attention_tiled — LLMCORE attend op

• Data kinds: tokens × tokens × tokens → tokens

• Call: import fullseye as fs; fs.ledger.attention_tiled(query, key, value, tile: 'int' = 64, mask=None, window=None, scale=None) -> 'np.ndarray' (to call the implementation directly, import llmcore; llmcore.attention_tiled(query, key, value, tile: 'int' = 64, mask=None, window=None, scale=None) -> 'np.ndarray'; from the registry, opsllmcore.get("attention_tiled"))

Usage

> This operator's description has not been translated yet. The original text follows as it is.

タイル + online softmax の注意(FlashAttention の芯)→ tokens。近似ではない。

Dao ら 2022。key/value を `tile` 行ずつ読み、走っている最大値と分母をその場で

補正しながら足す。作る中間行列は (T, `tile`) だけで、(T, T) を一度も作らない ——

速さの出どころは近似ではなく、置き場所。`attention_softmax` との相対差は

タイル 128 枚でも 1.2e-15(タイル 1 枚で 8.2e-16 なので、枚数を 128 倍しても 1.5 倍)。

Args:

query: (T, d) float。

key: (S, d) float。

value: (S, dv) float。

tile: 一度に読む key/value の行数。小さいほど中間行列が小さく、枚数が増える。

mask: None / `"causal" / "window"` / (T, S) の bool 配列。

window: `mask="window"` のときの窓幅。

scale: スコアの倍率。None なら 1/√d。

Returns:

tokens (T, dv) float: `attention_softmax` と機械精度で一致する。

Detailed usage guide

• llmcore family guide

References (sample data, literature)

• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).

• Operator provenance and references — the sources of the research/methods this op family came from.

• The canonical algorithm (author, year) and its uses are named in the family usage guide above.

Runnable examples (verified samples that actually call this op)

• poc_attention_identities — py -3.11 examples/poc_attention_identities.py

Ops the type connects to (they accept tokens as input)

rms_norm · rope_rotate · attention_scores · attention_apply · attention_softmax · attention_linear · attention_grouped · kv_cache_decode

Same category (attend)

attention_softmax · attention_linear · attention_grouped


*Provenance: llmcore.py — LLMCORE operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*

© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.