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)

    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.78.*"]

… lines 649–695 omitted …

        if inputs_embeds is None:
            inputs_embeds: torch.Tensor = self.embed_tokens(input_ids)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        if position_ids is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
            position_ids = position_ids.unsqueeze(0)

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
                "allow_is_causal_skip": False,  # Always force creation to account for causality in the indexer
            }
            causal_mask_mapping = {"indexed_attention": create_causal_mask(**mask_kwargs)}

        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)

        topk_indices = None  # MAIN DIFF with DSV3.2
        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
            hidden_states, topk_indices = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping["indexed_attention"],
                position_embeddings=position_embeddings,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                prev_topk_indices=topk_indices,  # MAIN DIFF with DSV3.2
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )

… lines 740–748 omitted …

    def __init__(self, config):
        super().__init__(config)
        self.model = GlmMoeDsaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

… lines 754–798 omitted …

        hidden_states = outputs.last_hidden_state
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

class GlmMoeDsaDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: GlmMoeDsaConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = GlmMoeDsaAttention(config, layer_idx)

        self.mlp = GlmMoeDsaMoE(config) if config.mlp_layer_types[layer_idx] == "sparse" else GlmMoeDsaMLP(config)

        self.input_layernorm = GlmMoeDsaRMSNorm(config.hidden_size, config.rms_norm_eps)
        self.post_attention_layernorm = GlmMoeDsaRMSNorm(config.hidden_size, config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        prev_topk_indices: torch.Tensor | None = None,  # MAIN DIFF with DSV3.2
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Self Attention
        hidden_states, _, topk_indices = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            position_embeddings=position_embeddings,
            prev_topk_indices=prev_topk_indices,  # MAIN DIFF with DSV3.2
            **kwargs,
        )
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states, topk_indices