vllm.model_executor.models.bart

class BartCrossAttention(nn.Module):

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        bias: bool = True,
        config: Optional[BartConfig] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.d_model = config.d_model
        self.embed_dim = embed_dim
        self.total_num_heads = num_heads
        self.total_num_kv_heads = self.total_num_heads
        self.head_dim = embed_dim // num_heads
        self.config = config

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(f"embed_dim must be divisible by num_heads "
                             f"(got `embed_dim`: {self.embed_dim}"
                             f" and `num_heads`: {num_heads}).")
        self.scaling = self.head_dim**-0.5

        # TP sharding sizes is accounted for within "*Parallel" layers.
        self.qkv_proj = QKVCrossParallelLinear(self.d_model,
                                               self.d_model //
                                               self.total_num_heads,
                                               self.total_num_heads,
                                               self.total_num_kv_heads,
                                               bias,
                                               quant_config=quant_config)

        self.out_proj = RowParallelLinear(
            embed_dim,
            embed_dim,
            bias=bias,
            quant_config=quant_config,
        )

        tp_world_size = get_tensor_model_parallel_world_size()
        assert self.total_num_heads % tp_world_size == 0
        self.num_heads = self.total_num_heads // tp_world_size

        if self.total_num_kv_heads >= tp_world_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_world_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_world_size % self.total_num_kv_heads == 0
        self.num_kv_heads = self.num_heads  # No GQA in bart
        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",
                              attn_type=AttentionType.ENCODER_DECODER)

    def forward(
        self,
        decoder_hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        """Input shape: Batch x Time x Channel"""

        q, k, v = self.qkv_proj(decoder_hidden_states, encoder_hidden_states)

        attn_output = self.attn(q, k, v)

        output, _ = self.out_proj(attn_output)
        return output

class BartDecoder(nn.Module):
    """
    Transformer decoder consisting of *config.decoder_layers* layers.
    Each layer is a [`BartDecoderLayer`]
    Args:
        config: BartConfig
        embed_tokens (nn.Embedding): output embedding
    """

    def __init__(
        self,
        config: BartConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        lora_config: Optional[LoRAConfig] = None,
        embed_tokens: Optional[nn.Embedding] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.cache_config = cache_config
        self.quant_config = quant_config
        self.lora_config = lora_config
        self.max_target_positions = config.max_position_embeddings
        embed_scale = math.sqrt(
            config.d_model) if config.scale_embedding else 1.0

        self.embed_tokens = BartScaledWordEmbedding(config.vocab_size,
                                                    config.d_model,
                                                    embed_scale=embed_scale)

        if embed_tokens is not None:
            self.embed_tokens.weight = embed_tokens.weight

        self.embed_positions = BartLearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
        )

        self.layers = nn.ModuleList(
            [BartDecoderLayer(config,cache_config,quant_config,
            prefix=f"{prefix}.layers.{layer_idx}") \
             for layer_idx in range(config.decoder_layers)])

        self.layernorm_embedding = nn.LayerNorm(config.d_model)

    def forward(
        self,
        decoder_input_ids: torch.Tensor,
        decoder_positions: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        r"""
        Args:
            decoder_input_ids
                Indices of *decoder* input sequence tokens in the vocabulary.
                Padding will be ignored by default should you
                provide it.
            decoder_positions
                Positions of *decoder* input sequence tokens.
            encoder_hidden_states:
                Tensor of encoder output embeddings
        Returns:
            Decoder output torch.Tensor
        """
        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(decoder_input_ids)
        else:
            decoder_positions = inputs_embeds[:, -1]

        # embed positions
        embed_pos = self.embed_positions(decoder_positions)
        embed_pos = embed_pos.to(inputs_embeds.device)

        hidden_states = inputs_embeds + embed_pos
        hidden_states = self.layernorm_embedding(hidden_states)

        # decoder layers

        for decoder_layer in self.layers:
            hidden_states = decoder_layer(
                decoder_hidden_states=hidden_states,
                encoder_hidden_states=encoder_hidden_states,
            )

        return hidden_states

class BartDecoderLayer(nn.Module):

    def __init__(
        self,
        config: BartConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = BartDecoderSelfAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            config=config,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
        )
        self.activation_fn = get_act_fn(config.activation_function)

        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        '''
        afeldman-nm: personally I would call this "cross-attention",
        however I left the name as "encoder_attn" to maintain consistency
        with the name of the pretrained weights.
        '''
        self.encoder_attn = BartCrossAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            config=config,
            prefix=f"{prefix}.encoder_attn",
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)

        ffn_hidden_size = self.embed_dim
        ffn_intermediate_size = config.encoder_ffn_dim
        ffn_has_bias = True
        self.fc1 = ColumnParallelLinear(
            ffn_hidden_size,
            ffn_intermediate_size,
            bias=ffn_has_bias,
            quant_config=quant_config,
        )
        self.fc2 = RowParallelLinear(
            ffn_intermediate_size,
            ffn_hidden_size,
            bias=ffn_has_bias,
            quant_config=quant_config,
        )

        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        decoder_hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        r"""
        Args:
            decoder_hidden_states
                torch.Tensor of *decoder* input embeddings.
            encoder_hidden_states
                torch.Tensor of *encoder* input embeddings.
        Returns:
            Decoder layer output torch.Tensor
        """
        residual = decoder_hidden_states

        # Self Attention
        hidden_states = self.self_attn(hidden_states=decoder_hidden_states)

        hidden_states = residual + hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Cross-Attention Block

        residual = hidden_states

        hidden_states = self.encoder_attn(
            decoder_hidden_states=hidden_states,
            encoder_hidden_states=encoder_hidden_states,
        )

        hidden_states = residual + hidden_states
        hidden_states = self.encoder_attn_layer_norm(hidden_states)

        # Fully Connected
        residual = hidden_states
        fc1_out, _ = self.fc1(hidden_states)
        hidden_states = self.activation_fn(fc1_out)

        hidden_states, _ = self.fc2(hidden_states)

        hidden_states = residual + hidden_states
        hidden_states = self.final_layer_norm(hidden_states)

        return hidden_states

class BartDecoderSelfAttention(nn.Module):

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        bias: bool = True,
        config: Optional[BartConfig] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.d_model = config.d_model
        self.embed_dim = embed_dim
        self.total_num_heads = num_heads
        self.total_num_kv_heads = self.total_num_heads
        self.head_dim = embed_dim // num_heads
        self.config = config

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(f"embed_dim must be divisible by num_heads "
                             f"(got `embed_dim`: {self.embed_dim}"
                             f" and `num_heads`: {num_heads}).")
        self.scaling = self.head_dim**-0.5

        self.qkv_proj = QKVParallelLinear(
            self.d_model,
            self.d_model // self.total_num_heads,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=bias,
            quant_config=quant_config,
        )

        self.out_proj = RowParallelLinear(
            embed_dim,
            embed_dim,
            bias=bias,
            quant_config=quant_config,
        )

        tp_world_size = get_tensor_model_parallel_world_size()
        assert self.total_num_heads % tp_world_size == 0
        self.num_heads = self.total_num_heads // tp_world_size

        if self.total_num_kv_heads >= tp_world_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_world_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_world_size % self.total_num_kv_heads == 0
        self.num_kv_heads = self.num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim

        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",
                              attn_type=AttentionType.DECODER)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        """Input shape: Batch x Time x Channel"""

        qkv, _ = self.qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)

        attn_output = self.attn(q, k, v)

        output, _ = self.out_proj(attn_output)
        return output

class BartEncoder(nn.Module):
    """
    Transformer encoder consisting of *config.encoder_layers*
    self attention layers. Each layer is a [`BartEncoderLayer`].
    Args:
        config: BartConfig
        embed_tokens (nn.Embedding): output embedding
    """

    def __init__(self,
                 config: BartConfig,
                 cache_config: Optional[CacheConfig] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 lora_config: Optional[LoRAConfig] = None,
                 embed_tokens: Optional[nn.Embedding] = None,
                 prefix: str = ""):
        super().__init__()

        self.cache_config = cache_config
        self.quant_config = quant_config
        self.lora_config = lora_config
        embed_dim = config.d_model
        self.max_source_positions = config.max_position_embeddings
        embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0

        self.embed_tokens = BartScaledWordEmbedding(config.vocab_size,
                                                    embed_dim,
                                                    embed_scale=embed_scale)

        if embed_tokens is not None:
            self.embed_tokens.weight = embed_tokens.weight

        self.embed_positions = BartLearnedPositionalEmbedding(
            config.max_position_embeddings,
            embed_dim,
        )
        self.layers = nn.ModuleList([
            BartEncoderLayer(config,
                             cache_config,
                             quant_config,
                             prefix=f"{prefix}.layers.{layer_idx}")
            for layer_idx in range(config.encoder_layers)
        ])

        self.layernorm_embedding = nn.LayerNorm(embed_dim)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        r"""
        Args:
            input_ids
                Indices of *encoder* input sequence tokens in the vocabulary.
                Padding will be ignored by default should you
                provide it.
            positions
                Positions of *encoder* input sequence tokens.
        Returns:
            Decoder output torch.Tensor
        """
        # retrieve input_ids and inputs_embeds
        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        embed_pos = self.embed_positions(positions)
        embed_pos = embed_pos.to(inputs_embeds.device)

        hidden_states = inputs_embeds + embed_pos
        hidden_states = self.layernorm_embedding(hidden_states)

        for encoder_layer in self.layers:
            hidden_states = encoder_layer(hidden_states=hidden_states)

        return hidden_states