attend op• Data kinds: tokens × tokens × tokens → tokens
• Call: import fullseye as fs; fs.ledger.attention_grouped(query, key, value, n_heads: 'int', n_kv_heads: 'int', mask=None, window=None) -> 'np.ndarray' (to call the implementation directly, import llmcore; llmcore.attention_grouped(query, key, value, n_heads: 'int', n_kv_heads: 'int', mask=None, window=None) -> 'np.ndarray'; from the registry, opsllmcore.get("attention_grouped"))
> This operator's description has not been translated yet. The original text follows as it is.
頭を束ねる注意(GQA)→ tokens。n_kv_heads == n_heads で MHA、== 1 で MQA に厳密一致。
Ainslie ら 2023。query は `n_heads 本に割り、key/value は n_kv_heads` 本しか
持たずに使い回す。両端が既存の 2 つ(MHA / MQA)に厳密に落ちる(どちらも 0.0e+00)
ので、「中間を取る」という主張が端で検算できる。
Args:
query: (T, n_heads·dh) float。頭ごとに幅 dh で並んでいる。
key: (S, n_kv_heads·dh) float。
value: (S, n_kv_heads·dh) float。
n_heads: query の頭の数。`n_kv_heads` の倍数であること。
n_kv_heads: key/value の頭の数。1 なら MQA、`n_heads` なら MHA。
mask: None / `"causal" / "window"` / (T, S) の bool 配列。
window: `mask="window"` のときの窓幅。
Returns:
tokens (T, n_heads·dh) float: 頭ごとの出力を横に連結したもの。
• 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.
• poc_attention_identities — py -3.11 examples/poc_attention_identities.py
tokens as input)rms_norm · rope_rotate · attention_scores · attention_apply · attention_softmax · attention_tiled · attention_linear · kv_cache_decode
attend)attention_softmax · attention_tiled · attention_linear
*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.