Skip to content

聚合问题 #3

@Wyhls

Description

@Wyhls

“ours” 方法聚合时残差并没有使用SVD分解,此时传输数据量不是和全量参数传输一样吗?您提供的聚合代码如下:
residue = M - lora_B_avg @ lora_A_avg

global_dict[name + ".lora_A.default.weight"] = lora_A_avg
global_dict[name + ".lora_B.default.weight"] = lora_B_avg
global_dict[name + ".base_layer.weight"] += torch.transpose(residue * scaling_factor, 1, 0)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions