vllm.config.engram ¶

@config
class EngramConfig:
    """Configuration for Engram embedding storage and sharding."""

    cpu_offload: bool = Field(default_factory=_default_cpu_offload)
    """Store embedding weights in pinned CPU memory for UVA lookup.
    Defaults to VLLM_PLE_CPU_OFFLOAD, which is enabled by default. An explicit
    value takes precedence over the environment variable."""

    embedding_across_dp: bool = False
    """Shard embeddings across TP and all DP ranks when enabled.
    Otherwise, each DP rank has a separate TP-sharded embedding replica."""

    dp_shared_memory: bool = False
    """Share CPU-offloaded embedding weights between co-located
    DP replicas. Each node stores one copy of every TP shard, reducing host
    memory without per-step Engram DP collectives. Requires sufficient
    /dev/shm capacity and a shared IPC namespace."""

    @model_validator(mode="after")
    def _validate_shared_memory(self) -> Self:
        if self.dp_shared_memory and not self.cpu_offload:
            raise ValueError("dp_shared_memory requires cpu_offload=True")
        return self

    def verify_model_config(self, model_config: "ModelConfig | None") -> None:
        """Reject Engram configuration for models without n-gram embeddings."""
        from vllm.platforms import current_platform

        field = (
            _NGRAM_LAYER_FIELDS.get(model_config.architecture)
            if model_config is not None
            else None
        )
        if (
            model_config is None
            or field is None
            or not current_platform.is_cuda()
            or not getattr(model_config.hf_text_config, field, None)
        ):
            raise ValueError(
                "EngramConfig requires a model with supported Engram "
                "embeddings, non-empty n-gram layer ids, and CUDA."
            )

    def verify_parallel_config(self, parallel_config: "ParallelConfig") -> None:
        """Reject unsupported embedding parallel topologies."""
        if self.dp_shared_memory:
            if parallel_config.data_parallel_size <= 1:
                raise ValueError("dp_shared_memory requires data_parallel_size > 1.")
            if parallel_config.enable_elastic_ep:
                raise ValueError("dp_shared_memory is not supported with elastic EP.")
        if (
            self.embedding_across_dp
            and parallel_config.data_parallel_size > 1
            and parallel_config.enable_elastic_ep
        ):
            raise ValueError(
                "Engram embedding_across_dp is not supported with elastic EP yet."
            )

    def verify_load_config(self, load_config: "LoadConfig") -> None:
        """Shared tables require a loader that invokes parameter weight callbacks."""
        if self.dp_shared_memory and load_config.load_format not in (
            "auto",
            "safetensors",
            "pt",
        ):
            raise ValueError(
                "dp_shared_memory requires load_format 'auto', "
                f"'safetensors' or 'pt'; got {load_config.load_format!r}."
            )

    def get_parallel_size(self, parallel_config: "ParallelConfig") -> int:
        """Derive the embedding group size from the parallel configuration."""
        size = parallel_config.tensor_parallel_size
        if self.embedding_across_dp and parallel_config.data_parallel_size > 1:
            size *= parallel_config.data_parallel_size
        return size

    def compute_hash(self) -> str:
        """Hash settings that affect embedding execution and graph structure."""
        return hash_factors(get_hash_factors(self, set()))