perf: widen hidden_dim_base to 1024 — H100 has headroom at 0.7ms/step
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -502,7 +502,7 @@ impl ParameterSpace for DQNParams {
|
||||
(0.005_f64.ln(), 0.01_f64.ln()), // 20: tau (log scale, Polyak soft update)
|
||||
|
||||
// GPU-dynamic network sizing (1D)
|
||||
(128.0, 512.0), // 21: hidden_dim_base (128-512: covers RTX3050→H100 range, 1024+ is wasteful)
|
||||
(128.0, 1024.0), // 21: hidden_dim_base (128-1024: 4x production default, H100 has headroom at 0.7ms/step)
|
||||
|
||||
// CQL regularization (1D)
|
||||
(0.0, 1.0), // 22: cql_alpha (0=disabled, 0.1=mild, 1.0=full offline-RL strength)
|
||||
|
||||
Reference in New Issue
Block a user