DeepSeek V4.1

class Block(nn.Module):
    """A block whose residual stream is `hc_mult` parallel copies (Hyper-Connections).

    Attention and FFN each sit between `hc_pre` (collapse the copies into one sublayer input) and
    `hc_post` (expand back out, mixing the residual in through `comb`). `hc_mixes` derives all three
    coefficient sets from the stream itself, `comb` made doubly stochastic by Sinkhorn.

    The coefficients a sublayer computes are used by the *next* one -- see `forward`.
    """

    attention_cls = Attention

    def __init__(
        self,
        layer_id: int,
        args: ModelArgs,
        engram_layout: EngramLayout | None = None,
    ):
        super().__init__()
        self.layer_id = layer_id
        self.norm_eps = args.norm_eps
        self.attn = self.attention_cls(layer_id, args)
        self.ffn = MoE(layer_id, args)
        self.engram = None
        if engram_layout is not None and layer_id in engram_layout.layer_ids:
            self.engram = Engram(args, layer_id, engram_layout)
        self.attn_norm = RMSNorm(args.dim, self.norm_eps)
        self.ffn_norm = RMSNorm(args.dim, self.norm_eps)
        self.hc_mult = hc_mult = args.hc_mult
        self.hc_sinkhorn_iters = args.hc_sinkhorn_iters
        self.hc_eps = args.hc_eps
        mix_hc = (2 + hc_mult) * hc_mult
        hc_dim = hc_mult * args.dim
        with set_dtype(torch.float32):
            self.hc_attn_fn = nn.Parameter(torch.empty(mix_hc, hc_dim))
            self.hc_ffn_fn = nn.Parameter(torch.empty(mix_hc, hc_dim))
            self.hc_attn_base = nn.Parameter(torch.empty(mix_hc))
            self.hc_ffn_base = nn.Parameter(torch.empty(mix_hc))
            self.hc_attn_scale = nn.Parameter(torch.empty(3))
            self.hc_ffn_scale = nn.Parameter(torch.empty(3))

    def forward(
        self,
        x: torch.Tensor,
        start_pos: int,
        pre_mix: torch.Tensor,
        image_mask: torch.Tensor | None,
        *attn_args,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """`pre_mix` collapses the hc_mult copies down to one input for this block's attention. Each
        sub-block's own `hc_mixes` produces the mix for the *next* one, so attention uses what the
        previous layer's FFN produced and the FFN uses what this attention produced.

        image_mask: [b, s] bool, True inside image spans (selects the VL routing bias)."""
        residual = x
        attn_pre, attn_post, attn_comb = self.hc_mixes(x, self.hc_attn_fn, self.hc_attn_scale, self.hc_attn_base)
        x = self.hc_pre(x, pre_mix)
        x = self.attn_norm(x)
        x = self.attn(x, start_pos, *attn_args)
        x = self.hc_post(x, residual, attn_post, attn_comb)

        residual = x
        ffn_pre, ffn_post, ffn_comb = self.hc_mixes(x, self.hc_ffn_fn, self.hc_ffn_scale, self.hc_ffn_base)
        x = self.hc_pre(x, attn_pre)
        x = self.ffn_norm(x)
        x = self.ffn(x, image_mask)
        x = self.hc_post(x, residual, ffn_post, ffn_comb)
        return x, ffn_pre

    def hc_mixes(self, x: torch.Tensor, hc_fn: torch.Tensor, hc_scale: torch.Tensor, hc_base: torch.Tensor):
        """x: [b,s,hc,d], hc_fn: [mix_hc, hc*d], hc_scale: [3], hc_base: [mix_hc]. Returns the
        pre / post / comb coefficients, split out of one projection of the flattened stream."""
        # normalized over the whole flattened hc*d stream, one statistic per token
        x = x.flatten(2).float()
        rsqrt = torch.rsqrt(x.square().mean(-1, keepdim=True) + self.norm_eps)
        mixes = F.linear(x, hc_fn) * rsqrt
        return hc_split_sinkhorn(mixes, hc_scale, hc_base, self.hc_mult, self.hc_sinkhorn_iters, self.hc_eps)

    def hc_pre(self, x: torch.Tensor, pre_mix: torch.Tensor):
        """Collapse the hc copies into one, weighted by pre_mix. [b,s,hc,d] x [b,s,hc] -> [b,s,d]"""
        y = torch.sum(pre_mix.unsqueeze(-1) * x.float(), dim=2)
        return y.to(x.dtype)

    def hc_post(self, x: torch.Tensor, residual: torch.Tensor, post: torch.Tensor, comb: torch.Tensor):
        """Expand the sublayer output back to hc copies and mix the residual in through `comb`.
        x: [b,s,d], residual: [b,s,hc,d], post: [b,s,hc], comb: [b,s,hc,hc] -> [b,s,hc,d]"""
        y = post.unsqueeze(-1) * x.unsqueeze(-2) + torch.sum(comb.unsqueeze(-1) * residual.unsqueeze(-2), dim=2)
        return y.type_as(x)

class Engram(nn.Module):
    """Writes an n-gram lookup into the residual stream, gated by how well it matches that stream.

    The hash ids fetch `n_hash_cols` rows; `wkv` turns them into one key per hc copy plus a shared
    value. The gate is a normalized dot product of stream against key.
    """

    def __init__(self, args: ModelArgs, layer_id: int, layout: EngramLayout):
        super().__init__()
        self.layer_id = layer_id
        self.layer_hash_index = layout.layer_ids.index(layer_id)
        self.dim = args.dim
        self.hc_mult = args.hc_mult
        self.clamp_value = 1e-6

        self.embed = ParallelEngramEmbedding(layout.num_embeddings[self.layer_hash_index], layout.head_dim)
        n_hash_cols = (layout.max_ngram_size - 1) * layout.n_heads
        self.wkv = Linear(n_hash_cols * layout.head_dim, args.dim * (args.hc_mult + 1))
        self.eps = args.norm_eps
        self.q_weight = nn.Parameter(torch.ones(args.hc_mult, args.dim))
        self.k_weight = nn.Parameter(torch.ones(args.hc_mult, args.dim))

    def forward(self, x: torch.Tensor, hash_ids: torch.Tensor, token_mask: torch.Tensor | None = None) -> torch.Tensor:
        """x: [B, L, hc_mult, dim]; hash_ids: [B, L, n_hash_cols]; token_mask: [B, L], False shuts
        the gate so those positions pass through untouched."""
        kv = self.wkv(self.embed(hash_ids).flatten(-2))
        key, value = kv.split([self.hc_mult * self.dim, self.dim], dim=-1)
        key = key.float().unflatten(-1, (self.hc_mult, self.dim))
        weight = self.q_weight.float() * self.k_weight.float()  # only ever used as a product
        h, eps = x.float(), self.eps
        # normalized per (token, hc copy) over `dim`, NOT jointly over the copies
        rstd = torch.rsqrt(h.square().mean(-1) + eps) * torch.rsqrt(key.square().mean(-1) + eps)
        dot = (h * weight * key).sum(-1) * rstd * self.dim**-0.5
        # signed sqrt before the sigmoid, matching the training kernel
        gate = torch.sigmoid(torch.copysign(dot.abs().clamp_min(self.clamp_value).sqrt(), dot))
        if token_mask is not None:
            gate = gate.masked_fill(~token_mask.unsqueeze(-1), 0)
        return (h + gate.unsqueeze(-1) * value.float().unsqueeze(-2)).to(x.dtype)