File size: 933 Bytes
80692f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 | """Canonical shape sweep for TopK.
Mix of:
- decoder vocab top-k (single sequence, very large n, moderate k) — pure
bandwidth test; the input read dominates everything.
- prefill / batched attention top-k (many rows, moderate n, small k) — tests
per-row parallelism and shared-memory bitonic networks.
- non-power-of-2 n stress case — bitonic sort networks naturally want
powers of two; this forces the agent to handle padding or partial sorts.
- small-k limit — k=1 (argmax) is a degenerate but useful sanity case.
"""
SHAPES = [
{"batch": 1, "n": 131072, "k": 64}, # decoder vocab top-k (Llama vocab ~128k)
{"batch": 64, "n": 8192, "k": 8}, # prefill / attention top-k
{"batch": 32, "n": 16384, "k": 32}, # mid-size batched
{"batch": 16, "n": 12000, "k": 16}, # non-power-of-2 n stress
{"batch": 128, "n": 4096, "k": 1}, # batched argmax (k=1 corner case)
]
|