Files
xserv/crates
Gahow Wang 99b212e6c1 model/sampling: NaN-safe argmax + optional repetition penalty
Make argmax skip NaN logits (warn once) instead of panicking the engine
thread on a single NaN. Add sample_greedy_penalized() applying an
HF-style repetition penalty over recent ids on the greedy path, to break
greedy repetition loops on reasoning models without touching the forward
pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-05-31 00:56:27 +08:00
..