vllm.model_executor.model_loader.weight_cache.artifact_store

Storage backends for daemon warmup-cache artifacts (disk and shared stores).

Classes:

• ArtifactStore –

  One warmup-cache artifact store: bytes in and out, keyed by artifact_key.

• DiskArtifactStore –

  The local disk store: one directory tree per warmup-cache kind.

• MooncakeArtifactStore –

  MooncakeDistributedStore wrapper via MOONCAKE_CONFIG_PATH.

Functions:

• artifact_key –

  Canonical key for one warmup-cache artifact, valid in every store.

ArtifactStore

Bases: Protocol

One warmup-cache artifact store: bytes in and out, keyed by artifact_key.

Methods:

• get –

  Fetch the artifact, or None on a miss.

• put –

  Publish the artifact under the key.

Source code in vllm/model_executor/model_loader/weight_cache/artifact_store.py

class ArtifactStore(Protocol):
    """One warmup-cache artifact store: bytes in and out, keyed by artifact_key."""

    name: str

    def get(self, key: str) -> bytes | None:
        """Fetch the artifact, or None on a miss."""
        ...

    def put(self, key: str, data: bytes) -> None:
        """Publish the artifact under the key."""
        ...

get(key)

Fetch the artifact, or None on a miss.

Source code in vllm/model_executor/model_loader/weight_cache/artifact_store.py

def get(self, key: str) -> bytes | None:
    """Fetch the artifact, or None on a miss."""
    ...

put(key, data)

Publish the artifact under the key.

Source code in vllm/model_executor/model_loader/weight_cache/artifact_store.py

def put(self, key: str, data: bytes) -> None:
    """Publish the artifact under the key."""
    ...

DiskArtifactStore

Bases: ArtifactStore

The local disk store: one directory tree per warmup-cache kind.

Source code in vllm/model_executor/model_loader/weight_cache/artifact_store.py

class DiskArtifactStore(ArtifactStore):
    """The local disk store: one directory tree per warmup-cache kind."""

    name = "disk"

    def __init__(self, kind: str, root: Path):
        self._prefix = f"{_KEY_PREFIX}/{kind}/"
        self._root = root

    def _path(self, key: str) -> Path:
        rel = key.removeprefix(self._prefix)
        parts = rel.split("/")
        if rel == key or len(parts) != 2 or any(p in ("", ".", "..") for p in parts):
            raise ValueError(f"key {key!r} is outside this store")
        return self._root / parts[0] / parts[1]

    def get(self, key: str) -> bytes | None:
        path = self._path(key)
        return path.read_bytes() if path.is_file() else None

    def put(self, key: str, data: bytes) -> None:
        path = self._path(key)
        path.parent.mkdir(parents=True, exist_ok=True)
        fd, tmp_path = tempfile.mkstemp(
            dir=path.parent, suffix=".tmp", prefix=f".{path.name}."
        )
        try:
            with os.fdopen(fd, "wb") as f:
                f.write(data)
            os.replace(tmp_path, path)
        except BaseException:
            with suppress(OSError):
                os.unlink(tmp_path)
            raise

MooncakeArtifactStore

Bases: ArtifactStore

MooncakeDistributedStore wrapper via MOONCAKE_CONFIG_PATH.

Methods:

• get –

  Fetch the artifact; a miss or a native error reads as empty.

• put –

  Publish the artifact; first writer wins, so publishers are idempotent.

Source code in vllm/model_executor/model_loader/weight_cache/artifact_store.py

class MooncakeArtifactStore(ArtifactStore):
    """MooncakeDistributedStore wrapper via MOONCAKE_CONFIG_PATH."""

    name = "mooncake"

    def __init__(self) -> None:
        self._store = MooncakeDistributedStore()
        config = MooncakeStoreConfig.load_from_config()
        local_hostname = rdma_utils.get_requester_local_hostname(get_ip())
        setup_mooncake_store(self._store, config, local_hostname)

    def get(self, key: str) -> bytes | None:
        """Fetch the artifact; a miss or a native error reads as empty."""
        return self._store.get(key) or None

    def put(self, key: str, data: bytes) -> None:
        """Publish the artifact; first writer wins, so publishers are idempotent."""
        if self._store.is_exist(key) == 1:
            return
        ret = self._store.put(key, data)
        if ret != 0 and self._store.is_exist(key) != 1:
            logger.warning("Mooncake artifact put %s failed (ret=%d)", key, ret)

get(key)

Fetch the artifact; a miss or a native error reads as empty.

Source code in vllm/model_executor/model_loader/weight_cache/artifact_store.py

def get(self, key: str) -> bytes | None:
    """Fetch the artifact; a miss or a native error reads as empty."""
    return self._store.get(key) or None

put(key, data)

Publish the artifact; first writer wins, so publishers are idempotent.

Source code in vllm/model_executor/model_loader/weight_cache/artifact_store.py

def put(self, key: str, data: bytes) -> None:
    """Publish the artifact; first writer wins, so publishers are idempotent."""
    if self._store.is_exist(key) == 1:
        return
    ret = self._store.put(key, data)
    if ret != 0 and self._store.is_exist(key) != 1:
        logger.warning("Mooncake artifact put %s failed (ret=%d)", key, ret)

artifact_key(kind, config_hash, filename)

Canonical key for one warmup-cache artifact, valid in every store.

Source code in vllm/model_executor/model_loader/weight_cache/artifact_store.py

def artifact_key(kind: str, config_hash: str, filename: str) -> str:
    """Canonical key for one warmup-cache artifact, valid in every store."""
    return f"{_KEY_PREFIX}/{kind}/{config_hash}/{filename}"