`nemo_rl.data_plane.nixl.blob_format`#

Blob codec: many values → one self-describing byte blob, no per-tensor serialization.

Tensors and numpy arrays are stored as their raw contiguous bytes. What the bytes mean (kind, dtype, shape) lives in two places that we own: the per-entry location meta TransferQueue hands back on every get, and the blob’s trailing index so a unit can read its own blobs without the controller (scan, checkpoint shards). Only non-tensor objects are pickled.

Layout::

[ entry_0 | pad | entry_1 | pad | ... | index (msgpack) | footer 16 B ]

Entries are 64-byte aligned.

Module Contents#

Classes#

`Entry`
`BlobPlan`

Functions#

`_align_up`
`_tensor_bytes`
`_numpy_bytes`
`encode_value`
`plan`
`write_tail`	Write only the index and footer (`len(p.index_bytes) + FOOTER_SIZE` bytes).
`write`
`_torch_dtype`
`decode_entry`	Reconstruct a value from its bytes. Tensors and arrays are zero-copy views.
`materialize`	Copy a decoded zero-copy view out of the transfer buffer.
`_footer`	`(index_off, index_len, tag)` from a v3 or v2 footer at the end of `blob`.
`read_index`
`read_index_file`	func: `read_index` of the blob stored at `off` in binary file `f`.
`_entries`
`footer_is_valid`

Data#

`ALIGN`
`MAGIC`
`VERSION`
`_FOOTER_FMT`
`FOOTER_SIZE`
`_FOOTER_V2_FMT`
`FOOTER_V2_SIZE`
`NO_TAG`
`KIND_TENSOR`
`KIND_NUMPY`
`KIND_PICKLE`

API#

nemo_rl.data_plane.nixl.blob_format.ALIGN#: 64

nemo_rl.data_plane.nixl.blob_format.MAGIC#: 1314276432

nemo_rl.data_plane.nixl.blob_format.VERSION#: 3

nemo_rl.data_plane.nixl.blob_format._FOOTER_FMT#: ‘<IHHII16s’

nemo_rl.data_plane.nixl.blob_format.FOOTER_SIZE#: ‘calcsize(…)’

nemo_rl.data_plane.nixl.blob_format._FOOTER_V2_FMT#: ‘<IHHII’

nemo_rl.data_plane.nixl.blob_format.FOOTER_V2_SIZE#: ‘calcsize(…)’

nemo_rl.data_plane.nixl.blob_format.NO_TAG#: ‘bytes(…)’

nemo_rl.data_plane.nixl.blob_format.KIND_TENSOR#: ‘t’

nemo_rl.data_plane.nixl.blob_format.KIND_NUMPY#: ‘n’

nemo_rl.data_plane.nixl.blob_format.KIND_PICKLE#: ‘p’

nemo_rl.data_plane.nixl.blob_format._align_up(n: int) → int#

class nemo_rl.data_plane.nixl.blob_format.Entry#

key: str#: None

off: int#: None

len: int#: None

kind: str#: None

dtype: str | None#: None

shape: tuple[int, ...] | None#: None

meta() → dict[str, Any]#: Location meta stored by TQ as custom_backend_meta (blob id added by the client).

static from_meta( key: str, m: dict[str, Any], ) → nemo_rl.data_plane.nixl.blob_format.Entry#

class nemo_rl.data_plane.nixl.blob_format.BlobPlan#

keys: list[str]#: None

buffers: list[memoryview]#: None

keepalive: list[Any]#: None

entries: list[nemo_rl.data_plane.nixl.blob_format.Entry]#: None

index_bytes: bytes#: None

nbytes: int#: None

tag: bytes#: None

property index_off: int#

nemo_rl.data_plane.nixl.blob_format._tensor_bytes(t: torch.Tensor) → tuple[memoryview, Any]#

nemo_rl.data_plane.nixl.blob_format._numpy_bytes(a: numpy.ndarray) → tuple[memoryview, Any]#

nemo_rl.data_plane.nixl.blob_format.encode_value( key: str, v: Any, ) → tuple[memoryview, Any, str, str | None, tuple[int, ...] | None]#

nemo_rl.data_plane.nixl.blob_format.plan( keys: Sequence[str], values: Sequence[Any], tag: bytes = NO_TAG, ) → nemo_rl.data_plane.nixl.blob_format.BlobPlan#

nemo_rl.data_plane.nixl.blob_format.write_tail( p: nemo_rl.data_plane.nixl.blob_format.BlobPlan, out: memoryview, ) → None#

Write only the index and footer (len(p.index_bytes) + FOOTER_SIZE bytes).

For a blob whose values are sent separately. out maps to blob offset p.index_off.

nemo_rl.data_plane.nixl.blob_format.write( p: nemo_rl.data_plane.nixl.blob_format.BlobPlan, out: memoryview, ) → None#

nemo_rl.data_plane.nixl.blob_format._torch_dtype(name: str) → torch.dtype#

nemo_rl.data_plane.nixl.blob_format.decode_entry( view: memoryview, e: nemo_rl.data_plane.nixl.blob_format.Entry, ) → Any#: Reconstruct a value from its bytes. Tensors and arrays are zero-copy views.

nemo_rl.data_plane.nixl.blob_format.materialize(v: Any) → Any#: Copy a decoded zero-copy view out of the transfer buffer.

nemo_rl.data_plane.nixl.blob_format._footer(blob: memoryview) → tuple[int, int, bytes] | None#: (index_off, index_len, tag) from a v3 or v2 footer at the end of blob.

nemo_rl.data_plane.nixl.blob_format.read_index( blob: memoryview, ) → list[nemo_rl.data_plane.nixl.blob_format.Entry]#

nemo_rl.data_plane.nixl.blob_format.read_index_file( f: Any, off: int, nbytes: int, ) → list[nemo_rl.data_plane.nixl.blob_format.Entry]#

Func:: read_index of the blob stored at off in binary file f.

nemo_rl.data_plane.nixl.blob_format._entries( raw: bytes, ) → list[nemo_rl.data_plane.nixl.blob_format.Entry]#

nemo_rl.data_plane.nixl.blob_format.footer_is_valid(blob_tail: memoryview) → bool#

nemo_rl.data_plane.nixl.blob_format#