vllm.model_executor.models.minimax_text_01

class MiniMaxText01Attention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        head_dim: int,
        num_kv_heads: int,
        rotary_dim: int,
        max_position: int = 4096 * 32,
        rope_theta: float = 10000,
        sliding_window: Optional[int] = None,
        quant_config: Optional[QuantizationConfig] = None,
        layer_idx: int = None,
        cache_config: Optional[CacheConfig] = None,
        prefix: str = "mha",
    ) -> None:
        super().__init__()
        self.layer_idx = layer_idx

        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:
            assert self.total_num_kv_heads % tp_size == 0
        else:
            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

        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.sliding_window = sliding_window
        self.prefix = prefix

        self.qkv_proj = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=False,
            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.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",
        )
        self.rotary_emb = get_rope(
            head_size=self.head_dim,
            rotary_dim=rotary_dim,
            max_position=max_position,
            base=int(rope_theta),
            is_neox_style=True,
            dtype=torch.float32,
        )
        return

    def forward(self, hidden_states: torch.Tensor, output: torch.Tensor,
                positions: torch.Tensor, **kwargs) -> None:
        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output[:], _ = self.o_proj(attn_output)

class MiniMaxText01DecoderLayer(nn.Module):

    def __init__(
        self,
        config: MiniMaxConfig,
        model_config: Optional[ModelConfig] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        expert_num: int = 1,
        layer_id: int = None,
        linear_layer_id: Optional[int] = None,
        prefix: str = "decoder",
    ) -> None:
        self._ilayer = layer_id
        self._irank = get_tensor_model_parallel_rank()
        self.prefix = prefix
        super().__init__()

        self.hidden_size = config.hidden_size
        self.expert_num = expert_num

        rope_theta = getattr(config, "rope_theta", 10000)

        head_dim = getattr(config, "head_dim", None)
        if head_dim is None:
            head_dim = config.hidden_size // config.num_attention_heads
        if hasattr(config, "max_model_len") and isinstance(
                config.max_model_len, int):
            max_position_embeddings = min(config.max_position_embeddings,
                                          config.max_model_len)
        if config.attention_type == 0:
            use_headxdim = True
            hidden_inner = (head_dim * config.num_attention_heads
                            if use_headxdim else config.hidden_size)
            self.self_attn = MiniMaxText01LinearAttention(
                hidden_size=self.hidden_size,
                hidden_inner_size=hidden_inner,
                num_heads=config.num_attention_heads,
                head_dim=head_dim,
                max_position=max_position_embeddings,
                block_size=config.block if hasattr(config, "block") else 256,
                num_hidden_layer=config.num_hidden_layers,
                model_config=model_config,
                cache_config=cache_config,
                quant_config=quant_config,
                layer_idx=self._ilayer,
                linear_layer_idx=linear_layer_id,
                prefix=prefix)
        elif config.attention_type == 1:
            self.self_attn = MiniMaxText01Attention(
                hidden_size=self.hidden_size,
                num_heads=config.num_attention_heads,
                head_dim=head_dim,
                rotary_dim=config.rotary_dim
                if hasattr(config, "rotary_dim") else head_dim,
                num_kv_heads=config.num_key_value_heads,
                max_position=max_position_embeddings,
                rope_theta=rope_theta,
                sliding_window=config.sliding_window,
                quant_config=quant_config,
                layer_idx=self._ilayer,
                cache_config=cache_config,
                prefix=prefix)
        else:
            raise ValueError(
                f"Unsupported attention type: {self.config.attention_type}")

        if expert_num == 1:
            self.mlp = MiniMaxText01MLP(
                hidden_size=self.hidden_size,
                intermediate_size=config.intermediate_size,
                quant_config=quant_config,
                layer_idx=self._ilayer,
                prefix=prefix)
        else:
            self.block_sparse_moe = MiniMaxText01MoE(
                num_experts=expert_num,
                top_k=config.num_experts_per_tok,
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                layer_idx=self._ilayer,
                quant_config=quant_config,
                prefix=prefix)

        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)
        if config.attention_type == 0:
            self.layernorm_attention_alpha = getattr(
                config, 'layernorm_linear_attention_alpha',
                getattr(config, 'linear_attn_alpha_factor', 1))
            self.layernorm_attention_beta = getattr(
                config, 'layernorm_linear_attention_beta',
                getattr(config, 'linear_attn_beta_factor', 1))
        else:
            self.layernorm_attention_alpha = getattr(
                config, 'layernorm_full_attention_alpha',
                getattr(config, 'full_attn_alpha_factor', 1))
            self.layernorm_attention_beta = getattr(
                config, 'layernorm_full_attention_beta',
                getattr(config, 'full_attn_beta_factor', 1))
        self.layernorm_mlp_alpha = getattr(
            config, 'layernorm_mlp_alpha',
            getattr(config, 'mlp_alpha_factor', 1))
        self.layernorm_mlp_beta = getattr(
            config, 'layernorm_mlp_beta', getattr(config, 'mlp_beta_factor',
                                                  1))
        self.postnorm = getattr(config, 'postnorm', False)
        self.shared_moe = False

        shared_intermediate = getattr(config, 'shared_intermediate_size', 0)
        if isinstance(shared_intermediate, list):
            shared_intermediate = shared_intermediate[
                layer_id] if layer_id < len(shared_intermediate) else 0
        if shared_intermediate > 0:
            self.shared_moe = True
            self.shared_mlp = MiniMaxText01MLP(
                hidden_size=self.hidden_size,
                intermediate_size=shared_intermediate,
                quant_config=quant_config,
                layer_idx=self._ilayer,
                prefix=prefix)
            self.coefficient = ReplicatedLinear(
                self.hidden_size,
                1,
                bias=False,
                quant_config=quant_config,
                params_dtype=torch.float32,
            )
            self.coefficient.weight.weight_loader = (
                self.shared_moe_coefficient_loader)
            self.shared_moe_mode = getattr(config, 'shared_moe_mode',
                                           'softmax')
        return

    def forward(self,
                hidden_states: torch.Tensor,
                positions: torch.Tensor,
                kv_caches: Union[list[dict], Optional[torch.Tensor]],
                attn_metadata: AttentionMetadata,
                residual: Optional[torch.Tensor],
                is_warmup: bool = False,
                **kwargs) -> tuple[torch.Tensor, torch.Tensor]:

        layernorm_input = hidden_states
        layernorm_output = self.input_layernorm(layernorm_input)
        residual = layernorm_output if self.postnorm else layernorm_input
        self_attention_output = torch.empty_like(layernorm_output)
        self.self_attn(
            hidden_states=layernorm_output,
            output=self_attention_output,
            positions=positions,
            kv_caches=kv_caches,
        )

        residual = residual * self.layernorm_attention_alpha
        self_attention_output = (self_attention_output *
                                 self.layernorm_attention_beta)

        layernorm_input = residual + self_attention_output
        layernorm_output = self.post_attention_layernorm(layernorm_input)
        residual = layernorm_output if self.postnorm else layernorm_input

        if self.expert_num == 1:
            hidden_states = self.mlp(layernorm_output)
        else:
            moe_layernorm_output = layernorm_output.clone()
            moe_hidden_states = self.block_sparse_moe(moe_layernorm_output)
            if self.shared_moe:
                before_moe_dtype = layernorm_output.dtype
                moe_hidden_fp32 = moe_hidden_states.to(torch.float32)
                output_mlp = self.shared_mlp(layernorm_output).to(
                    torch.float32)

                coef, _ = self.coefficient(layernorm_output.to(torch.float32))

                if self.shared_moe_mode == 'softmax':
                    coef = torch.nn.functional.softmax(coef, dim=-1)
                    hidden_states = moe_hidden_fp32 * (
                        1 - coef) + output_mlp * coef
                elif self.shared_moe_mode == 'sigmoid':
                    coef = torch.nn.functional.sigmoid(coef)
                    hidden_states = moe_hidden_fp32 * (
                        1 - coef) + output_mlp * coef

                hidden_states = hidden_states.to(before_moe_dtype)
            else:
                hidden_states = moe_hidden_states

        residual = residual * self.layernorm_mlp_alpha
        hidden_states = hidden_states * self.layernorm_mlp_beta

        hidden_states = residual + hidden_states

        return hidden_states, None

    @staticmethod
    def shared_moe_coefficient_loader(param: torch.Tensor,
                                      loaded_weight: torch.Tensor) -> None:
        assert param.size() == loaded_weight.size()

        param.data.copy_(loaded_weight.to(torch.float32))
        return