Ayush Chaurasia
242ade8017
## How packing works
Consider four tokenized documents:
[1]
[2]
[10, 11, 12, 13, 14, 15, 16, 17]
[20]
With:
```
StreamingDataset(
table,
shuffle=False,
columns=["tokens"],
num_splits=2,
pack_sequences=5,
eos_id=9,
pad_id=0,
blocks_per_epoch=6,
)
```
the documents are assigned to two fixed logical splits. Each split
maintains an independent token buffer, appends eos_id after every
document, and emits blocks of five tokens.
Because blocks_per_epoch=6, each split emits exactly three blocks:
Cycl/e 1:
Split 0: [1, 9, 2, 9, 0] # 9 is eos, 0 is padding
Split 1: [10, 11, 12, 13, 14]
Cycle 2:
Split 0: [0, 0, 0, 0, 0]
Split 1: [15, 16, 17, 9, 20]
Cycle 3:
Split 0: [0, 0, 0, 0, 0]
Split 1: [9, 0, 0, 0, 0]
If a split runs out of tokens early, it emits padded blocks through the
fixed budget. This prevents one rank from finishing before another.
Logical splits are independent of rank and worker ownership. A
checkpoint records each split’s consumed-document count, emitted-block
count, remaining tokens, and document boundaries. Merging
those per-split states allows the same packed stream to resume after the
topology changes.
doc_ids identifies document segments, including continuations across
block boundaries. It is not a padding mask: padding retains the
preceding document ID, so callers must mask padding using a
reserved pad_id.
blocks_per_epoch="auto" is also available. It estimates the budget from
a deterministic bounded sample and warns that the result is approximate.
WIP pre-training tests:
```
┌────────────────────────────────────┬────────────────────────┬─────────────────────────────────────┐
│ │ GPT-2 124M │ GPT-2 medium 354M │
├────────────────────────────────────┼────────────────────────┼─────────────────────────────────────┤
│ Corpus │ 2.4M docs / 12GB table │ 9.67M docs / 45GB table │
│ Tokens (Chinchilla) │ 2.43B │ 7.0B │
├────────────────────────────────────┼────────────────────────┼─────────────────────────────────────┤
│ Data prep (ingest→curate→tokenize) │ ~12 min │ ~51 min │
├────────────────────────────────────┼────────────────────────┼─────────────────────────────────────┤
│ Training wall time │ ~50 min │ 3h 06m │
├────────────────────────────────────┼────────────────────────┼─────────────────────────────────────┤
│ Throughput / MFU │ 1.60M tok/s / 35% │ 684k tok/s / 42.0%, │
├────────────────────────────────────┼────────────────────────┼─────────────────────────────────────┤
│ Final val loss │ 3.230 │ 2.840 │
└────────────────────────────────────┴────────────────────────┴─────────────────────────────────────┘
```
---------
Co-authored-by: OpenAI Codex <codex@openai.com>
LanceDB Python SDK
A Python library for LanceDB.
Installation
pip install lancedb
Pre-Haswell x86_64 hosts: lancedb-compat
The default lancedb wheel targets x86-64-haswell (AVX2 + FMA + F16C) for full performance on modern hardware. Pre-Haswell hosts — Intel Sandy Bridge / Ivy Bridge / Westmere; AMD Bulldozer / Piledriver / Steamroller — don't have AVX2 and crash with Illegal instruction at import lancedb.
For those hosts, install the lancedb-compat package instead:
pip install lancedb-compat
Same Python API (import lancedb works as usual). The compat wheel is compiled at the x86-64-v2 baseline (Nehalem-class) and uses runtime SIMD dispatch in the embedded lance crate to pick the right kernel tier (scalar / AVX / AVX+FMA / AVX2+FMA / AVX-512) at load time, so it still goes fast on modern hardware while running cleanly on the pre-Haswell silicon. Use lance.simd_info() from Python to verify which tier was selected.
lancedb and lancedb-compat install to the same lancedb/ namespace and conflict at install time. Pick one. To switch, pip uninstall lancedb first, then pip install lancedb-compat (or vice-versa).
If you need a custom baseline (or lancedb-compat isn't yet published for your platform), build from source with the override:
RUSTFLAGS="-C target-cpu=x86-64-v2" maturin build --release
pip install ./target/wheels/lancedb-*.whl
Preview Releases
Stable releases are created about every 2 weeks. For the latest features and bug fixes, you can install the preview release. These releases receive the same level of testing as stable releases, but are not guaranteed to be available for more than 6 months after they are released. Once your application is stable, we recommend switching to stable releases.
pip install --pre --extra-index-url https://pypi.fury.io/lancedb/ lancedb
Threading in CPU-limited containers
LanceDB uses separate pools for compute work and storage I/O. On a container with two visible CPUs, current releases intentionally use one compute worker by default; no manual configuration is needed. If every query logs an I/O core reservation warning on a two-CPU container, upgrade from LanceDB 0.21.1 or earlier.
The two commonly tuned environment variables control different resources:
LANCE_CPU_THREADSoverrides the number of compute workers. One worker is the appropriate setting for a two-CPU container when an explicit override is needed.LANCE_IO_THREADScontrols concurrent storage operations, not reserved CPU cores. Its default can be greater than the number of CPUs because I/O workers spend much of their time waiting for storage.
Keep the defaults unless measurements show that the workload benefits from an override. See the Lance threading model for the current defaults and tuning guidance.
Usage
Basic Example
import lancedb
db = lancedb.connect('<PATH_TO_LANCEDB_DATASET>')
table = db.open_table('my_table')
results = table.search([0.1, 0.3]).limit(20).to_list()
print(results)
Development
See CONTRIBUTING.md for information on how to contribute to LanceDB.