DeepSeek-V3.2-Exp · Rotated shared key

61 transformer blocks › 58 MoE blocks › Sparse latent attention › Key heads

Rotated shared key

tensork_pe · 568

The position-carrying key features every head uses.

shape: $k^{R}$ [B batch, T positions, 64 rotary]

Sources

Contextual evidence from Rotate shared key

Rotary embedding multiplies feature pairs by the coefficients; the indexer uses the non-interleaved layout.

apply-rope · deepseek-ai/DeepSeek-V3.2-Exp / inference/model.py
Pinned revision 87e509a2e5a1 ↗ · lines 405–425

Read code excerpt

def apply_rotary_emb(x: torch.Tensor, freqs_cis: torch.Tensor, interleaved: bool = True) -> torch.Tensor:
    """
    Applies rotary positional embeddings to the input tensor.

    Args:
        x (torch.Tensor): Input tensor with positional embeddings to be applied.
        freqs_cis (torch.Tensor): Precomputed complex exponential values for positional embeddings.

    Returns:
        torch.Tensor: Tensor with rotary embeddings applied.
    """
    dtype = x.dtype
    shape = x.shape
    if not interleaved:
        x = x.view(*shape[:-1], 2, -1).transpose(-1, -2).contiguous()
    x = torch.view_as_complex(x.float().view(*shape[:-1], -1, 2))
    freqs_cis = freqs_cis.view(1, x.size(1), 1, x.size(-1))
    y = torch.view_as_real(x * freqs_cis).flatten(3)
    if not interleaved:
        y = torch.cat([y[..., 0::2], y[..., 1::2]], dim=-1)
    return y.to(dtype)

Contextual evidence from Rotate shared key

Rotary coefficients are precomputed once for the maximum sequence length, with YaRN frequency interpolation beyond the original 4,096 positions.

rope · deepseek-ai/DeepSeek-V3.2-Exp / inference/model.py
Pinned revision 87e509a2e5a1 ↗ · lines 334–402

Read code excerpt

    dim = args.qk_rope_head_dim
    seqlen = args.max_seq_len
    beta_fast = args.beta_fast
    beta_slow = args.beta_slow
    base = args.rope_theta
    factor = args.rope_factor

… lines 340–392 omitted …

    freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
    if seqlen > args.original_seq_len:
        low, high = find_correction_range(beta_fast, beta_slow, dim, base, args.original_seq_len)
        smooth = 1 - linear_ramp_factor(low, high, dim // 2)
        freqs = freqs / factor * (1 - smooth) + freqs * smooth

    t = torch.arange(seqlen)
    freqs = torch.outer(t, freqs)
    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
    return freqs_cis

Contextual evidence from Join key parts

In the full-sequence (prefill) path, the latent is up-projected into per-head keys and values; each head’s key joins its non-rotated part with the shared rotary key; scores are scaled, masked by the indexer’s selection and the causal mask, normalized and used to mix values.

mla-prefill · deepseek-ai/DeepSeek-V3.2-Exp / inference/model.py
Pinned revision 87e509a2e5a1 ↗ · lines 574–589

Read code excerpt

        if mask is not None:    # MHA prefill
            q = torch.cat([q_nope, q_pe], dim=-1)
            kv = self.wkv_b(kv)
            kv = kv.view(bsz, seqlen, self.n_local_heads, self.qk_nope_head_dim + self.v_head_dim)
            k_nope, v = torch.split(kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
            k = torch.cat([k_nope, k_pe.expand(-1, -1, self.n_local_heads, -1)], dim=-1)
            scores = torch.einsum("bshd,bthd->bsht", q, k).mul_(self.softmax_scale)

            # indexer
            topk_indices = self.indexer(x, qr, start_pos, freqs_cis, mask)
            index_mask = torch.full((bsz, seqlen, seqlen), float("-inf"), device=x.device).scatter_(-1, topk_indices, 0)
            index_mask += mask
            scores += index_mask.unsqueeze(2)

            scores = scores.softmax(dim=-1)
            x = torch.einsum("bsht,bthd->bshd", scores, v)