perf(rl): migrate 25 hot-path launches to raw cuLaunchKernel

Browse Source

Replace all launch_builder().arg().launch() outside CUDA graph capture
with raw_launch() calling cuLaunchKernel directly. Eliminates ~200μs
per-launch cudarc overhead (trait dispatch, Arc, event guards).

25 launches migrated: PER push/sample/update/rebuild, HER track/inject
/forward, fused controllers, EMA producers, LR controller, spectral
norm, Q-bias, per-branch LR, grad accumulate, reduce_axis0.

61 launches inside graph capture regions kept as launch_builder
(only execute during warmup/capture steps 0-1).

Expected: ~7ms/step host overhead eliminated.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

...

This commit is contained in:

jgrusewski

2026-05-26 10:15:40 +02:00

parent 7e12099934

commit 41881f7724

1 changed files with 397 additions and 444 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

841

crates/ml-alpha/src/trainer/integrated.rs

File diff suppressed because it is too large Load Diff