vllm.model_executor.models.deepseek_eagle3 ¶

class DeepseekV2Eagle3DecoderLayer(nn.Module):
    """
    Eagle3 decoder layer for Deepseek that:
    1. Always uses MLP (not MoE)
    2. First layer accepts concatenated embeds + hidden_states
    """

    def __init__(
        self,
        vllm_config: VllmConfig,
        prefix: str,
        config: DeepseekV2Config | DeepseekV3Config | None = None,
        layer_idx: int = 0,
    ) -> None:
        super().__init__()

        if config is None:
            config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = get_draft_quant_config(vllm_config)

        self.hidden_size = config.hidden_size
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings", 8192)

        self.layer_idx = layer_idx

        # MLA attention parameters
        qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0)
        qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0)
        v_head_dim = getattr(config, "v_head_dim", 0)
        kv_lora_rank = getattr(config, "kv_lora_rank", 0)
        config = copy.copy(config)
        if rope_scaling:
            rope_params = rope_scaling.copy()
            rope_params["rope_type"] = "deepseek_yarn"
        else:
            rope_params = {"rope_type": "default"}
        config.rope_parameters = rope_params
        self.self_attn = DeepseekV2MLAAttention(
            vllm_config=vllm_config,
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            qk_nope_head_dim=qk_nope_head_dim,
            qk_rope_head_dim=qk_rope_head_dim,
            v_head_dim=v_head_dim,
            q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None,
            kv_lora_rank=kv_lora_rank,
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
            input_size=2 * self.hidden_size if layer_idx == 0 else self.hidden_size,
        )

        # Always use MLP (not MoE) for Eagle3
        self.mlp = DeepseekV2MLP(
            hidden_size=config.hidden_size,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            quant_config=quant_config,
            prefix=f"{prefix}.mlp",
        )

        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
        )

        self.hidden_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        if getattr(config, "norm_before_residual", False):
            self._residual_norm = self._norm_before_residual
        else:
            self._residual_norm = self._norm_after_residual

    def _norm_before_residual(
        self, hidden_states: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor]:
        hidden_states = self.hidden_norm(hidden_states)
        residual = hidden_states
        return hidden_states, residual

    def _norm_after_residual(
        self, hidden_states: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor]:
        residual = hidden_states
        hidden_states = self.hidden_norm(hidden_states)
        return hidden_states, residual

    def forward(
        self,
        positions: torch.Tensor,
        embeds: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: torch.Tensor | None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        if self.layer_idx == 0:
            # First layer: concatenate embeds with hidden_states
            embeds = self.input_layernorm(embeds)
            hidden_states, residual = self._residual_norm(hidden_states=hidden_states)
            hidden_states = torch.cat([embeds, hidden_states], dim=-1)
        else:
            # Subsequent layers: process hidden_states and residuals only
            hidden_states, residual = self.input_layernorm(hidden_states, residual)

        # Self Attention
        hidden_states = self.self_attn(
            positions=positions,
            hidden_states=hidden_states,
            llama_4_scaling=None,
        )

        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)

        # Fully Connected (MLP, not MoE)
        hidden_states = self.mlp(hidden_states)

        return hidden_states, residual

class Eagle3DeepseekV2ForCausalLM(DeepseekV2ForCausalLM):
    """Eagle3 speculative decoding model for DeepseekV2/V3."""

    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
        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
        )

        # Store target layer count in draft config
        self.config.target_layer_count = target_layer_num

        self.model = DeepseekV2Eagle3Model(
            vllm_config=vllm_config, prefix="model", start_layer_id=target_layer_num
        )

        logit_scale = getattr(self.config, "logit_scale", 1.0)
        self.lm_head = ParallelLMHead(
            self.config.draft_vocab_size,
            self.config.hidden_size,
            prefix=maybe_prefix(prefix, "lm_head"),
        )
        self.logits_processor = LogitsProcessor(
            self.config.draft_vocab_size, scale=logit_scale
        )
        self.draft_id_to_target_id = nn.Parameter(
            torch.zeros(self.config.draft_vocab_size, dtype=torch.long),
            requires_grad=False,
        )

    def embed_input_ids(
        self,
        input_ids: torch.Tensor,
        multimodal_embeddings: NestedTensors | None = None,
        is_multimodal: torch.Tensor | None = None,
    ) -> torch.Tensor:
        return self.model.embed_input_ids(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]:
        return self.model(input_ids, positions, hidden_states, inputs_embeds)

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor | None:
        logits = self.logits_processor(self.lm_head, hidden_states)
        if self.draft_id_to_target_id is None:
            assert logits.shape[1] == self.config.vocab_size, (
                "Expected logits to have shape "
                f"(*, {self.config.vocab_size}), but got {logits.shape}"
            )
            return logits

        base = torch.arange(self.config.draft_vocab_size, device=logits.device)
        targets = base + self.draft_id_to_target_id
        logits_new = logits.new_full(
            (
                logits.shape[0],
                self.config.vocab_size,
            ),
            float("-inf"),
        )
        logits_new[:, targets] = logits
        return logits_new

    def combine_hidden_states(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        # Combine multiple auxiliary hidden states returned by Eagle3
        return self.model.fc(hidden_states)

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
        model_weights = {}
        includes_draft_id_mapping = False
        includes_embed_tokens = False

        for name, loaded_weight in weights:
            if "t2d" in name:
                continue
            if "d2t" in name:
                name = name.replace("d2t", "draft_id_to_target_id")
                includes_draft_id_mapping = True
            elif "lm_head" not in name:
                name = "model." + name
            if "embed_tokens" in name:
                includes_embed_tokens = True
            model_weights[name] = loaded_weight
            process_eagle_weight(self, name)

        skip_substrs = []
        if not includes_draft_id_mapping:
            skip_substrs.append("draft_id_to_target_id")
        if not includes_embed_tokens:
            skip_substrs.append("embed_tokens")

        loader = AutoWeightsLoader(
            self,
            skip_prefixes=None,
            skip_substrs=skip_substrs,
        )
        loader.load_weights(model_weights.items())