vllm.v1.worker.gpu.spec_decode.gemma4.speculator ¶

class Gemma4Speculator(AutoRegressiveSpeculator):
    @property
    def advance_draft_positions(self) -> bool:
        # Gemma4 MTP is Q-only and reads K/V from the target's existing cache.
        # No new KV slots are written, so positions and seq_lens stay fixed.
        return False

    def load_draft_model(
        self,
        target_model: nn.Module,
        target_attn_layer_names: set[str],
    ) -> nn.Module:
        draft_vllm_config = self._create_draft_vllm_config()
        with set_model_tag("eagle_head"):
            draft_model = get_model(
                vllm_config=draft_vllm_config,
                model_config=self.speculative_config.draft_model_config,
                load_config=self.speculative_config.draft_load_config,
            )
        self._setup_gemma4_kv_sharing(draft_model, target_attn_layer_names)
        self._share_embeddings(draft_model, target_model)
        return draft_model

    def _create_draft_vllm_config(self) -> VllmConfig:
        """Preserve the target's forced TRITON_ATTN backend for draft layers.

        Gemma4 forces TRITON_ATTN due to heterogeneous head dimensions
        (head_dim=256 sliding, global_head_dim=512 full). The base class
        resets attention_config.backend to None for draft models, causing
        sliding layers to fall back to FLASH_ATTN which cannot handle
        KV-shared cache. Override to carry the target's backend through.
        """
        draft_model_config = self.speculative_config.draft_model_config
        draft_vllm_config = replace(
            self.vllm_config,
            model_config=draft_model_config,
        )
        target_backend = self.vllm_config.attention_config.backend
        if target_backend is not None:
            draft_vllm_config = replace(
                draft_vllm_config,
                attention_config=replace(
                    draft_vllm_config.attention_config,
                    backend=target_backend,
                ),
            )
        return draft_vllm_config

    def _setup_gemma4_kv_sharing(
        self,
        model: nn.Module,
        target_attn_layer_names: set[str],
    ) -> None:
        """Wire draft layers to share KV and KV scales with the target model.

        Each draft decoder layer is mapped to the last non-KV-shared
        target layer of the same attention type (sliding or full).
        """
        draft_config = self.speculative_config.draft_model_config.hf_config
        draft_text_config = draft_config.get_text_config()
        target_config = self.vllm_config.model_config.hf_config
        target_text_config = target_config.get_text_config()
        target_layer_types = getattr(target_text_config, "layer_types", [])

        if not (hasattr(model, "model") and hasattr(model.model, "layers")):
            return

        target_num_kv_shared = getattr(target_text_config, "num_kv_shared_layers", 0)
        num_non_shared = len(target_layer_types) - target_num_kv_shared
        target_names_by_index: dict[int, str] = {}
        for name in target_attn_layer_names:
            _, separator, layer_suffix = name.partition(".layers.")
            if not separator:
                continue
            layer_index, _, _ = layer_suffix.partition(".")
            if layer_index.isdigit():
                target_names_by_index[int(layer_index)] = name

        type_to_target_names: dict[str, list[str]] = defaultdict(list)
        for idx, lt in enumerate(target_layer_types[:num_non_shared]):
            if target_name := target_names_by_index.get(idx):
                type_to_target_names[lt].append(target_name)

        draft_layer_types = getattr(draft_text_config, "layer_types", [])
        for draft_idx, layer in enumerate(model.model.layers):
            if not hasattr(layer, "self_attn"):
                continue
            attn = getattr(layer.self_attn, "attn", None)
            if attn is None:
                continue

            draft_layer_type = (
                draft_layer_types[draft_idx]
                if draft_idx < len(draft_layer_types)
                else "full_attention"
            )
            candidates = type_to_target_names.get(draft_layer_type, [])
            if not candidates:
                logger.warning(
                    "No target layer of type '%s' for draft layer %d",
                    draft_layer_type,
                    draft_idx,
                )
                continue

            target_layer_name = candidates[-1]
            attn.kv_sharing_target_layer_name = target_layer_name

            # KV-cache sharing aliases the cache tensor during allocation, but
            # the quantization scales live on the Attention modules themselves.
            # The BF16 draft model has no quantization config, so its K/V scales
            # otherwise remain at the default 1.0 while it reads the target's
            # calibrated FP8 cache.
            target_attn = self.vllm_config.compilation_config.static_forward_context[
                target_layer_name
            ]
            _copy_target_kv_scales(attn, target_attn)

            logger.info(
                "Gemma4 MTP: draft layer %d (%s) -> %s",
                draft_idx,
                draft_layer_type,
                target_layer_name,
            )

    def _share_embeddings(
        self,
        draft_model: nn.Module,
        target_model: nn.Module,
    ) -> None:
        target_language_model = (
            target_model.get_language_model()
            if hasattr(target_model, "get_language_model")
            else target_model
        )
        if get_pp_group().world_size == 1:
            target_embed = getattr(target_language_model.model, "embed_tokens", None)
            if target_embed is not None:
                del draft_model.model.embed_tokens
                draft_model.model.embed_tokens = target_embed