Programmatic Dependent Launch (PDL) for more performance on newer NVIDIA GPUs (Hopper+)#22522

Merged

JohannesGaessler merged 58 commits into

ggml-org:masterfrom

aendk:akieslinger/pdl-cuda-lc-experiments

May 20, 2026

Commits on Feb 5, 2026

Moves pdl logic to separate function, removes some whitespace
aendk
committed

Commits on Feb 13, 2026

Commits on Feb 19, 2026

Commits on Mar 11, 2026

Commits on Mar 24, 2026

Merge branch 'master' into akieslinger/pdl-cuda
aendk
committed

Commits on Mar 25, 2026

Added LC according to internal guidance and tested kernel performance.
aendk
committed

Commits on Apr 2, 2026

Enrols missing qwen3-5 kernels passively into PDL.
aendk
committed

Commits on Apr 10, 2026

Commits on Apr 16, 2026

Merge branch 'master' into akieslinger/pdl-cuda-lc-experiments
aendk
committed

Commits on Apr 20, 2026

Merge branch 'master' into akieslinger/pdl-cuda-lc-experiments
aendk
committed

Commits on Apr 29, 2026

Commits on May 11, 2026

Adds stricter checks to enable PDL, adds env-check to disable it, and removes now superfluous compile option to enable PDL.
aendk
committed

Commits on May 12, 2026

Fix: Correct PDL en/disablement based on device-side arch check. Host
aendk
committed

Commits on May 13, 2026

Commits on May 18, 2026

Commits on May 19, 2026

Commits on May 20, 2026

[Whitespace] Adds final newline to common.cuh to make editorconfig CI job happy
aendk
committed