vllm.model_executor.models.glm4_moe

class Glm4MoE(nn.Module):

    def __init__(
        self,
        config: Glm4MoeConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        enable_eplb: bool = False,
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor

        self.ep_group = get_ep_group().device_group
        self.ep_rank = self.ep_group.rank()
        self.ep_size = self.ep_group.size()
        self.n_routed_experts: int = config.n_routed_experts
        self.n_shared_experts: int = config.n_shared_experts

        if config.hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {config.hidden_act}. "
                             "Only silu is supported for now.")
        # NOTE In the transformers implementation, the gate isn't an nn.Linear,
        # so we cannot use ReplicatedLinear here.
        # See: https://github.com/huggingface/transformers/blob/v4.55.1/src/transformers/models/glm4_moe/modeling_glm4_moe.py#L260
        self.gate = nn.Linear(
            config.hidden_size,
            config.n_routed_experts,
            bias=False,
            dtype=torch.float32,
        )
        self.gate.e_score_correction_bias = nn.Parameter(
            torch.empty(config.n_routed_experts, dtype=torch.float32))

        # Load balancing settings.
        vllm_config = get_current_vllm_config()
        eplb_config = vllm_config.parallel_config.eplb_config
        self.enable_eplb = enable_eplb

        self.n_redundant_experts = eplb_config.num_redundant_experts
        self.n_logical_experts = self.n_routed_experts
        self.n_physical_experts = (self.n_logical_experts +
                                   self.n_redundant_experts)
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

        self.physical_expert_start = (self.ep_rank *
                                      self.n_local_physical_experts)
        self.physical_expert_end = (self.physical_expert_start +
                                    self.n_local_physical_experts)

        self.experts = FusedMoE(
            num_experts=config.n_routed_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            reduce_results=False,
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            use_grouped_topk=True,
            num_expert_group=config.n_group,
            topk_group=config.topk_group,
            prefix=f"{prefix}.experts",
            scoring_func="sigmoid",
            e_score_correction_bias=self.gate.e_score_correction_bias,
            enable_eplb=self.enable_eplb,
            num_redundant_experts=self.n_redundant_experts)

        if config.n_shared_experts is not None:
            intermediate_size = (config.moe_intermediate_size *
                                 config.n_shared_experts)
            self.shared_experts = Glm4MoeMLP(
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
                reduce_results=self.experts.must_reduce_shared_expert_outputs(
                ),
                prefix=f"{prefix}.shared_experts",
            )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)

        if self.n_shared_experts is not None:
            shared_output = self.shared_experts(hidden_states)
        router_logits = self.gate(hidden_states.to(dtype=torch.float32))
        final_hidden_states = self.experts(
            hidden_states=hidden_states,
            router_logits=router_logits) * self.routed_scaling_factor
        if shared_output is not None:
            final_hidden_states = final_hidden_states + shared_output
        if self.tp_size > 1:
            final_hidden_states = (
                self.experts.maybe_all_reduce_tensor_model_parallel(
                    final_hidden_states))
        return final_hidden_states.view(num_tokens, hidden_dim)

class Glm4MoeAttention(nn.Module):

    def __init__(
        self,
        config: Glm4MoeConfig,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embeddings: int = 131072,
        head_dim: Optional[int] = None,
        rms_norm_eps: float = 1e-05,
        qkv_bias: bool = False,
        use_qk_norm: bool = False,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = num_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = num_kv_heads
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = head_dim or (hidden_size // self.total_num_heads)
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings
        self.use_qk_norm = use_qk_norm

        self.qkv_proj = QKVParallelLinear(hidden_size,
                                          self.head_dim,
                                          self.total_num_heads,
                                          self.total_num_kv_heads,
                                          bias=qkv_bias,
                                          quant_config=quant_config,
                                          prefix=f"{prefix}.qkv_proj")

        self.o_proj = RowParallelLinear(self.total_num_heads * self.head_dim,
                                        hidden_size,
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")

        partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
            partial_rotary_factor=partial_rotary_factor,
        )
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            self.scaling,
            num_kv_heads=self.num_kv_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

        if self.use_qk_norm:
            self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
            self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        if self.use_qk_norm:
            q = self.q_norm(q.reshape(-1, self.num_heads,
                                      self.head_dim)).reshape(q.shape)
            k = self.k_norm(k.reshape(-1, self.num_kv_heads,
                                      self.head_dim)).reshape(k.shape)

        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

class Glm4MoeDecoderLayer(nn.Module):

    def __init__(
        self,
        config: Glm4MoeConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        enable_eplb: bool = False,
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          131072)
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
        layer_idx = int(prefix.split(sep='.')[-1])
        self.layer_idx = layer_idx

        self.self_attn = Glm4MoeAttention(
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=config.num_key_value_heads,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            head_dim=config.head_dim,
            rms_norm_eps=config.rms_norm_eps,
            qkv_bias=config.attention_bias,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            use_qk_norm=config.use_qk_norm,
        )

        if (config.n_routed_experts is not None
                and layer_idx >= config.first_k_dense_replace):
            self.mlp = Glm4MoE(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
                enable_eplb=enable_eplb,
            )
        else:
            self.mlp = Glm4MoeMLP(hidden_size=config.hidden_size,
                                  intermediate_size=config.intermediate_size,
                                  hidden_act=config.hidden_act,
                                  quant_config=quant_config,
                                  prefix=f"{prefix}.mlp")

        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)
        self.routed_scaling_factor = config.routed_scaling_factor

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if residual is None:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
        else:
            hidden_states, residual = self.input_layernorm(
                hidden_states, residual)
        hidden_states = self.self_attn(positions=positions,
                                       hidden_states=hidden_states)
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual

class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    fall_back_to_pt_during_load = False

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        self.config = config
        self.quant_config = quant_config
        self.model = Glm4MoeModel(vllm_config=vllm_config,
                                  prefix=maybe_prefix(prefix, "model"))
        if get_pp_group().is_last_rank:
            self.lm_head = ParallelLMHead(config.vocab_size,
                                          config.hidden_size,
                                          quant_config=quant_config)
        else:
            self.lm_head = PPMissingLayer()
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
        self.expert_weights = []

        # Set MoE hyperparameters
        self.num_moe_layers = (config.num_hidden_layers -
                               config.first_k_dense_replace)
        self.num_expert_groups = config.n_group

        self.moe_layers: list[FusedMoE] = []
        example_moe = None
        for layer in self.model.layers:
            if isinstance(layer, PPMissingLayer):
                continue

            assert isinstance(layer, Glm4MoeDecoderLayer)
            if isinstance(layer.mlp, Glm4MoE):
                # Pick last one layer since the first ones may be dense layers.
                example_moe = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

        if example_moe is None:
            raise RuntimeError("No Glm4MoE layer found in model.layers.")

        self.num_logical_experts = example_moe.n_logical_experts
        self.num_physical_experts = example_moe.n_physical_experts
        self.num_local_physical_experts = example_moe.n_local_physical_experts
        self.num_routed_experts = example_moe.n_routed_experts
        self.num_shared_experts = example_moe.n_shared_experts
        self.num_redundant_experts = example_moe.n_redundant_experts

    def set_eplb_state(
        self,
        expert_load_view: torch.Tensor,
        logical_to_physical_map: torch.Tensor,
        logical_replica_count: torch.Tensor,
    ) -> None:
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
                                   inputs_embeds)
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)

    def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
        return self.model.get_expert_mapping()