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

perf: IQN quantiles 64→32 — saves 4.3GB VRAM, halves IQN GEMMs

Browse Source 
Academic literature shows diminishing returns past 32 quantiles.
64 quantiles allocated 8.6GB for tiled intermediates (B×Q×hidden).
32 quantiles: 4.3GB (half), ~13 GEMMs halved in aux_child.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-18 22:28:23 +02:00
parent 1481ba2699
commit 76cfc0d185
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
crates/ml/src/trainers/dqn/config.rs
View File

@@ -1739,7 +1739,7 @@ pub(crate) fn dqn_default_config() -> DQNConfig {
// CQL + IQN (2026 modernization)
cql_alpha: 1.0,
iqn_num_quantiles: 64,
iqn_num_quantiles: 32, // 64→32: saves 4.3GB VRAM, halves IQN GEMMs, diminishing returns past 32
iqn_kappa: 1.0,
iqn_embedding_dim: 64,
cvar_alpha: 0.05,