DeepSeek V4.1

def make_identity_pre_mix(x: torch.Tensor, hc_mult: int) -> torch.Tensor:
    """initial one-hot mix"""
    pre_mix = x.new_zeros(x.size(0), x.size(1), hc_mult, dtype=torch.float32)
    pre_mix[:, :, 0] = 1.0
    return pre_mix

    @torch.inference_mode()
    def forward(
        self, input_ids: torch.Tensor, start_pos: int = 0, images=None, token_types: torch.Tensor | None = None
    ):
        """input_ids: [b, s], every entry a real token id -- generate.py only ever passes positions it
        has already filled, so the padding it uses internally never reaches here. token_types /
        images carry the VL inputs built by image_processor.prepare_vl_inputs; image spans must lie
        inside the first (start_pos 0) chunk."""
        image_mask = None if token_types is None else token_types >= 0  # TEXT is -1
        # image tokens take no part in an n-gram and get no engram contribution; text-only needs no mask
        engram_mask = None if image_mask is None else ~image_mask
        engram_hashes = self.engram_hash(input_ids, start_pos, engram_mask) if self.engram_hash is not None else None
        h = self.embed(input_ids)
        if images is not None:
            assert start_pos == 0, "image spans must be prefilled in a single chunk"
            self.merge_image_embeddings(images, h)
        # Expand to hc_mult copies for Hyper-Connections
        h = h.unsqueeze(2).repeat(1, 1, self.hc_mult, 1)
        main_hiddens = []
        pre_mix = make_identity_pre_mix(h, self.hc_mult)
        for i, layer in enumerate(self.layers):
            if layer.engram is not None:
                h = layer.engram(h, engram_hashes[:, :, layer.engram.layer_hash_index, :], engram_mask)
            # the MTP head reads the attention input of its target layers, not their output
            if i in self.target_layer_ids:
                main_hiddens.append(h.mean(dim=2))
            h, pre_mix = layer(h, start_pos, pre_mix, image_mask)
        h = layer.hc_pre(h, pre_mix)
        logits = self.head(self.norm(h))
        output_ids = sample(logits, self.temperature)
        main_hidden = torch.cat(main_hiddens, dim=-1) if main_hiddens else None
        return output_ids, logits, main_hidden

    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)

    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