vllm.model_executor.models.bert_with_rope

@support_torch_compile
@default_pooling_type("CLS")
class BertWithRope(nn.Module, SupportsQuant):
    hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})

    def __init__(self,
                 *,
                 vllm_config: VllmConfig,
                 prefix: str = "",
                 add_pooling_layer: bool = False):
        super().__init__()
        self.vllm_config = vllm_config
        self.add_pooling_layer = add_pooling_layer
        self.config = vllm_config.model_config.hf_config
        self.embeddings = BertWithRopeEmbedding(self.config)
        self.encoder = BertWithRopeEncoder(
            vllm_config=vllm_config,
            bias=getattr(self.config, "bias", True),
            rotary_kwargs=self.config.rotary_kwargs,
            prefix=f"{prefix}.encoder")
        self.pooler = BertPooler(self.config) if add_pooling_layer else None

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.embeddings(input_ids=input_ids,
                                            token_type_ids=token_type_ids)
        return self.encoder(positions, hidden_states)

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

        if self.config.hidden_act in ["silu", "geglu"]:
            stacked_params_mapping = [
                # (param_name, shard_name, shard_id)
                ("gate_up_proj", "gate_proj", 0),
                ("gate_up_proj", "up_proj", 1),
            ]
        else:
            stacked_params_mapping = []

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if not self.add_pooling_layer and "pooler" in name:
                continue
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                if name.endswith((".w1", ".w2")):
                    # Nomic-MoE has fused experts weights
                    weight_loader(param, loaded_weight, name)
                else:
                    weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

class BertWithRopeAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_attention_heads: int,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        bias: bool = True,
        rotary_kwargs: Optional[dict] = None,
        prefix: str = "",
    ):
        super().__init__()

        self.hidden_size = hidden_size
        tp_size = get_tensor_model_parallel_world_size()

        self.total_num_heads = num_attention_heads
        assert self.total_num_heads % tp_size == 0

        self.num_heads = self.total_num_heads // tp_size
        self.total_num_kv_heads = self.total_num_heads
        self.head_dim = self.hidden_size // self.total_num_heads
        assert self.head_dim * self.total_num_heads == self.hidden_size

        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)

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

        self.rotary_emb = get_rope(**rotary_kwargs)

        self.attn = EncoderOnlyAttention(num_heads=self.num_heads,
                                         head_size=self.head_dim,
                                         scale=self.scaling,
                                         num_kv_heads=self.num_kv_heads,
                                         cache_config=cache_config,
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.attn")

        self.out_proj = RowParallelLinear(input_size=hidden_size,
                                          output_size=hidden_size,
                                          bias=bias,
                                          quant_config=quant_config,
                                          prefix=f"{prefix}.dense")

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

class BertWithRopeBlock(nn.Module):

    def __init__(self,
                 config: PretrainedConfig,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 moe: bool = False,
                 bias: bool = True,
                 rotary_kwargs: Optional[dict] = None,
                 prefix: str = ""):
        super().__init__()
        self.attn = BertWithRopeAttention(
            hidden_size=config.hidden_size,
            num_attention_heads=config.num_attention_heads,
            cache_config=cache_config,
            quant_config=quant_config,
            bias=bias,
            rotary_kwargs=rotary_kwargs,
            prefix=f"{prefix}.attention")

        if moe:
            self.mlp = NomicMoE(num_experts=config.num_experts,
                                top_k=config.moe_top_k,
                                hidden_size=config.hidden_size,
                                intermediate_size=config.intermediate_size,
                                hidden_act=config.hidden_act)
        else:
            if config.hidden_act in ["silu", "geglu"]:
                self.mlp = BertWithRopeGatedMLP(
                    hidden_size=config.hidden_size,
                    intermediate_size=config.intermediate_size,
                    hidden_act=config.hidden_act,
                    bias=bias,
                    quant_config=quant_config,
                    prefix=f"{prefix}.mlp")
            else:
                self.mlp = BertWithRopeMLP(
                    hidden_size=config.hidden_size,
                    intermediate_size=config.intermediate_size,
                    hidden_act=config.hidden_act,
                    bias=bias,
                    quant_config=quant_config,
                    prefix=f"{prefix}.mlp")

        self.attn_ln = nn.LayerNorm(config.hidden_size,
                                    eps=config.layer_norm_eps)
        self.mlp_ln = nn.LayerNorm(config.hidden_size,
                                   eps=config.layer_norm_eps)

    def forward(self, positions: torch.Tensor, hidden_states: torch.Tensor):
        attn_output = self.attn(positions, hidden_states)
        hidden_states = self.attn_ln(hidden_states + attn_output)
        mlp_out = self.mlp(hidden_states)
        hidden_states = self.mlp_ln(hidden_states + mlp_out)
        return hidden_states

class BertWithRopeEmbedding(nn.Module):

    def __init__(self, config: PretrainedConfig):

        super().__init__()
        if config.position_embedding_type not in ["rope", "rotary"]:
            raise ValueError("Only 'rotary'('rope') position_embedding_type" +
                             " is supported")

        self.word_embeddings = VocabParallelEmbedding(config.vocab_size,
                                                      config.hidden_size)
        if config.type_vocab_size > 0:
            self.token_type_embeddings = VocabParallelEmbedding(
                config.type_vocab_size, config.hidden_size)
        else:
            self.token_type_embeddings = None

        self.LayerNorm = nn.LayerNorm(config.hidden_size,
                                      eps=config.layer_norm_eps)

    def forward(
        self,
        input_ids: torch.Tensor,
        token_type_ids: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        input_shape = input_ids.size()
        inputs_embeds = self.word_embeddings(input_ids)

        embeddings = inputs_embeds
        if self.token_type_embeddings is not None:
            if token_type_ids is None:
                token_type_ids = torch.zeros(input_shape,
                                             dtype=torch.long,
                                             device=inputs_embeds.device)

            token_type_embeddings = self.token_type_embeddings(token_type_ids)
            embeddings += token_type_embeddings

        embeddings = self.LayerNorm(embeddings)
        return embeddings

class BertWithRopeEncoder(nn.Module):

    def __init__(self,
                 vllm_config: VllmConfig,
                 bias: bool = True,
                 rotary_kwargs: Optional[dict] = None,
                 prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
        every_n = getattr(config, "moe_every_n_layers", 0)
        self.layers = nn.ModuleList([
            BertWithRopeBlock(config=config,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              bias=bias,
                              moe=every_n > 0 and (layer_idx % every_n == 1),
                              rotary_kwargs=rotary_kwargs,
                              prefix=f"{prefix}.layer.{layer_idx}")
            for layer_idx in range(config.num_hidden_layers)
        ])

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        for layer in self.layers:
            hidden_states = layer(positions, hidden_states)
        return hidden_states

class BertWithRopeGatedMLP(nn.Module):

    def __init__(self,
                 hidden_size: int,
                 intermediate_size: int,
                 hidden_act: str,
                 bias: bool = True,
                 quant_config: Optional[QuantizationConfig] = None,
                 prefix: str = ""):
        super().__init__()
        self.act_fn = get_act_and_mul_fn(hidden_act)
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size,
            [intermediate_size] * 2,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.down_proj = RowParallelLinear(input_size=intermediate_size,
                                           output_size=hidden_size,
                                           bias=bias,
                                           quant_config=quant_config,
                                           prefix=f"{prefix}.down_proj")

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        gate_up, _ = self.gate_up_proj(hidden_states)
        hidden_states = self.act_fn(gate_up)
        hidden_states, _ = self.down_proj(hidden_states)
        return hidden_states