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

fix: remove redundant temporal/ISV re-execution in reduce_current_q_stats

Browse Source 
reduce_current_q_stats() re-ran 10 kernels (mamba2_step, predictive_coding,
regime_dropout, ISV forward/gate/route, recursive_confidence, trade_plan,
plan_noise, risk_budget) every training step OUTSIDE the graph replay.

These ops already executed inside the graph — the logit buffers are current.
The redundant execution added ~4s of GPU compute per step (matching the
graph replay cost), doubling total epoch time.

Fix: compute expected_q directly from the existing logit buffers.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-18 13:42:35 +02:00
parent c3cdb39081
commit 144e70ca29
1 changed files with 2 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
View File

@@ -7256,35 +7256,8 @@ impl GpuDqnTrainer {
let support_ptr = self.per_sample_support_ptr;
let q_var_ptr = self.q_var_buf_trainer.raw_ptr();
// Mamba2: enrich h_s2 with temporal context before Q-value computation
self.mamba2_step(batch_size)?;
// G12: Predictive coding — temporal smoothness on enriched trunk
self.compute_predictive_coding_loss(batch_size)?;
// G9: Regime-conditioned dropout on h_s2 (training path only)
self.apply_regime_dropout(batch_size, true)?;
// ISV forward: encoder MLP → branch gate + gamma mod
self.launch_isv_forward()?;
// ISV temporal routing: update per-feature temporal_weight for next mamba2 call
// One-step lag (mamba2 already ran above), same pattern as ISV signals.
self.launch_isv_temporal_route()?;
// ISV feature gate: modulate h_s2 features based on ISV regime embedding
self.launch_isv_feature_gate(batch_size)?;
// Recursive confidence: predict own TD-error from h_s2
self.launch_recursive_confidence_forward(batch_size)?;
// Trade plan forward: h_s2 → plan_params [B, 6]
self.launch_trade_plan_forward(batch_size)?;
// Plan noise: ±5% temporal diversity on plan_params
self.launch_plan_noise_inject(batch_size)?;
// Risk budget forward: h_s2 → risk_budget R ∈ (0,1) before Q-value computation
self.risk_budget_forward(batch_size)?;
// All temporal/ISV ops already ran inside the graph replay.
// Logit buffers (on_v_logits, on_b_logits) are current — just compute Q-stats.
unsafe {
self.stream