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

cleanup(dqn): remove val-collapse diagnostic printf — fix verified

Browse Source 
Diagnostic kernel printf in backtest_env_step_batch identified the gate
(Kelly cap warm-branch deadlock) and verified the fix
(0c9d1ee39: max(kelly_f, warmup_floor)) on train-4r6p8:

  Before fix (train-4fpzx, fresh model epoch 0):
    val_picked_dir_dist [short=0.19 hold=0.20 long=0.39 flat=0.21]
    val_dir_dist        [short=0.0001 hold=0.20 long=0.0000 flat=0.80]
    trade_count = 23 over 214K bars  (active_frac = 0.0001)

  After fix (train-4r6p8, fresh model epoch 0):
    val_picked_dir_dist [short=0.24 hold=0.17 long=0.42 flat=0.17]
    val_dir_dist        [short=0.24 hold=0.17 long=0.42 flat=0.17]   1:1
    trade_count = 139,695 over 214K bars  (active_frac = 0.6590)

  Picked and realised distributions now bit-identical — every Boltzmann
  pick translates faithfully to actual_dir.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-26 21:52:08 +02:00
parent 0c9d1ee39e
commit 2f4bd8e58b
2 changed files with 1 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
crates/ml/src/cuda_pipeline/backtest_env_kernel.cu
View File

@@ -486,34 +486,6 @@ extern "C" __global__ void backtest_env_step_batch(
&actual_mag,
&trail_triggered
);
/* DIAG (val-collapse): sample bars across the window for window 0.
* First 10 bars = initial behavior. Then every 100 bars from step 100
* to 1000 (chunk 1-15 transition zone), every 1000 bars to 10000,
* every 10000 bars to 200000. Plus a "mismatch trigger" — print on
* every bar where action picks Long/Short but actual_dir is not
* Long/Short (the gate event). Limit total to first 200 mismatches
* to keep printf buffer manageable. Removed once gate identified. */
int dir_dbg = action_val / (b1_size * b2_size * b3_size);
int mag_dbg = (action_val / (b2_size * b3_size)) % b1_size;
bool sample_print = (current_step < 10
|| (current_step < 1000 && current_step % 100 == 0)
|| (current_step < 10000 && current_step % 1000 == 0)
|| current_step % 10000 == 0
|| current_step == 213000
|| current_step == 214000);
bool mismatch = (dir_dbg == 0 || dir_dbg == 2) /* picked Short or Long */
&& actual_dir != dir_dbg
&& current_step >= 10
&& current_step % 137 == 0; /* prime stride to avoid pattern */
if (w == 0 && (sample_print || mismatch)) {
printf("[VALDIAG step=%d] act=%d dir=%d mag=%d close=%.2f maxpos=%.4f "
"prevpos=%.4f pos_post=%.4f actual_dir=%d actual_mag=%d trail=%d "
"conv=%.3f health=%.3f value=%.2f maxeq=%.2f cash=%.2f\n",
current_step, action_val, dir_dbg, mag_dbg, close, max_position,
prev_position, position, actual_dir, actual_mag, trail_triggered,
conviction_core, health_k_batch, value, max_equity, cash);
}
(void)prev_position; (void)prev_position_sign; (void)trail_triggered;
/* ── Post-trade floor check: catch intra-step breaches ────────── */

2
docs/dqn-wire-up-audit.md
View File

@@ -259,7 +259,7 @@ P5T5 Phase E (2026-04-26): per-fold reset for `MetricBandsRegistry` regression-d
| `attention_backward_kernel.cu` | `gpu_attention.rs`, `gpu_tlob.rs` (reused for grad_norm+Adam kernels) | Wired | Attention backward pass | — |
| `tlob_kernel.cu` (`tlob_sdp_forward`, `tlob_sdp_backward`, `tlob_write_states`, `tlob_read_grad_from_states`) | `gpu_tlob.rs` | Wired | D.8 Plan 2 Task 6C — TLOB SDP forward/backward, state scatter/read kernels | — |
| `training_guard_kernel.cu` (`training_guard_check_and_accumulate`) | `gpu_training_guard.rs` | Wired | NaN / guard check kernel | — |
| `backtest_env_kernel.cu` (`backtest_env_step`) | `gpu_backtest_evaluator.rs` (`ENV_CUBIN`) | Wired | Backtest environment step. Diagnostic printf at `backtest_env_step_batch` sampled across full 214K val window for window 0 — tracking val-collapse Long/Short→Flat gate (first 10 bars confirmed kernel-correct; gate must activate beyond first chunk); removed once root cause identified. | — |
| `backtest_env_kernel.cu` (`backtest_env_step`) | `gpu_backtest_evaluator.rs` (`ENV_CUBIN`) | Wired | Backtest environment step | — |
| `backtest_metrics_kernel.cu` | `gpu_backtest_evaluator.rs` (`METRICS_CUBIN`) | Wired | Backtest metrics reduction | — |
| `backtest_plan_kernel.cu` (`backtest_plan_diag_reduce`, `backtest_plan_state_isv`) | `gpu_backtest_evaluator.rs` (`BACKTEST_PLAN_CUBIN`) | Wired | Plan-ISV diagnostic reduction in val. D.6 (Plan 2 Task 6A): `backtest_plan_state_isv` now writes `pisv[6] = remaining_fraction`; stride updated to `SL_PORTFOLIO_PLAN_DIM=7`. | — |
| `backtest_forward_ppo_kernel.cu` (`backtest_forward_ppo_kernel`) | `gpu_backtest_evaluator.rs::evaluate_ppo` (`PPO_FORWARD_CUBIN`) | OUT-of-DQN-scope | PPO-path backtest only; correct-as-is | — |