GLM-5

class GlmMoeDsaRotaryEmbedding(nn.Module):
    @deprecate_kwarg("device", version="5.18")
    def __init__(self, config: GlmMoeDsaConfig, device=None):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config

        self.rope_type = self.config.rope_parameters["rope_type"]
        rope_init_fn: Callable = self.compute_default_rope_parameters
        if self.rope_type != "default":
            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)

        self.inv_freq = nn.Buffer(inv_freq, persistent=False)
        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)

    @staticmethod
    @deprecate_kwarg("device", version="5.18")
    def compute_default_rope_parameters(config: GlmMoeDsaConfig, device=None, **kwargs) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

        attention_factor = 1.0  # Unused in this type of RoPE
        # Compute the inverse frequencies
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        return inv_freq.to(device), attention_factor

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = (
            self.inv_freq[None, :, None].expand(position_ids.shape[0], -1, 1).to(dtype=torch.float, device=x.device)
        )
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) else "cpu"
        # Disable any outside autocast context if any, to really force fp32
        with maybe_autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)

    def __post_init__(self, **kwargs):
        # Per-layer indexer mode: a pattern (e.g. `"FSSF..."`) overrides the freq/offset schedule.
        if self.indexer_types is None:
            pattern = kwargs.get("index_topk_pattern")
            if pattern is not None:
                self.indexer_types = (
                    [{"F": "full", "S": "shared"}[c] for c in pattern] if isinstance(pattern, str) else list(pattern)
                )
            else:
                freq = max(kwargs.get("index_topk_freq", 1), 1)
                offset = kwargs.get("index_skip_topk_offset", 2)
                self.indexer_types = [
                    "full" if (max(i - offset + 1, 0) % freq) == 0 else "shared" for i in range(self.num_hidden_layers)
                ]
        self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
        # RoPE applies only to the rope slice, so point `head_dim` at it: the inherited (Llama) rotary
        # embedding reads `config.head_dim` and then computes the right frequencies with no override needed.
        self.head_dim = self.qk_rope_head_dim
        # MLP layer types: the first `first_k_dense_replace` layers are dense, the rest are MoE.
        if self.mlp_layer_types is None:
            n_dense = min(self.first_k_dense_replace, self.num_hidden_layers)
            self.mlp_layer_types = ["dense"] * n_dense + ["sparse"] * (self.num_hidden_layers - n_dense)

def yarn_get_mscale(scale=1, mscale=1):
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


def yarn_apply_mscale(rope_parameters, scaling):
    if rope_parameters.get("rope_type", "default") != "default":
        mscale_all_dim = rope_parameters.get("mscale_all_dim", 0)
        scaling_factor = rope_parameters["factor"]
        if mscale_all_dim:
            mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
            scaling = scaling * mscale * mscale
    return scaling

… lines 307–325 omitted …

        self.q_lora_rank = config.q_lora_rank
        self.qk_rope_head_dim = config.qk_rope_head_dim
        self.kv_lora_rank = config.kv_lora_rank
        self.v_head_dim = config.v_head_dim
        self.qk_nope_head_dim = config.qk_nope_head_dim
        self.qk_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads

        self.is_causal = True
        self.q_a_proj = nn.Linear(self.hidden_size, self.q_lora_rank, bias=config.attention_bias)
        self.q_a_layernorm = GlmMoeDsaRMSNorm(self.q_lora_rank)
        self.q_b_proj = nn.Linear(self.q_lora_rank, self.num_heads * self.qk_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 = GlmMoeDsaRMSNorm(config.kv_lora_rank)
        self.kv_b_proj = nn.Linear(
            config.kv_lora_rank,
            self.num_heads * (self.qk_nope_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.scaling = yarn_apply_mscale(config.rope_parameters, self.qk_head_dim ** (-0.5))
        # Refer: https://arxiv.org/abs/2603.12201 for more details.
        self.skip_topk = config.indexer_types[layer_idx] == "shared"
        self.indexer = None if self.skip_topk else GlmMoeDsaIndexer(config, layer_idx)

def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights

… lines 291–403 omitted …

        cos, sin = position_embeddings
        q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)

        # Cache read / write is performed while latent KV is still compressed
        if past_key_values is not None:
            k_pass, k_rot = past_key_values.update(k_pass, k_rot, self.layer_idx)

        query_states = torch.cat((q_pass, q_rot), dim=-1)

        key_states, value_states = self.expand_kv(k_pass, k_rot)

        # DSA: select this layer's top-k tokens, or reuse the previous full layer's on `"shared"` layers.
        if self.indexer is not None:
            topk_indices = self.indexer(
                hidden_states,
                q_resid,
                position_embeddings,
                attention_mask[:, 0, :, :],
                position_ids,  # Kept for BC
                past_key_values=past_key_values,
            )  # [B, S, topk]
        else:
            if prev_topk_indices is None:
                raise ValueError("Shared DSA layers require top-k indices from a previous full indexer layer.")
            topk_indices = prev_topk_indices

        sparse_indices = None
        if self.config._attn_implementation in ("eager", "sdpa"):
            index_mask = (
                topk_indices.new_ones((batch_size, seq_length, key_states.shape[2]), dtype=torch.bool)
                .scatter(-1, topk_indices.long(), False)
                .unsqueeze(1)
            )

            if attention_mask.dtype == torch.bool:
                attention_mask = attention_mask & ~index_mask
            else:
                attention_mask = attention_mask.masked_fill(index_mask, torch.finfo(hidden_states.dtype).min)
        else:
            sparse_indices = topk_indices

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            indices=sparse_indices,  # consumed by flash_mla_with_kvcache; ignored by eager / SDPA
            **kwargs,
        )

        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights, topk_indices