apply_rope.default",
"tensor_cast.attention.default",
"tensor_cast.matmul_all_reduce.default",
"tensor_cast.reshape_and_cache.default",
"tensor_cast.rms_norm.default",
"tensor_cast.swiglu.default"
],
"ops": [
{
"op_name": "aten.mm.default",
"count": 129,
"total_duration_us": 3234.0,
"unique_shape_variants": 3,
"shape_variants": [
{
"input_shapes": [
"(16, 5120)",
"(5120, 768)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(16, 768)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 5120)",
"(5120, 3200)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(16, 3200)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 5120)",
"(5120, 9496)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(16, 9496)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "MatMulV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.attention.default",
"count": 64,
"total_duration_us": 2368.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(16, 512)",
"(513, 128, 1, 128)",
"(513, 128, 1, 128)",
"(16, 33)"
],
"input_dtypes": [
"float16",
"float16",
"float16",
"int64"
],
"output_shapes": [
"(16, 512)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "FusedInferAttentionScore",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.matmul_all_reduce.default",
"count": 128,
"total_duration_us": 2368.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(16, 512)",
"(512, 5120)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(16, 5120)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 1600)",
"(1600, 5120)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(16, 5120)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": false,
"composite": true,
"category": null,
"sub_kernels": [
"MatMulV2",
"hcom_allReduce_"
]
},
{
"op_name": "tensor_cast.rms_norm.default",
"count": 193,
"total_duration_us": 386.0,
"unique_shape_variants": 3,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 5120)",
"(5120)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 16, 5120)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 4, 128)",
"(128)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 16, 4, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 1, 128)",
"(128)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 16, 1, 128)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "RmsNorm",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.add_rms_norm2.default",
"count": 64,
"total_duration_us": 192.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 5120)",
"(16, 5120)",
"(5120)"
],
"input_dtypes": [
"float16",
"float16",
"float16"
],
"output_shapes": [
"(1, 16, 5120)",
"(1, 16, 5120)"
],
"output_dtypes": [
"float16",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "AddRmsNorm",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.apply_rope.default",
"count": 64,
"total_duration_us": 128.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 1, 16, 128)",
"(1, 4, 16, 128)",
"(1, 16, 128)",
"(1, 16, 128)"
],
"input_dtypes": [
"float16",
"float16",
"float16",
"float16"
],
"output_shapes": [
"(1, 16, 1, 128)",
"(1, 16, 4, 128)"
],
"output_dtypes": [
"float16",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "InterleaveRope",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.reshape_and_cache.default",
"count": 64,
"total_duration_us": 128.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(16, 128)",
"(16, 128)",
"(2, 513, 128, 1, 128)",
"(16)"
],
"input_dtypes": [
"float16",
"float16",
"float16",
"int64"
],
"output_shapes": [],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "ReshapeAndCacheNdKernel",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.swiglu.default",
"count": 64,
"total_duration_us": 128.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 1600)",
"(1, 16, 1600)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 16, 1600)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "SwiGlu",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.add.Tensor",
"count": 64,
"total_duration_us": 128.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 5120)",
"(16, 5120)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 16, 5120)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "Add",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.copy_.default",
"count": 1,
"total_duration_us": 98.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(2, 513, 128, 1, 128)",
"(2, 513, 128, 1, 128)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(2, 513, 128, 1, 128)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "TensorMove",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.all_gather.default",
"count": 1,
"total_duration_us": 45.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 9496)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 151936)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "hcom_allGather_",
"zero_cost": false,
"composite": false,
"category": "communication",
"sub_kernels": null
},
{
"op_name": "aten.index.Tensor",
"count": 2,
"total_duration_us": 24.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(40960, 256)",
"(16)"
],
"input_dtypes": [
"float16",
"int64"
],
"output_shapes": [
"(16, 256)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 5120)",
"(16)"
],
"input_dtypes": [
"float16",
"int64"
],
"output_shapes": [
"(1, 16, 5120)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "Index",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.all_reduce.default",
"count": 1,
"total_duration_us": 16.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 5120)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 5120)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "hcom_allReduce_",
"zero_cost": false,
"composite": false,
"category": "communication",
"sub_kernels": null
},
{
"op_name": "aten.ge.Scalar",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(1, 16)"
],
"output_dtypes": [
"bool"
]
}
],
"is_mapped": false,
"kernel_type": null,
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.lt.Scalar",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(1, 16)"
],
"output_dtypes": [
"bool"
]
}
],
"is_mapped": false,
"kernel_type": null,
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.bitwise_and.Tensor",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16)",
"(1, 16)"
],
"input_dtypes": [
"bool",
"bool"
],
"output_shapes": [
"(1, 16)"
],
"output_dtypes": [
"bool"
]
}
],
"is_mapped": false,
"kernel_type": null,
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.sub.Tensor",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(1, 16)"
],
"output_dtypes": [
"int64"
]
}
],
"is_mapped": false,
"kernel_type": null,
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.full.default",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [],
"input_dtypes": [],
"output_shapes": [
"(1, 16)"
],
"output_dtypes": [
"int64"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.where.self",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16)",
"(1, 16)",
"(1, 16)"
],
"input_dtypes": [
"bool",
"int64",
"int64"
],
"output_shapes": [
"(1, 16)"
],
"output_dtypes": [
"int64"
]
}
],
"is_mapped": true,
"kernel_type": "SelectV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.embedding.default",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(9496, 5120)",
"(1, 16)"
],
"input_dtypes": [
"float16",
"int64"
],
"output_shapes": [
"(1, 16, 5120)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "GatherV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "prims.convert_element_type.default",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 1)"
],
"input_dtypes": [
"bool"
],
"output_shapes": [
"(1, 16, 1)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.mul.Tensor",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 5120)",
"(1, 16, 1)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 16, 5120)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "Mul",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.unsqueeze.default",
"count": 1,
"total_duration_us": 0.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16)"
],
"input_dtypes": [
"bool"
],
"output_shapes": [
"(1, 16, 1)"
],
"output_dtypes": [
"bool"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.view.default",
"count": 1028,
"total_duration_us": 0.0,
"unique_shape_variants": 13,
"shape_variants": [
{
"input_shapes": [
"(1, 16)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(16)"
],
"output_dtypes": [
"int64"
]
},
{
"input_shapes": [
"(16, 256)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 256)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 5120)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 5120)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 512)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 512)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 512)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 4, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 1, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 4, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 512)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 1, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 512)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 512)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 1600)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 1600)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 1600)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 1600)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 9496)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 9496)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.split.Tensor",
"count": 1,
"total_duration_us": 0.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 256)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 128)",
"(1, 16, 128)"
],
"output_dtypes": [
"float16",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.split_with_sizes.default",
"count": 128,
"total_duration_us": 0.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(16, 768)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 512)",
"(16, 128)",
"(16, 128)"
],
"output_dtypes": [
"float16",
"float16",
"float16"
]
},
{
"input_shapes": [
"(16, 3200)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 1600)",
"(16, 1600)"
],
"output_dtypes": [
"float16",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.permute.default",
"count": 256,
"total_duration_us": 0.0,
"unique_shape_variants": 3,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 4, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 4, 16, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16, 1, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 1, 16, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 1, 16, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 1, 128)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.select.int",
"count": 128,
"total_duration_us": 0.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(2, 513, 128, 1, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(513, 128, 1, 128)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.slice.Tensor",
"count": 1,
"total_duration_us": 0.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 5120)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 5120)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.alias.default",
"count": 1,
"total_duration_us": 0.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 16, 151936)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 16, 151936)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
}
]
}
