Kimi K2 · Add shared output

61 transformer blocks › 60 MoE blocks › Mixture of experts

Add shared output

no parametersy + self.shared_experts(identity)

Add the shared expert’s output to the routed output.

{\htmlData{light=output add}{\Delta_{\mathrm{ffn}}}}={\htmlData{light=routed combine}{y^{\mathrm{routed}}}}+{\htmlData{light=shared shared-expert}{y^{\mathrm{shared}}}}

in: $y^{\mathrm{routed}}$ [B batch, T positions, 7168 d] $y^{\mathrm{shared}}$ [B batch, T positions, 7168 d]
out: $\Delta_{\mathrm{ffn}}$ [B batch, T positions, 7168 d]

Codemodeling_deepseek.py 523–533, 536–609

modeling_deepseek.py 523–533

    def forward(self, hidden_states):
        identity = hidden_states
        orig_shape = hidden_states.shape
        topk_idx, topk_weight = self.gate(hidden_states)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        flat_topk_idx = topk_idx.view(-1)
        if not self.training:
            y = self.moe_infer(hidden_states, topk_idx, topk_weight).view(*orig_shape)
        if self.config.n_shared_experts is not None:
            y = y + self.shared_experts(identity)
        return y

modeling_deepseek.py 536–609

    def moe_infer(self, x, topk_ids, topk_weight):
        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
        cnts.scatter_(1, topk_ids, 1)
        tokens_per_expert = cnts.sum(dim=0)
        idxs = topk_ids.view(-1).argsort()
        sorted_tokens = x[idxs // topk_ids.shape[1]]
        sorted_tokens_shape = sorted_tokens.shape
        if self.ep_size > 1:
            tokens_per_ep_rank = tokens_per_expert.view(self.ep_size, -1).sum(dim=1)
            tokens_per_expert_group = tokens_per_expert.new_empty(
                tokens_per_expert.shape[0]
            )
            dist.all_to_all_single(tokens_per_expert_group, tokens_per_expert)
            output_splits = (
                tokens_per_expert_group.view(self.ep_size, -1)
                .sum(1)
                .cpu()
                .numpy()
                .tolist()
            )
            gathered_tokens = sorted_tokens.new_empty(
                tokens_per_expert_group.sum(dim=0).cpu().item(), sorted_tokens.shape[1]
            )
            input_split_sizes = tokens_per_ep_rank.cpu().numpy().tolist()
            dist.all_to_all(
                list(gathered_tokens.split(output_splits)),
                list(sorted_tokens.split(input_split_sizes)),
            )
            tokens_per_expert_post_gather = tokens_per_expert_group.view(
                self.ep_size, self.experts_per_rank
            ).sum(dim=0)
            gatherd_idxs = np.zeros(shape=(gathered_tokens.shape[0],), dtype=np.int32)
            s = 0
            for i, k in enumerate(tokens_per_expert_group.cpu().numpy()):
                gatherd_idxs[s : s + k] = i % self.experts_per_rank
                s += k
            gatherd_idxs = gatherd_idxs.argsort()
            sorted_tokens = gathered_tokens[gatherd_idxs]
            tokens_per_expert = tokens_per_expert_post_gather
        tokens_per_expert = tokens_per_expert.cpu().numpy()

        outputs = []
        start_idx = 0
        for i, num_tokens in enumerate(tokens_per_expert):
            end_idx = start_idx + num_tokens
            if num_tokens == 0:
                continue
            expert = self.experts[i + self.ep_rank * self.experts_per_rank]
            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
            expert_out = expert(tokens_for_this_expert)
            outputs.append(expert_out)
            start_idx = end_idx

        outs = torch.cat(outputs, dim=0) if len(outputs) else sorted_tokens.new_empty(0)
        if self.ep_size > 1:
            new_x = torch.empty_like(outs)
            new_x[gatherd_idxs] = outs
            gathered_tokens = new_x.new_empty(*sorted_tokens_shape)
            dist.all_to_all(
                list(gathered_tokens.split(input_split_sizes)),
                list(new_x.split(output_splits)),
            )
            outs = gathered_tokens

        new_x = torch.empty_like(outs)
        new_x[idxs] = outs
        final_out = (
            new_x.view(*topk_ids.shape, -1)
            .type(topk_weight.dtype)
            .mul_(topk_weight.unsqueeze(dim=-1))
            .sum(dim=1)
            .type(new_x.dtype)
        )
        return final_out

Sources1

Direct evidence

Each token’s output is the router-weighted sum of its selected experts plus the shared expert applied to every token; tokens are sorted by expert, each expert runs on its tokens, and outputs are scattered back.

moe-forward · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 523–609

Read code excerpt

    def forward(self, hidden_states):
        identity = hidden_states
        orig_shape = hidden_states.shape
        topk_idx, topk_weight = self.gate(hidden_states)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        flat_topk_idx = topk_idx.view(-1)
        if not self.training:
            y = self.moe_infer(hidden_states, topk_idx, topk_weight).view(*orig_shape)
        if self.config.n_shared_experts is not None:
            y = y + self.shared_experts(identity)
        return y

… lines 534–535 omitted …

    def moe_infer(self, x, topk_ids, topk_weight):
        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
        cnts.scatter_(1, topk_ids, 1)
        tokens_per_expert = cnts.sum(dim=0)
        idxs = topk_ids.view(-1).argsort()
        sorted_tokens = x[idxs // topk_ids.shape[1]]
        sorted_tokens_shape = sorted_tokens.shape
        if self.ep_size > 1:
            tokens_per_ep_rank = tokens_per_expert.view(self.ep_size, -1).sum(dim=1)
            tokens_per_expert_group = tokens_per_expert.new_empty(
                tokens_per_expert.shape[0]
            )
            dist.all_to_all_single(tokens_per_expert_group, tokens_per_expert)
            output_splits = (
                tokens_per_expert_group.view(self.ep_size, -1)
                .sum(1)
                .cpu()
                .numpy()
                .tolist()
            )
            gathered_tokens = sorted_tokens.new_empty(
                tokens_per_expert_group.sum(dim=0).cpu().item(), sorted_tokens.shape[1]
            )
            input_split_sizes = tokens_per_ep_rank.cpu().numpy().tolist()
            dist.all_to_all(
                list(gathered_tokens.split(output_splits)),
                list(sorted_tokens.split(input_split_sizes)),
            )
            tokens_per_expert_post_gather = tokens_per_expert_group.view(
                self.ep_size, self.experts_per_rank
            ).sum(dim=0)
            gatherd_idxs = np.zeros(shape=(gathered_tokens.shape[0],), dtype=np.int32)
            s = 0
            for i, k in enumerate(tokens_per_expert_group.cpu().numpy()):
                gatherd_idxs[s : s + k] = i % self.experts_per_rank
                s += k
            gatherd_idxs = gatherd_idxs.argsort()
            sorted_tokens = gathered_tokens[gatherd_idxs]
            tokens_per_expert = tokens_per_expert_post_gather
        tokens_per_expert = tokens_per_expert.cpu().numpy()

        outputs = []
        start_idx = 0
        for i, num_tokens in enumerate(tokens_per_expert):
            end_idx = start_idx + num_tokens
            if num_tokens == 0:
                continue
            expert = self.experts[i + self.ep_rank * self.experts_per_rank]
            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
            expert_out = expert(tokens_for_this_expert)
            outputs.append(expert_out)
            start_idx = end_idx

        outs = torch.cat(outputs, dim=0) if len(outputs) else sorted_tokens.new_empty(0)
        if self.ep_size > 1:
            new_x = torch.empty_like(outs)
            new_x[gatherd_idxs] = outs
            gathered_tokens = new_x.new_empty(*sorted_tokens_shape)
            dist.all_to_all(
                list(gathered_tokens.split(input_split_sizes)),
                list(new_x.split(output_splits)),
            )
            outs = gathered_tokens

        new_x = torch.empty_like(outs)
        new_x[idxs] = outs
        final_out = (
            new_x.view(*topk_ids.shape, -1)
            .type(topk_weight.dtype)
            .mul_(topk_weight.unsqueeze(dim=-1))
            .sum(dim=1)
            .type(new_x.dtype)
        )
        return final_out