Routing

未关闭
#118 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
30/100
Issue 类型
文档
描述清晰度
需要澄清
活跃度
停滞
技术栈
python, pytorch

调研方向

先阅读 issue 中展示的 router 的 forward 方法和链接的 noisy top-k 论文,然后追踪 moe_jitter_eps 如何影响 scores 和 expert selection。记录 input jitter 是否等同于 score noise;当 routing behavior 以及与论文的任何差异都得到清晰解释时,即视为完成。

由索引模型根据 Issue 内容生成。

描述

Is the router implemented the noisy top k routing suggested by the OUTRAGEOUSLY LARGE NEURAL NETWORKS:
THE SPARSELY-GATED MIXTURE-OF-EXPERTS LAYER
paper?

In the router code you seem to apply the noise at the input of the router and not at the router scores like in the paper above:

 def forward(self, x):
        if self.training and self.args.moe_jitter_eps is not None:
            x = x * self.jitter(x)

        scores = self.layer(x.view(-1, x.shape[-1])).softmax(dim=-1)
        expert_weights, expert_indices = self._top_k(scores)
        if self.args.moe_normalize_expert_weights:
            expert_weights = expert_weights / torch.norm(
                expert_weights, p=self.args.moe_normalize_expert_weights,dim=-1, keepdim=True)

        expert_indices = (
            _uniform_expert_assignment(expert_indices, self.args.moe_num_experts)
            if self.args.uniform_expert_assignment else expert_indices
        )
        return scores, expert_weights, expert_indices

In the aforementioned paper the noisy top k works like:
image

Is this somehting equivalent? I am not trying to argue that it is wrong, but i was just trying to figure out if this is the same.

主要语言
Python
星标
1.6k
派生
227
PR 合并指标
30 天内没有已合并 PR

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。