vllm.model_executor.models.step3_text

class FusedMoEBlock(nn.Module):

    def __init__(self,
                 config: ModelConfig,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()

        if self.tp_size > config.moe_num_experts:
            raise ValueError(
                f"Tensor parallel size {self.tp_size} is greater than "
                f"the number of experts {config.moe_num_experts}.")

        self.experts = FusedMoE(num_experts=config.moe_num_experts,
                                top_k=config.moe_top_k,
                                hidden_size=config.hidden_size,
                                intermediate_size=config.moe_intermediate_size,
                                reduce_results=False,
                                renormalize=config.norm_expert_weight,
                                quant_config=quant_config,
                                prefix=f"{prefix}.experts")
        self.gate = ReplicatedLinear(config.hidden_size,
                                     config.moe_num_experts,
                                     bias=False,
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")

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

        router_logits, _ = self.gate(hidden_states)

        final_hidden_states = self.experts(hidden_states=hidden_states,
                                           router_logits=router_logits)
        if self.tp_size > 1:
            final_hidden_states = tensor_model_parallel_all_reduce(
                final_hidden_states)

        return final_hidden_states.view(orig_shape)

class Step3TextAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        num_kv_heads: int,
        norm_eps: float,
        rope_theta: int,
        share_q_dim: Optional[int] = None,
        rope_scaling: Optional[dict[str, Any]] = None,
        max_position_embedding: int = 8192,
        head_dim: int = 256,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        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

        if num_kv_heads != 1:
            raise ValueError(f"Step3TextAttention num_kv_heads must be 1, "
                             f"but got {num_kv_heads}.")
        self.num_kv_heads = num_kv_heads

        self.head_dim = head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.q_size = share_q_dim if share_q_dim else self.head_dim

        self.qkv_proj = ReplicatedLinear(
            hidden_size,
            self.q_size + self.kv_size * 2,
            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.inter_norm = RMSNorm(self.q_size, eps=norm_eps)
        self.wq = ColumnParallelLinear(
            self.q_size,
            self.head_dim * self.total_num_heads,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.wq",
        )
        self.rotary_emb = get_rope(self.head_dim,
                                   rotary_dim=self.head_dim,
                                   max_position=max_position_embedding,
                                   base=rope_theta,
                                   rope_scaling=rope_scaling)
        scaling = self.head_dim**-0.5
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              scaling,
                              self.num_kv_heads,
                              cache_config=cache_config,
                              prefix=f"{prefix}.attn")

    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)
        q = self.inter_norm(q)
        q = self.wq(q)[0]
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        residual, _ = self.o_proj(attn_output)
        return residual

class Step3TextDecoderLayer(nn.Module):

    def __init__(self,
                 config: ModelConfig,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = "") -> None:
        super().__init__()
        config = config.hf_config
        self.hidden_size = config.hidden_size
        rope_scaling = getattr(config, "rope_scaling", None)

        self.self_attn = Step3TextAttention(
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            num_kv_heads=1,
            cache_config=cache_config,
            quant_config=quant_config,
            norm_eps=config.rms_norm_eps,
            max_position_embedding=config.max_position_embedding,
            head_dim=config.head_dim,
            share_q_dim=config.share_q_dim,
            rope_theta=config.rope_theta,
            rope_scaling=rope_scaling,
            prefix=f"{prefix}.self_attn")

        layer_idx = int(prefix.split("layers.")[1].split(".")[0])
        moe_layers_enum = getattr(config, "moe_layers_enum", None)
        if moe_layers_enum is not None:
            moe_layers_idx = [
                int(i) for i in moe_layers_enum.strip().split(',')
            ]
        else:
            # Default to 1dense.
            moe_layers_idx = [i for i in range(1, config.num_hidden_layers)]

        if layer_idx in moe_layers_idx:
            self.moe = FusedMoEBlock(config=config,
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.moe")
            self.share_expert = Step3TextMLP(
                hidden_size=self.hidden_size,
                intermediate_size=config.share_expert_dim,
                hidden_act="silu",
                quant_config=quant_config,
                prefix=f"{prefix}.share_expert")
            self.use_moe = True
        else:
            self.mlp = Step3TextMLP(hidden_size=config.hidden_size,
                                    intermediate_size=config.intermediate_size,
                                    hidden_act="silu",
                                    quant_config=quant_config,
                                    prefix=f"{prefix}.mlp")
            self.use_moe = False
        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]:
        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)

        if self.use_moe:
            share_output = self.share_expert(hidden_states)
            moe_output = self.moe(hidden_states)
            hidden_states = share_output + moe_output
        else:
            hidden_states = self.mlp(hidden_states)

        return hidden_states, residual

class Step3TextForCausalLM(nn.Module, SupportsPP):

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
    ):
        super().__init__()
        config = vllm_config.model_config.hf_config
        lora_config = vllm_config.lora_config
        self.config = config
        self.vllm_config = vllm_config

        self.model = Step3TextModel(vllm_config=vllm_config, prefix=prefix)

        if get_pp_group().is_last_rank:
            self.unpadded_vocab_size = config.vocab_size
            if lora_config:
                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
            self.lm_head = ParallelLMHead(
                self.unpadded_vocab_size,
                config.hidden_size,
                org_num_embeddings=config.vocab_size,
                padding_size=DEFAULT_VOCAB_PADDING_SIZE
                if not lora_config else lora_config.lora_vocab_padding_size,
            )
            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
                                                    config.vocab_size)
            self.sampler = get_sampler()
        else:
            self.lm_head = PPMissingLayer()

        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)

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

    def sample(
        self,
        logits: Optional[torch.Tensor],
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
        next_tokens = self.sampler(logits, sampling_metadata)
        return next_tokens

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        qkv_params_mapping = [
            # (param_name, shard_name, relative_start_idx, relative_end_idx)
            (".qkv_proj", ".q_proj", 0, self.config.share_q_dim /
             (self.config.share_q_dim + self.config.head_dim * 2)),
            (".qkv_proj", ".k_proj", self.config.share_q_dim /
             (self.config.share_q_dim + self.config.head_dim * 2),
             (self.config.share_q_dim + self.config.head_dim) /
             (self.config.share_q_dim + self.config.head_dim * 2)),
            (".qkv_proj", ".v_proj",
             (self.config.share_q_dim + self.config.head_dim) /
             (self.config.share_q_dim + self.config.head_dim * 2),
             (self.config.share_q_dim + self.config.head_dim * 2) /
             (self.config.share_q_dim + self.config.head_dim * 2)),
        ]
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()

        expert_params_mapping = [
            (".moe.experts.w13_weight", ".moe.gate_proj.weight", "w1"),
            (".moe.experts.w13_weight", ".moe.up_proj.weight", "w3"),
            (".moe.experts.w2_weight", ".moe.down_proj.weight", "w2")
        ]

        disable_moe_stacked_params = [
            data[1] for data in expert_params_mapping
        ]

        for name, loaded_weight in weights:
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
                if any(disable_moe_stacked_param in name
                       for disable_moe_stacked_param in
                       disable_moe_stacked_params):
                    continue
                name = name.replace(weight_name, param_name)
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                loaded_params.add(name)
                break
            else:
                for mapping in expert_params_mapping:
                    param_name, weight_name, shard_id = mapping
                    if weight_name not in name:
                        continue
                    name = name.replace(weight_name, param_name)
                    # Skip layers on other devices.
                    if is_pp_missing_parameter(name, self):
                        continue
                    # Skip loading extra bias for GPTQ models.
                    if ((name.endswith(".bias") or name.endswith("_bias"))
                            and name not in params_dict):
                        continue
                    param = params_dict[name]
                    weight_loader = param.weight_loader
                    for expert_id in range(loaded_weight.shape[0]):
                        loaded_weight_expert = loaded_weight[expert_id]
                        weight_loader(param,
                                      loaded_weight_expert,
                                      name,
                                      shard_id=shard_id,
                                      expert_id=expert_id)
                    loaded_params.add(name)
                    break
                else:
                    for (param_name, weight_name, start_idx,
                         end_idx) in qkv_params_mapping:
                        if weight_name not in name:
                            continue
                        name = name.replace(weight_name, param_name)
                        if is_pp_missing_parameter(name, self):
                            continue
                        param = params_dict[name]
                        dim = param.shape[param.output_dim]
                        begin_idx = int(start_idx * dim)
                        end_idx = int(end_idx * dim)
                        param_slice = param.narrow(param.output_dim, begin_idx,
                                                   end_idx - begin_idx)
                        param_slice.copy_(loaded_weight)
                        loaded_params.add(name)
                        break
                    else:
                        if is_pp_missing_parameter(name, self):
                            continue
                        param = params_dict[name]
                        weight_loader = getattr(param, "weight_loader",
                                                default_weight_loader)
                        weight_loader(param, loaded_weight)
                        loaded_params.add(name)
        return loaded_params