{
"model": "deepseek-ai/DeepSeek-V3",
"scenario": "prefill",
"total_ops": 18360,
"unique_ops": 46,
"mapped_count": 40,
"unmapped_count": 6,
"unmapped_ops": [
"aten.bitwise_and.Tensor",
"aten.constant_pad_nd.default",
"aten.copy.default",
"aten.ge.Scalar",
"aten.lt.Scalar",
"aten.sub.Tensor"
],
"mapped_ops": [
"aten.add.Tensor",
"aten.alias.default",
"aten.bitwise_not.default",
"aten.cat.default",
"aten.clone.default",
"aten.copy_.default",
"aten.div.Tensor",
"aten.embedding.default",
"aten.expand.default",
"aten.full.default",
"aten.gather.default",
"aten.index.Tensor",
"aten.mm.default",
"aten.mul.Tensor",
"aten.scatter.value",
"aten.sigmoid.default",
"aten.slice.Tensor",
"aten.split.Tensor",
"aten.split_with_sizes.default",
"aten.sum.dim_IntList",
"aten.topk.default",
"aten.unsqueeze.default",
"aten.view.default",
"aten.where.self",
"prims.convert_element_type.default",
"tensor_cast.add_rms_norm2.default",
"tensor_cast.add_rms_norm_quant2.default",
"tensor_cast.all_gather.default",
"tensor_cast.all_reduce.default",
"tensor_cast.all_to_all.default",
"tensor_cast.concat_and_cache_mla.default",
"tensor_cast.mlapo_quant.default",
"tensor_cast.multihead_latent_attention.default",
"tensor_cast.init_routing_v2.default",
"tensor_cast.quantize.default",
"tensor_cast.rms_norm.default",
"tensor_cast.static_quant_linear.default",
"tensor_cast.static_quant_linear_all_reduce.default",
"tensor_cast.swiglu.default",
"tensor_cast.unpermute_tokens.default"
],
"ops": [
{
"op_name": "tensor_cast.all_gather.default",
"count": 2843,
"total_duration_us": 60739.0,
"unique_shape_variants": 4,
"shape_variants": [
{
"input_shapes": [
"(16, 2048)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(128, 2048)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(128, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(2048, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 1, 16160)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 1, 129280)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "hcom_allGather_",
"zero_cost": false,
"composite": false,
"category": "communication",
"sub_kernels": null
},
{
"op_name": "tensor_cast.static_quant_linear.default",
"count": 1975,
"total_duration_us": 51856.0,
"unique_shape_variants": 5,
"shape_variants": [
{
"input_shapes": [
"(256, 7168)",
"(7168, 4608)"
],
"input_dtypes": [
"int8",
"int8"
],
"output_shapes": [
"(256, 4608)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 7168)",
"(7168, 4096)"
],
"input_dtypes": [
"int8",
"int8"
],
"output_shapes": [
"(16, 4096)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(16, 2048)",
"(2048, 7168)"
],
"input_dtypes": [
"int8",
"int8"
],
"output_shapes": [
"(16, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 7168)",
"(7168, 4096)"
],
"input_dtypes": [
"int8",
"int8"
],
"output_shapes": [
"(256, 4096)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 2048)",
"(2048, 7168)"
],
"input_dtypes": [
"int8",
"int8"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "QuantBatchMatmulV3",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.all_to_all.default",
"count": 116,
"total_duration_us": 26216.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(2048, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(2048, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "hcom_alltoallv_",
"zero_cost": false,
"composite": false,
"category": "communication",
"sub_kernels": null
},
{
"op_name": "aten.mul.Tensor",
"count": 117,
"total_duration_us": 5055.0,
"unique_shape_variants": 3,
"shape_variants": [
{
"input_shapes": [
"(1, 256, 7168)",
"(1, 256, 1)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 256, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 8)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(256, 8, 7168)",
"(256, 8, 1)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 8, 7168)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "Mul",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.quantize.default",
"count": 2036,
"total_duration_us": 4487.0,
"unique_shape_variants": 6,
"shape_variants": [
{
"input_shapes": [
"(256, 16, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 16, 128)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(1, 256, 2304)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 256, 2304)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(16, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 7168)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(16, 2048)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 2048)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(256, 2048)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 2048)"
],
"output_dtypes": [
"int8"
]
}
],
"is_mapped": true,
"kernel_type": "AscendQuantV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.static_quant_linear_all_reduce.default",
"count": 64,
"total_duration_us": 4096.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(256, 2048)",
"(2048, 7168)"
],
"input_dtypes": [
"int8",
"int8"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 2304)",
"(2304, 7168)"
],
"input_dtypes": [
"int8",
"int8"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": false,
"composite": true,
"category": null,
"sub_kernels": [
"QuantBatchMatmulV3",
"hcom_allReduce_"
]
},
{
"op_name": "aten.sum.dim_IntList",
"count": 174,
"total_duration_us": 4002.0,
"unique_shape_variants": 3,
"shape_variants": [
{
"input_shapes": [
"(256, 8, 2)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(256, 8)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 1)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(256, 8, 7168)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "ReduceSumD",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.cat.default",
"count": 58,
"total_duration_us": 3364.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)"
],
"input_dtypes": [
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16"
],
"output_shapes": [
"(2048, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "ConcatD",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.unpermute_tokens.default",
"count": 58,
"total_duration_us": 3364.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(2048, 7168)",
"(256, 8)"
],
"input_dtypes": [
"float16",
"int64"
],
"output_shapes": [
"(256, 8, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "MoeDistributeCombineV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.swiglu.default",
"count": 989,
"total_duration_us": 3089.0,
"unique_shape_variants": 3,
"shape_variants": [
{
"input_shapes": [
"(1, 256, 2304)",
"(1, 256, 2304)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 256, 2304)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(128, 2048)",
"(128, 2048)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(128, 2048)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 2048)",
"(256, 2048)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(256, 2048)"
],
"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": 235,
"total_duration_us": 2066.0,
"unique_shape_variants": 5,
"shape_variants": [
{
"input_shapes": [
"(1, 256, 7168)",
"(256, 7168)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 256, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 256)",
"(1, 256)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 256)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(256, 1)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 1)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(256, 7168)",
"(256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(1, 256, 7168)",
"(1, 256, 7168)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 256, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "Add",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.constant_pad_nd.default",
"count": 58,
"total_duration_us": 1914.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(2048, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": false,
"kernel_type": null,
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.init_routing_v2.default",
"count": 58,
"total_duration_us": 1914.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 7168)",
"(256, 8)"
],
"input_dtypes": [
"float16",
"int64"
],
"output_shapes": [
"(2048, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "MoeDistributeDispatchV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.copy.default",
"count": 928,
"total_duration_us": 1856.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(16, 2048)",
"(16, 2048)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(16, 2048)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": false,
"kernel_type": null,
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.mlapo_quant.default",
"count": 61,
"total_duration_us": 1586.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 7168)",
"(1, 256, 64)",
"(1, 256, 64)",
"(1536, 7168)",
"(1536)",
"(384, 1536)",
"(576, 7168)",
"(512)"
],
"input_dtypes": [
"float16",
"float16",
"float16",
"int8",
"float16",
"int8",
"int8",
"float16"
],
"output_shapes": [
"(256, 16, 192)",
"(256, 512)",
"(256, 64)"
],
"output_dtypes": [
"float16",
"float16",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": false,
"composite": true,
"category": null,
"sub_kernels": [
"QuantBatchMatmulV3",
"KvRmsNormRopeCache"
]
},
{
"op_name": "prims.convert_element_type.default",
"count": 175,
"total_duration_us": 1510.0,
"unique_shape_variants": 4,
"shape_variants": [
{
"input_shapes": [
"(1, 256, 1)"
],
"input_dtypes": [
"bool"
],
"output_shapes": [
"(1, 256, 1)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(256, 256)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 256)"
],
"output_dtypes": [
"bool"
]
},
{
"input_shapes": [
"(256, 7168)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.mm.default",
"count": 59,
"total_duration_us": 1327.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(256, 7168)",
"(7168, 256)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 256)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(1, 7168)",
"(7168, 16160)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 16160)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "MatMulV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.add_rms_norm2.default",
"count": 58,
"total_duration_us": 928.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 256, 7168)",
"(256, 7168)",
"(7168)"
],
"input_dtypes": [
"float16",
"float16",
"float16"
],
"output_shapes": [
"(1, 256, 7168)",
"(1, 256, 7168)"
],
"output_dtypes": [
"float16",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "AddRmsNorm",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.multihead_latent_attention.default",
"count": 61,
"total_duration_us": 732.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 16, 192)",
"(3, 128, 576)",
"(1, 2)",
"(16, 128, 512)",
"(16, 512, 128)",
"(512, 4096)"
],
"input_dtypes": [
"float16",
"float16",
"int64",
"float16",
"float16",
"float16"
],
"output_shapes": [
"(256, 16, 128)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": false,
"composite": true,
"category": null,
"sub_kernels": [
"TransposeBatchMatMul",
"FusedInferAttentionScore"
]
},
{
"op_name": "tensor_cast.rms_norm.default",
"count": 62,
"total_duration_us": 558.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 256, 7168)",
"(7168)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(1, 256, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "RmsNorm",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.topk.default",
"count": 174,
"total_duration_us": 348.0,
"unique_shape_variants": 3,
"shape_variants": [
{
"input_shapes": [
"(256, 8, 32)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8, 2)",
"(256, 8, 2)"
],
"output_dtypes": [
"int64"
]
},
{
"input_shapes": [
"(256, 8)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 4)",
"(256, 4)"
],
"output_dtypes": [
"int64"
]
},
{
"input_shapes": [
"(256, 256)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8)",
"(256, 8)"
],
"output_dtypes": [
"int64"
]
}
],
"is_mapped": true,
"kernel_type": "TopKV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.full.default",
"count": 117,
"total_duration_us": 234.0,
"unique_shape_variants": 3,
"shape_variants": [
{
"input_shapes": [],
"input_dtypes": [],
"output_shapes": [
"(1, 256)"
],
"output_dtypes": [
"int64"
]
},
{
"input_shapes": [],
"input_dtypes": [],
"output_shapes": [
"(256, 8)"
],
"output_dtypes": []
},
{
"input_shapes": [],
"input_dtypes": [],
"output_shapes": [
"(16, 2048)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.concat_and_cache_mla.default",
"count": 61,
"total_duration_us": 183.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 512)",
"(256, 64)",
"(3, 128, 576)",
"(256)"
],
"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": "aten.where.self",
"count": 59,
"total_duration_us": 176.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(1, 256)",
"(1, 256)",
"(1, 256)"
],
"input_dtypes": [
"bool",
"int64",
"int64"
],
"output_shapes": [
"(1, 256)"
],
"output_dtypes": [
"int64"
]
},
{
"input_shapes": [
"(256, 256)",
"()",
"(256, 256)"
],
"input_dtypes": [
"bool"
],
"output_shapes": [
"(256, 256)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "SelectV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.sigmoid.default",
"count": 58,
"total_duration_us": 116.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 256)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 256)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "Sigmoid",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.scatter.value",
"count": 58,
"total_duration_us": 116.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 8)",
"(256, 4)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(256, 8)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "ScatterElements",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.clone.default",
"count": 58,
"total_duration_us": 116.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 8, 32)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8, 32)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "TensorMove",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.bitwise_not.default",
"count": 58,
"total_duration_us": 116.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 256)"
],
"input_dtypes": [
"bool"
],
"output_shapes": [
"(256, 256)"
],
"output_dtypes": [
"bool"
]
}
],
"is_mapped": true,
"kernel_type": "BitwiseNot",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.gather.default",
"count": 58,
"total_duration_us": 116.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 256)",
"(256, 8)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(256, 8)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "GatherV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.div.Tensor",
"count": 58,
"total_duration_us": 116.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 8)",
"(256, 1)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": "Div",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.all_reduce.default",
"count": 1,
"total_duration_us": 64.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 256, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "hcom_allReduce_",
"zero_cost": false,
"composite": false,
"category": "communication",
"sub_kernels": null
},
{
"op_name": "aten.index.Tensor",
"count": 2,
"total_duration_us": 47.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(163840, 128)",
"(256)"
],
"input_dtypes": [
"float16",
"int64"
],
"output_shapes": [
"(256, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 256, 7168)",
"(1)"
],
"input_dtypes": [
"float16",
"int64"
],
"output_shapes": [
"(1, 1, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "Index",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "tensor_cast.add_rms_norm_quant2.default",
"count": 3,
"total_duration_us": 42.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 256, 7168)",
"(256, 7168)",
"(7168)"
],
"input_dtypes": [
"float16",
"float16",
"float16"
],
"output_shapes": [
"(1, 256, 7168)",
"(1, 256, 7168)"
],
"output_dtypes": [
"int8",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "AddRmsNormQuant",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.embedding.default",
"count": 1,
"total_duration_us": 9.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(16160, 7168)",
"(1, 256)"
],
"input_dtypes": [
"float16",
"int64"
],
"output_shapes": [
"(1, 256, 7168)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "GatherV2",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.copy_.default",
"count": 2,
"total_duration_us": 6.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(3, 128, 576)",
"(3, 128, 576)"
],
"input_dtypes": [
"float16",
"float16"
],
"output_shapes": [
"(3, 128, 576)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": "TensorMove",
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.ge.Scalar",
"count": 1,
"total_duration_us": 2.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(1, 256)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(1, 256)"
],
"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, 256)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(1, 256)"
],
"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, 256)",
"(1, 256)"
],
"input_dtypes": [
"bool",
"bool"
],
"output_shapes": [
"(1, 256)"
],
"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, 256)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(1, 256)"
],
"output_dtypes": [
"int64"
]
}
],
"is_mapped": false,
"kernel_type": null,
"zero_cost": false,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.unsqueeze.default",
"count": 117,
"total_duration_us": 0.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(1, 256)"
],
"input_dtypes": [
"bool"
],
"output_shapes": [
"(1, 256, 1)"
],
"output_dtypes": [
"bool"
]
},
{
"input_shapes": [
"(256, 8)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8, 1)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.view.default",
"count": 431,
"total_duration_us": 0.0,
"unique_shape_variants": 13,
"shape_variants": [
{
"input_shapes": [
"(1, 256)"
],
"input_dtypes": [
"int64"
],
"output_shapes": [
"(256)"
],
"output_dtypes": [
"int64"
]
},
{
"input_shapes": [
"(256, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 256, 128)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(256, 16, 128)"
],
"input_dtypes": [
"int8"
],
"output_shapes": [
"(1, 256, 2048)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(1, 256, 2048)"
],
"input_dtypes": [
"int8"
],
"output_shapes": [
"(256, 2048)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(1, 256, 7168)"
],
"input_dtypes": [
"int8"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(256, 2304)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 256, 2304)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 256, 2304)"
],
"input_dtypes": [
"int8"
],
"output_shapes": [
"(256, 2304)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(256, 256)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8, 32)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(256, 8, 32)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 256)"
],
"output_dtypes": []
},
{
"input_shapes": [
"(256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 256, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 1, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 16160)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 1, 16160)"
],
"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, 256, 128)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 256, 64)",
"(1, 256, 64)"
],
"output_dtypes": [
"float16",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.slice.Tensor",
"count": 2036,
"total_duration_us": 0.0,
"unique_shape_variants": 5,
"shape_variants": [
{
"input_shapes": [
"(3072, 1536)"
],
"input_dtypes": [
"int8"
],
"output_shapes": [
"(384, 1536)"
],
"output_dtypes": [
"int8"
]
},
{
"input_shapes": [
"(2048, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(128, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 7168)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(128, 2048)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 2048)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 256, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 256, 7168)"
],
"output_dtypes": [
"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": 1975,
"total_duration_us": 0.0,
"unique_shape_variants": 5,
"shape_variants": [
{
"input_shapes": [
"(256, 4608)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 2304)",
"(256, 2304)"
],
"output_dtypes": [
"float16",
"float16"
]
},
{
"input_shapes": [
"(2048, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)",
"(128, 7168)"
],
"output_dtypes": [
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16"
]
},
{
"input_shapes": [
"(16, 4096)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 2048)",
"(16, 2048)"
],
"output_dtypes": [
"float16",
"float16"
]
},
{
"input_shapes": [
"(128, 7168)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)",
"(8, 7168)"
],
"output_dtypes": [
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16",
"float16"
]
},
{
"input_shapes": [
"(256, 4096)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(256, 2048)",
"(256, 2048)"
],
"output_dtypes": [
"float16",
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.expand.default",
"count": 58,
"total_duration_us": 0.0,
"unique_shape_variants": 1,
"shape_variants": [
{
"input_shapes": [
"(256, 8, 1)"
],
"input_dtypes": [],
"output_shapes": [
"(256, 8, 32)"
],
"output_dtypes": []
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
},
{
"op_name": "aten.alias.default",
"count": 2785,
"total_duration_us": 0.0,
"unique_shape_variants": 2,
"shape_variants": [
{
"input_shapes": [
"(16, 2048)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(16, 2048)"
],
"output_dtypes": [
"float16"
]
},
{
"input_shapes": [
"(1, 1, 129280)"
],
"input_dtypes": [
"float16"
],
"output_shapes": [
"(1, 1, 129280)"
],
"output_dtypes": [
"float16"
]
}
],
"is_mapped": true,
"kernel_type": null,
"zero_cost": true,
"composite": false,
"category": null,
"sub_kernels": null
}
]
}
