vllm.model_executor.models.qwen3_moe

class Qwen3MoeAttention(nn.Module):

    def __init__(
        self,
        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 = 8192,
        head_dim: Optional[int] = None,
        rms_norm_eps: float = 1e-06,
        qkv_bias: bool = False,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        dual_chunk_attention_config: Optional[dict[str, Any]] = None,
    ) -> 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.dual_chunk_attention_config = dual_chunk_attention_config

        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")

        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,
            dual_chunk_attention_config=dual_chunk_attention_config,
        )
        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",
            **{
                "layer_idx": extract_layer_index(prefix),
                "dual_chunk_attention_config": dual_chunk_attention_config,
            } if dual_chunk_attention_config else {},
        )

        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)
        # Add qk-norm
        q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim,
                           self.head_dim)
        q_by_head = self.q_norm(q_by_head)
        q = q_by_head.view(q.shape)

        k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim,
                           self.head_dim)
        k_by_head = self.k_norm(k_by_head)
        k = k_by_head.view(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 Qwen3MoeDecoderLayer(nn.Module):

    def __init__(
        self,
        config: Qwen3MoeConfig,
        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",
                                          8192)
        dual_chunk_attention_config = getattr(config,
                                              "dual_chunk_attention_config",
                                              None)
        self.self_attn = Qwen3MoeAttention(
            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,
            rms_norm_eps=config.rms_norm_eps,
            qkv_bias=getattr(config, 'attention_bias', False),
            head_dim=getattr(config, 'head_dim', None),
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            dual_chunk_attention_config=dual_chunk_attention_config,
        )

        # `mlp_only_layers` in the config.
        layer_idx = extract_layer_index(prefix)
        mlp_only_layers = ([] if not hasattr(config, "mlp_only_layers") else
                           config.mlp_only_layers)
        if (layer_idx not in mlp_only_layers) and (
                config.num_experts > 0 and
            (layer_idx + 1) % config.decoder_sparse_step == 0):
            self.mlp = Qwen3MoeSparseMoeBlock(config=config,
                                              quant_config=quant_config,
                                              prefix=f"{prefix}.mlp",
                                              enable_eplb=enable_eplb)
        else:
            self.mlp = Qwen3MoeMLP(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)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        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,
        )

        # Fully Connected
        hidden_states, residual = self.post_attention_layernorm(
            hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual

class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA,
                          MixtureOfExperts):
    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 = Qwen3MoeModel(vllm_config=vllm_config,
                                   prefix=maybe_prefix(prefix, "model"))
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      quant_config=quant_config)
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.model.embed_tokens.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

        # Set MoE hyperparameters
        self.expert_weights = []

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

            assert isinstance(layer, Qwen3MoeDecoderLayer)
            if isinstance(layer.mlp, Qwen3MoeSparseMoeBlock):
                example_layer = layer.mlp
                self.moe_layers.append(layer.mlp.experts)

        if example_layer is None:
            raise RuntimeError("No Qwen3MoE layer found in the model.layers.")

        self.num_moe_layers = len(self.moe_layers)
        self.num_expert_groups = 1
        self.num_shared_experts = 0
        self.num_logical_experts = example_layer.n_logical_experts
        self.num_physical_experts = example_layer.n_physical_experts
        self.num_local_physical_experts = example_layer.n_local_physical_experts
        self.num_routed_experts = example_layer.n_routed_experts
        self.num_redundant_experts = example_layer.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 update_physical_experts_metadata(
        self,
        num_physical_experts: int,
        num_local_physical_experts: int,
    ) -> None:
        assert self.num_local_physical_experts == num_local_physical_experts
        self.num_physical_experts = num_physical_experts
        self.num_local_physical_experts = num_local_physical_experts
        self.num_redundant_experts = (num_physical_experts -
                                      self.num_logical_experts)
        for layer in self.model.layers:
            if isinstance(layer.mlp, Qwen3MoeSparseMoeBlock):
                moe = layer.mlp
                moe.n_local_physical_experts = num_local_physical_experts
                moe.n_physical_experts = num_physical_experts
                moe.n_redundant_experts = self.num_redundant_experts
                moe.experts.update_expert_map()

    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()