qat量化精度异常，是否是量化逻辑问题

oe3.5.0

实验超参: calibration_step: 500, observer: mse

calib_float是calibration与浮点对比分析结果，qua_float是qat与浮点对比分析结果

abnormal_layer_advisor 、compare_per_layer_out

超大数据范围"Total data range 130925.09375 maybe too large for quantization."

目前对应此信息，修改的是 hat/models/task_modules/sparsebevoe/blocks.py

"""

DeformableFeatureAggregationOE / project_points函数

depth_clamped = torch.clamp(points_2d[..., 2:3], min=0.01)

depth = self.reciprocal_op(depth_clamped)

depth = torch.clamp(depth, max=100.0)

xy = points_2d[..., :2]

points_2d = self.point_mul.mul(xy, depth)

points_2d = torch.clamp(points_2d, -1.1, 1.1)

"""

另外set_qconfig函数内也做了对应的修改，看表现存在没有生效的模块，是什么原因呢，是配置的地方不对还是整体解决思路有问题

"Current scale does not cover the data range" 上游的巨大误差传播到下游，导致会发生截断等操作，感觉根因还是上述配置没有生效导致的