--- title: "TensorsContainer" sidebar_label: "TensorsContainer" description: "The clika_runtime.io TensorsContainer class." --- A checkpoint's tensors, materialized on demand: get(name) reads one entry onto the load policy's device, load_all() reads every entry, prefetch() starts reads ahead of use, evict() drops a materialized entry back to disk. names(), len() and `in` read the header only. ## `__init__` ```python __init__(self, /, *args, **kwargs) ``` Initialize self. See help(type(self)) for accurate signature. ## `contains` ```python containscontains(self, name: str) -> bool ``` contains(self, name: str) -> bool contains(name) -> bool Whether `name` is an entry. ## `entry` ```python entryentry(self, name: str) -> clika_runtime._core.io.ContainerEntry | None ``` entry(self, name: str) -> clika_runtime._core.io.ContainerEntry | None entry(name) -> ContainerEntry | None The parsed metadata of one entry with no payload IO; None for an unknown or a composed name. ## `entry_bytes` ```python entry_bytesentry_bytes(self, name: str) -> int ``` entry_bytes(self, name: str) -> int entry_bytes(name) -> int The on-disk payload bytes of one entry. ## `evict` ```python evictevict(self, name: str) -> None ``` evict(self, name: str) -> None evict(name) -> None Drop a materialized entry back to disk; the next get() reads it again. ## `evict_all` ```python evict_allevict_all(self) -> None ``` evict_all(self) -> None evict_all() -> None Drop every materialized entry back to disk. ## `get` ```python getget(self, name: str) -> clika_runtime._core.Tensor ``` get(self, name: str) -> clika_runtime._core.Tensor get(name) -> Tensor The entry materialized on the load policy's device (read from disk on first use, then cached until evicted). ## `keys` ```python keyskeys(self) -> list[str] ``` keys(self) -> list[str] keys() -> list[str] Every entry name (the dict spelling of names()). ## `load_all` ```python load_allload_all(self) -> dict ``` load_all(self) -> dict load_all() -> dict[str, Tensor] Every entry materialized, as a dict. ## `load_into` ```python load_intoload_into(self, name: str, dst: clika_runtime._core.Tensor, where: object | None = None) -> None ``` load_into(self, name: str, dst: clika_runtime._core.Tensor, where: object | None = None) -> None load_into(name, dst, where=None) -> None Read the entry into an existing tensor of the same shape and dtype, which ends carrying the entry's quantization scheme (none for a plain entry). ## `names` ```python namesnames(self) -> list[str] ``` names(self) -> list[str] names() -> list[str] Every entry name, header order. ## `open` ```python openopen(path: str, *, device: object | None = None, sticky: bool = False) -> clika_runtime._core.io.TensorsContainer ``` open(path: str, *, device: object | None = None, sticky: bool = False) -> clika_runtime._core.io.TensorsContainer open(path, *, device=None, sticky=False) -> TensorsContainer Open a checkpoint by its suffix: .safetensors, a torch checkpoint (.pt / .pth / .bin / .ckpt) or .gguf. Entries materialize on `device` (a Device, a device string or a Stream; None = the CPU); `sticky` keeps their lineage on disk so derived tensors can be evicted and re-read. ## `prefetch` ```python prefetchprefetch(self, names: collections.abc.Sequence[str], where: object | None = None) -> None ``` prefetch(self, names: collections.abc.Sequence[str], where: object | None = None) -> None prefetch(names, where=None) -> None Start reading the named entries ahead of use, onto `where` (a Device, a device string or a Stream; None = the load policy's device). ## `release` ```python releaserelease(self, name: str) -> None ``` release(self, name: str) -> None release(name) -> None Drop the container's own handle to a materialized entry; tensors already handed out stay valid. ## `size` ```python sizesize(self) -> int ``` size(self) -> int size() -> int The entry count. ## `subset` ```python subsetsubset(self, prefix: str | None = None, *, regex: str | None = None, strip: bool = True, return_clone: bool = False) -> clika_runtime._core.io.TensorsContainer ``` subset(self, prefix: str | None = None, *, regex: str | None = None, strip: bool = True, return_clone: bool = False) -> clika_runtime._core.io.TensorsContainer subset(prefix=None, *, regex=None, strip=True, return_clone=False) -> TensorsContainer A sub-container over the entries a prefix selects (every name starting with it) or a regular expression selects (RE2 syntax, matched against the whole name), the one a block loads its weights from by its own names (`load_state_dict`'s strict check then reads the sub-container's names alone). strip=True serves the entries with the prefix removed, or under the pattern's first capture group (a pattern with no group raises); strip=False keeps the full names. By default the entries MOVE out of this container (one owner per name: they leave names(), and a second subset of the same names is empty); return_clone=True shares them and this container keeps serving them too (a weight two modules bind). No tensor data is read or copied either way, and a weight read through both containers is one tensor. An empty selection is an empty sub-container; two entries producing one name, an empty produced name or a pattern that does not compile raise. ## `total_bytes` ```python total_bytestotal_bytes(self) -> int ``` total_bytes(self) -> int total_bytes() -> int The payload bytes of every entry, summed.