flashinfer-ai/flashinfer

[Compilation] GDN decode cold kernel compilation time is too long on SM90/100/120/121

Ouverte

#4 110 ouverte le 23 juil. 2026

 (12 commentaires) (0 réaction) (1 personne assignée)Python (1 031 forks)github user discovery
good first issueop: gdn

Métriques du dépôt

Stars
 (5 756 étoiles)
Métriques de merge PR
 (Merge moyen 11j 23h) (186 PRs mergées en 30 j)

Description

Noticed by @bkryu : "In our recent CI telemetry, we have been seeing that tests/gdn/test_decode_delta_rule.py runs very long (30-50 mins on every arch) and has been straining our CI.

TL;DR of the cause: the reference and kernels are all fine. Most time is spent on cold kernel compilation. On a B200 I locally see ~800 cases that collapses to ~200 cute.compile calls based on the cache key logic.

For some kernels, B/tile_v/T/cache flags/pool-strides each key a new compile. Would be great if we can make the kernels batch and/or seqlen dynamic to systematically reduce unit testing time."

Guide contributeur