vllm.model_executor.models.olmo ¶

class OlmoAttention(nn.Module):
    """
    This is the attention block where the output is computed as
    `Attention(LN(x))` in `MLP(LN(x + Attention(LN(x))))`
    (plus another skip connection).
    """

    def __init__(
        self,
        config: OlmoConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.num_attention_heads

        assert self.hidden_size % self.total_num_heads == 0
        assert self.total_num_heads % tensor_model_parallel_world_size == 0

        self.num_heads = self.total_num_heads // tensor_model_parallel_world_size
        self.head_dim = self.hidden_size // self.total_num_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.clip_qkv = config.clip_qkv

        # Attention input projection. Projects x -> (q, k, v)
        self.qkv_proj = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            bias=config.attention_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

        # Rotary embeddings.
        self.rotary_emb = get_rope(
            self.head_dim,
            max_position=self.max_position_embeddings,
            rope_parameters=config.rope_parameters,
        )
        self.scaling = self.head_dim**-0.5
        self.attn = Attention(
            self.num_heads,
            self.head_dim,
            scale=self.scaling,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

        # Attention output projection.
        self.o_proj = RowParallelLinear(
            self.hidden_size,
            self.hidden_size,
            bias=config.attention_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.o_proj",
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(hidden_states)
        if self.clip_qkv is not None:
            qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.o_proj(attn_output)
        return output

class OlmoDecoderLayer(nn.Module):
    """
    This is a typical transformer block where the output is
    computed as `MLP(LN(x + Attention(LN(x))))`
    (plus another skip connection).
    """

    def __init__(
        self,
        config: OlmoConfig,
        cache_config: CacheConfig | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        # Attention block.
        self.self_attn = OlmoAttention(
            config, cache_config, quant_config, prefix=f"{prefix}.self_attn"
        )

        # MLP block.
        self.mlp = OlmoMLP(config, quant_config, prefix=f"{prefix}.mlp")

        # LayerNorm
        self.input_layernorm = nn.LayerNorm(
            config.hidden_size, elementwise_affine=False, bias=False
        )
        self.post_attention_layernorm = nn.LayerNorm(
            config.hidden_size, elementwise_affine=False, bias=False
        )

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
        # Attention block.
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(positions, hidden_states)
        hidden_states = hidden_states + residual

        # MLP block.
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states

class OlmoForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
    """
    Extremely barebones HF model wrapper.
    """

    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_stacked={
            # weight_name: (param_name, shard_id)
            ".q_proj": (".qkv_proj", "q"),
            ".k_proj": (".qkv_proj", "k"),
            ".v_proj": (".qkv_proj", "v"),
            ".gate_proj": (".gate_up_proj", 0),
            ".up_proj": (".gate_up_proj", 1),
        }
    )
    packed_modules_mapping = {
        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
        "gate_up_proj": ["gate_proj", "up_proj"],
    }

    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.model = OlmoModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        if config.tie_word_embeddings:
            self.lm_head = self.model.embed_tokens
        else:
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "lm_head"),
            )
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.embed_input_ids(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
        hidden_states = self.model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
        )
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor | None:
        logits = self.logits_processor(self.lm_head, hidden_states)
        return logits

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(
            self,
            skip_prefixes=(
                ["lm_head.weight"] if self.config.tie_word_embeddings else None
            ),
        )
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

class OlmoMLP(nn.Module):
    """
    This is the MLP block where the output is computed as
    `MLP(LN(x))` in `MLP(LN(x + Attention(LN(x))))`
    (plus another skip connection).
    """

    def __init__(
        self,
        config: OlmoConfig,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size

        # Feed-forward input projection.
        self.gate_up_proj = MergedColumnParallelLinear(
            self.hidden_size,
            [self.intermediate_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )

        # Activation function.
        self.act_fn = SiluAndMul()

        # Feed-forward output projection.
        self.down_proj = RowParallelLinear(
            self.intermediate_size,
            self.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.down_proj",
        )

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

@support_torch_compile
class OlmoModel(nn.Module):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.config = config

        self.embed_tokens = VocabParallelEmbedding(
            config.vocab_size, config.hidden_size
        )
        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: OlmoDecoderLayer(
                config, cache_config, quant_config, prefix=prefix
            ),
            prefix=f"{prefix}.layers",
        )
        self.norm = nn.LayerNorm(
            config.hidden_size, elementwise_affine=False, bias=False
        )
        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
            ["hidden_states"], config.hidden_size
        )

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
        """
        Args:
            input_ids: A tensor of shape `(batch_size, seq_len)`.
        """
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.embed_input_ids(input_ids)
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]

        # Apply blocks one-by-one.
        for layer in islice(self.layers, self.start_layer, self.end_layer):
            # shape: (batch_size, seq_len, d_model)
            hidden_states = layer(positions, hidden_states)

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})
        # Apply final layer norm.
        # shape: (batch_size, seq_len or 1, d_model)
        hidden_states = self.norm(hidden_states)
        return hidden_states