DeepSeek-V3.2-Exp · Attention input

61 transformer blocks › 58 MoE blocks › Sparse latent attention › Query compression

Attention input

tensorx · 545

The normalized features entering attention.

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

Sources

Contextual evidence from Attention RMSNorm

RMSNorm owns one gain per feature and computes in float32; with a residual argument it first adds the residual, fusing the residual addition into the next normalization.

rmsnorm · deepseek-ai/DeepSeek-V3.2-Exp / inference/model.py
Pinned revision 87e509a2e5a1 ↗ · lines 280–306

Read code excerpt

    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.dim = dim
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim, dtype=torch.float32))

    def forward(self, x: torch.Tensor, residual: Optional[torch.Tensor] = None):
        """
        Forward pass for RMSNorm.

        Args:
            x (torch.Tensor): Input tensor.

        Returns:
            torch.Tensor: Normalized tensor with the same shape as input.
        """
        dtype = x.dtype
        if residual is None:
            x = x.float()
            var = x.pow(2).mean(-1, keepdim=True)
            x = x * torch.rsqrt(var + self.eps)
            return (self.weight * x).to(dtype)
        else:
            x = residual = x.float() + residual.float()
            var = x.pow(2).mean(-1, keepdim=True)
            x = x * torch.rsqrt(var + self.eps)
            return (self.weight * x).to(dtype), residual.to(dtype)

Contextual evidence from Attention RMSNorm

A block applies attention-norm, MLA, ffn-norm and the FFN; the first n_dense_layers blocks use the dense MLP, the rest MoE. Residual additions happen inside the next normalization.

block · deepseek-ai/DeepSeek-V3.2-Exp / inference/model.py
Pinned revision 87e509a2e5a1 ↗ · lines 826–851

Read code excerpt

        self.attn = MLA(args)
        self.ffn = MLP(args.dim, args.inter_dim) if layer_id < args.n_dense_layers else MoE(args)
        self.attn_norm = RMSNorm(args.dim)
        self.ffn_norm = RMSNorm(args.dim)

… lines 830–843 omitted …

        if residual is None:
            x, residual = self.attn_norm(x), x
        else:
            x, residual = self.attn_norm(x, residual)
        x = self.attn(x, start_pos, freqs_cis, mask)
        x, residual = self.ffn_norm(x, residual)
        x = self.ffn(x)
        return x, residual

Contextual evidence from Query down-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 · deepseek-ai/DeepSeek-V3.2-Exp / inference/model.py
Pinned revision 87e509a2e5a1 ↗ · lines 523–539

Read code excerpt

        self.qk_head_dim = args.qk_nope_head_dim + args.qk_rope_head_dim
        self.v_head_dim = args.v_head_dim

        self.wq_a = Linear(self.dim, self.q_lora_rank)
        self.q_norm = RMSNorm(self.q_lora_rank)
        self.wq_b = ColumnParallelLinear(self.q_lora_rank, self.n_heads * self.qk_head_dim)
        self.wkv_a = Linear(self.dim, self.kv_lora_rank + self.qk_rope_head_dim)
        self.kv_norm = RMSNorm(self.kv_lora_rank)
        self.wkv_b = ColumnParallelLinear(self.kv_lora_rank, self.n_heads * (self.qk_nope_head_dim + self.v_head_dim))
        self.wo = RowParallelLinear(self.n_heads * self.v_head_dim, self.dim)
        self.softmax_scale = self.qk_head_dim ** -0.5
        self.scale_fmt = args.scale_fmt
        if args.max_seq_len > args.original_seq_len:
            mscale = 0.1 * args.mscale * math.log(args.rope_factor) + 1.0
            self.softmax_scale = self.softmax_scale * mscale * mscale

        self.indexer = Indexer(args)

Contextual evidence from Query down-projection

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

mla-queries · deepseek-ai/DeepSeek-V3.2-Exp / inference/model.py
Pinned revision 87e509a2e5a1 ↗ · lines 560–564

Read code excerpt

        qr = self.q_norm(self.wq_a(x))
        q = self.wq_b(qr)
        q = q.view(bsz, seqlen, self.n_local_heads, self.qk_head_dim)
        q_nope, q_pe = torch.split(q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
        q_pe = apply_rotary_emb(q_pe, freqs_cis)

Contextual evidence from Query down-projection

Linear layers have no bias unless asked. FP8 weights carry one float32 scale per 128 × 128 block; the scales are quantization metadata and are excluded from parameter counts.

linear · deepseek-ai/DeepSeek-V3.2-Exp / inference/model.py
Pinned revision 87e509a2e5a1 ↗ · lines 15–193

Read code excerpt

block_size = 128

… lines 16–178 omitted …

    def __init__(self, in_features: int, out_features: int, bias: bool = False, dtype = None):
        super().__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.weight = nn.Parameter(torch.empty(out_features, in_features, dtype=dtype or Linear.dtype))
        if self.weight.element_size() == 1:
            scale_out_features = (out_features + block_size - 1) // block_size
            scale_in_features = (in_features + block_size - 1) // block_size
            self.weight.scale = self.scale = nn.Parameter(torch.empty(scale_out_features, scale_in_features, dtype=torch.float32))
        else:
            self.register_parameter("scale", None)
        if bias:
            self.bias = nn.Parameter(torch.empty(out_features))
        else:
            self.register_parameter("bias", None)