weight2metis scale is zero if there are identical weight values

[johanesmikhael]

The weight2metis function may encounter issues when min_diff == 0, which can result from consecutive identical values in the input tensor. Is it the expected behavior?

[johanesmikhael]

I propose below change:

def weight2metis(weight):
    sorted_weight = weight.sort()[0].unique() # add unique() to avoid min_diff == 0
    diff = sorted_weight[1:] - sorted_weight[:-1]
    if diff.sum() == 0:
        return None
    weight_min, weight_max = sorted_weight[0], sorted_weight[-1]
    srange = weight_max - weight_min
    min_diff = diff.min()
    scale = (min_diff / srange).item()
    tick, arange = scale.as_integer_ratio()
    weight_ratio = (weight - weight_min).div_(srange).mul_(arange).add_(tick)
    return weight_ratio.to(torch.long)