vllm.model_executor.models.llama_eagle ¶

class EagleLlamaForCausalLM(LlamaForCausalLM):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        nn.Module.__init__(self)
        self.config = vllm_config.speculative_config.draft_model_config.hf_config
        # Ensure draft_vocab_size is set
        # default to the base vocab size when absent
        if getattr(self.config, "draft_vocab_size", None) is None:
            base_vocab_size = getattr(self.config, "vocab_size", None)
            self.config.draft_vocab_size = base_vocab_size
        target_layer_num = vllm_config.model_config.get_num_layers(
            vllm_config.parallel_config
        )
        self.model = LlamaModel(
            vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num
        )

        logit_scale = getattr(self.config, "logit_scale", 1.0)
        self.logits_processor = LogitsProcessor(
            self.config.vocab_size, scale=logit_scale
        )

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        inputs_embeds: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if inputs_embeds is not None:
            raise NotImplementedError(
                f"{type(self).__name__} does not support multimodal inputs yet."
            )
        return self.model(input_ids, positions, hidden_states)

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
        def transform(inputs):
            name, loaded_weight = inputs
            if "lm_head" not in name:
                name = "model." + name
            return name, loaded_weight

        loader = AutoWeightsLoader(
            self,
            skip_prefixes=None,
        )
        loader.load_weights(map(transform, weights))

@support_torch_compile
class LlamaModel(nn.Module):
    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        start_layer_id: int = 0,
    ) -> None:
        super().__init__()
        self.config = vllm_config.speculative_config.draft_model_config.hf_config
        self.vocab_size = self.config.vocab_size

        self.embed_tokens = VocabParallelEmbedding(
            self.config.vocab_size,
            self.config.hidden_size,
            prefix=maybe_prefix(prefix, "embed_tokens"),
        )

        self.layers = nn.ModuleList(
            [
                LlamaDecoderLayer(
                    vllm_config,
                    i == 0,
                    prefix=maybe_prefix(prefix, f"layers.{i + start_layer_id}"),
                    config=self.config,
                )
                for i in range(self.config.num_hidden_layers)
            ]
        )
        self.fc = torch.nn.Linear(
            self.config.hidden_size * 2, self.config.hidden_size, bias=False
        )

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        input_embeds = self.embed_tokens(input_ids)
        hidden_states = self.fc(torch.cat((input_embeds, hidden_states), dim=-1))
        residual = None
        for layer in self.layers:
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )
        hidden_states = hidden_states + residual
        return hidden_states, hidden_states

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            (".gate_up_proj", ".gate_proj", 0),
            (".gate_up_proj", ".up_proj", 1),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            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)
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # if PP disabled then draft will share embed with target
                if get_pp_group().world_size == 1 and "embed_tokens." in name:
                    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