gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: cap auto batch_size at 8192 — prevents 2M+ nonsense on H100

Browse Source 
AutoBatchSizer computes raw VRAM ceiling which can be millions on 80GB.
Cap at 8192: saturates H100 132 SMs for our GEMM sizes (80x256x128),
diminishing returns above this for DQN gradient quality.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-02 14:51:58 +02:00
parent 99392ffcb4
commit 3d53735363
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
crates/ml/src/trainers/dqn/trainer/constructor.rs
View File

@@ -95,7 +95,7 @@ impl DQNTrainer {
// batch_size == 0 -> auto-compute from VRAM (no artificial cap)
if hyperparams.batch_size == 0 {
hyperparams.batch_size = max_safe_batch;
hyperparams.batch_size = max_safe_batch.min(8192);
info!("AutoBatchSizer: batch_size auto-computed to {}", hyperparams.batch_size);
}
@@ -106,7 +106,7 @@ impl DQNTrainer {
"DQN batch_size capped from {} -> {} (VRAM ceiling)",
hyperparams.batch_size, max_safe_batch
);
hyperparams.batch_size = max_safe_batch;
hyperparams.batch_size = max_safe_batch.min(8192);
}
// Use override device if provided (hyperopt shares one CUDA context),