Kimi K2 · Query latent RMSNorm

61 transformer blocks › 60 MoE blocks › Latent attention › Query compression

Query latent RMSNorm

learnedself.q_a_layernorm(self.q_a_proj(hidden_states))

Rescale each position’s features using their root mean square, then apply learned gain.

\begin{aligned}m_2(x)&=\frac{1}{d}\sum_{j=1}^d x_j^2\\y&=\gamma_{{\htmlClass{vl-stream-q}{Q}}}\odot\frac{x}{\sqrt{m_2(x)+\varepsilon}}\\\varepsilon&=10^{-6}\end{aligned}

in: $c^{Q}$ [B batch, T positions, 1536 q latent]
out: $\hat c^{Q}$ [B batch, T positions, 1536 q latent]
parameters: 1,536 · × 60 MoE blocks = 92,160

Normalized in float32, cast back to the input type, then multiplied by the gain (modeling_deepseek.py 104–108).

Parameters

1,536 parameters per instance.

Exact per-instance accounting
Component	Calculation	Parameters
Learned gain	$1{,}536$	1,536
Per-instance total		1,536

Each normalization instance owns its learned gain vector.

Across 60 independent MoE blocks: $1{,}536 \times 60 = 92{,}160$ parameters.

Derived from the configured feature width; the model total matches the checkpoint’s tensors.

Codemodeling_deepseek.py 94–774

modeling_deepseek.py 94–111

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)

modeling_deepseek.py 662–688

            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()

modeling_deepseek.py 767–774

        if self.q_lora_rank is None:
            q = self.q_proj(hidden_states)
        else:
            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
        q = q.view(bsz, q_len, self.num_heads, self.q_head_dim).transpose(1, 2)
        q_nope, q_pe = torch.split(
            q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
        )

Sources3

Direct evidence

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)

Direct evidence

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()

Direct evidence

Queries are produced through a low-rank latent and split into a non-rotated and a rotated part.

mla-queries · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 767–774

Read code excerpt

        if self.q_lora_rank is None:
            q = self.q_proj(hidden_states)
        else:
            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
        q = q.view(bsz, q_len, self.num_heads, self.q_head_dim).transpose(1, 2)
        q_nope, q_pe = torch.split(
            q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
        )