Kimi K2 · Attention input · OpenModelAtlas

61 transformer blocks › 60 MoE blocks › Latent attention › Key heads

Attention input

supplied · no parametershidden_states

The normalized features entering attention.

shape: $x$ [B batch, 1 positions, 7168 d]

Sources

Contextual evidence from Attention RMSNorm

RMSNorm owns one gain per feature and computes in float32.

rmsnorm · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 94–111

Read code excerpt

class DeepseekV3RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        DeepseekV3RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)


ALL_LAYERNORM_LAYERS.append(DeepseekV3RMSNorm)

Contextual evidence from Attention RMSNorm

A block applies input RMSNorm, attention and a residual addition, then post-attention RMSNorm, the FFN and a second residual addition; the first first_k_dense_replace blocks use the dense MLP, the rest MoE.

block · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 1152–1218

Read code excerpt

        self.mlp = (
            DeepseekV3MoE(config)
            if (
                config.n_routed_experts is not None
                and layer_idx >= config.first_k_dense_replace
                and layer_idx % config.moe_layer_freq == 0
            )
            else DeepseekV3MLP(config)
        )
        self.input_layernorm = DeepseekV3RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )
        self.post_attention_layernorm = DeepseekV3RMSNorm(
            config.hidden_size, eps=config.rms_norm_eps
        )

… lines 1167–1197 omitted …

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

Contextual evidence from Rotary key projection

MLA owns a query down-projection, query RMSNorm and up-projection, a joint key–value down-projection that also emits the shared rotary key, a latent RMSNorm, a key–value up-projection and an output projection.

mla-init · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 662–688

Read code excerpt

            self.q_a_proj = nn.Linear(
                self.hidden_size, config.q_lora_rank, bias=config.attention_bias
            )
            self.q_a_layernorm = DeepseekV3RMSNorm(config.q_lora_rank)
            self.q_b_proj = nn.Linear(
                config.q_lora_rank, self.num_heads * self.q_head_dim, bias=False
            )

        self.kv_a_proj_with_mqa = nn.Linear(
            self.hidden_size,
            config.kv_lora_rank + config.qk_rope_head_dim,
            bias=config.attention_bias,
        )
        self.kv_a_layernorm = DeepseekV3RMSNorm(config.kv_lora_rank)
        self.kv_b_proj = nn.Linear(
            config.kv_lora_rank,
            self.num_heads
            * (self.q_head_dim - self.qk_rope_head_dim + self.v_head_dim),
            bias=False,
        )

        self.o_proj = nn.Linear(
            self.num_heads * self.v_head_dim,
            self.hidden_size,
            bias=config.attention_bias,
        )
        self._init_rope()

Contextual evidence from Rotary key projection

The joint down-projection yields the key–value latent and one rotary key shared by all heads; the latent is normalized and up-projected into per-head content keys and values.

mla-kv · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 776–789

Read code excerpt

        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
        compressed_kv, k_pe = torch.split(
            compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
        )
        k_pe = k_pe.view(bsz, q_len, 1, self.qk_rope_head_dim).transpose(1, 2)
        kv = (
            self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
            .view(bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
            .transpose(1, 2)
        )

        k_nope, value_states = torch.split(
            kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1
        )