Module ids 

PLE n-gram row ids: EOS-aware right-shift plus the multiply-XOR hash.

Pure token-id arithmetic — no device, no weights — which is why it is tested bit-exactly against the reference in tests.rs and runs in CI.

This does NOT transfer from LongCat (#746). LongCat accumulates a polynomial rolling hash (acc += shift[d] * m[d], then % rows). Qwen multiplies each shifted token by a SplitMix64-derived odd multiplier and XORs them, then takes a per-head prime modulus and adds a per-head offset into one 320M-row table. The two produce different rows from the same tokens, and both produce VALID rows — so a mix-up is silent.

Reference: Qwen4ExpTextNGramEmbedding.forward / _shift_right_ignore_eos, bench/qwen4_exp/ref/modeling_qwen4_exp.py.

Structs

PleIdDims

Geometry for one PLE site, read from the checkpoint rather than derived.

Functions

ple_ngram_ids

Row ids for every head, one row per token in tokens.