Two-phase kernel: q_mean_reduce computes global mean, q_mean_subtract
centers all Q-values. Runs after compute_expected_q, before Q-attention.
Fixes Q-mean drift 0→+0.47 observed in train-vpb4w.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Observed train-vpb4w freeze at epoch 23 (val_Sharpe=51.11):
- avg_grad=10.0 (nonzero!) but Q-stats FROZEN for 4+ epochs
- Root cause: Adam momentum converged to fixed point, not zero gradient
Component 10: Targeted Adam momentum reset for branch weights only
(indices 8-41). Lighter than full rewind, preserves trunk convergence.
Component 11: Q-gap target attractor for spread gradient. Amplifies
spread when Q-gap is below target, reduces when above. Prevents
spread from being too weak at high Q-gap levels.
Component 12: Cosine LR schedule with warm restarts (T=20 epochs).
LR decay prevents overshoot, restart jumps break fixed points.
Half-momentum at restart preserves gradient direction.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
entropy_coeff ramps when atom utilization EMA drops below 50%.
Linear ramp: 0 extra at 50%+, 1× base_entropy extra at 0%.
Prevents distributional collapse 100%→20% observed in train-vpb4w.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Task 10: q_mean_reduce + q_mean_subtract two-phase kernel. Zero params.
Fixes Q-mean drift 0→+0.47 from bootstrapping bias.
Task 11: Adaptive entropy_coeff based on utilization_ema. Zero params.
Fixes atom utilization collapse 100%→20%.
Execution order: 10, 11, then 1-9.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Observed in train-vpb4w (live H100 run):
- Q-mean drifts 0→+0.47 over 18 epochs (bootstrapping bias)
- Atom utilization drops 100%→20% (distributional collapse)
- val_Sharpe peaks 54.59 then decays to 31.52
Component 8: Mean-center Q-values before action selection pipeline.
Zero params, preserves ranking, prevents drift accumulation.
Component 9: Adaptive atom entropy regularization. Ramps entropy_coeff
when utilization drops below 50%. Uses existing c51_grad_kernel param.
Both are zero-param fixes moved to top of implementation order.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Softmax over SH2=256 features gives mask[i]=1/256=0.004 per feature,
collapsing the input signal 256×. Sigmoid gives mask[i]∈(0,1) independently
per feature. At W_vsn2=0 init: sigmoid(0)=0.5 → h_masked = 0.5*h_s2.
Shared memory reduced: SH2+R (no mask buffer needed, sigmoid is inline).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sigmoid(Q·K/√d) × V is a per-head feature gate, not standard multi-head
attention (seq_len=1 makes softmax trivial). Comments now accurately
describe the learned gating mechanism and its purpose.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
LOW 1: sel_t_buf CudaSlice + per-step memcpy_htod → sel_t_pinned
device-mapped. Zero copies, GPU reads directly from host memory.
LOW 2: is_q_gap_frozen uniform [q_gap; 4] → actual per-branch Q-gaps
from 48B DtoH readback in reduce_current_q_stats. Each branch now
independently detected as frozen/unfrozen. Liquid tau and trajectory
backtracking use real per-branch velocities.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
mag_concat now uses vsn_masked (feature-selected) instead of raw h_s2
for its first SH2 columns. strided_scatter kernel writes tight [B,SH2]
vsn_masked into wide [B,SH2+3] mag_concat preserving Q_dir in last 3.
launch_mag_concat → launch_mag_concat_from with explicit source ptr.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
CRITICAL: backward_full used state_dim for goff_b_s1 offset but
compute_param_sizes uses s1_input_dim (smaller with bottleneck active).
All downstream gradient offsets were misaligned. Fixed: sd → s1d.
HIGH: q_attn_params used alloc_f32 (potentially non-zero) instead of
alloc_zeros. Cross-branch Q-attention residual connection needs near-zero
init for stability. Fixed: alloc_zeros.
MEDIUM (deferred): VSN masking bypassed for magnitude branch — mag_concat
reads raw h_s2 instead of vsn_masked. No impact while W_vsn2 is zero-init
(identity mask). Will fix when VSN backward is wired.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
selectivity_forward: sigmoid(dot(W_sel, h_s2) + b_sel) per sample.
selectivity_backward: BCE gradient on (sel, per_sample_loss/mean_loss).
Separate Adam at LR=1e-4. PER priority integration deferred.
Also fixes kernel names (selectivity_gate_fwd→selectivity_forward,
selectivity_gate_bwd→selectivity_backward) to match experience_kernels.cu.
Adds sel_norm_buf, sel_norm_partials, sel_clip_buf, sel_t_buf fields
needed by dqn_adam_update_kernel. Wired into run_full_step after main
Adam step with mean_loss=1.0 placeholder and non-fatal error handling.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
c51_grad_kernel reads liquid_mod[d] per branch instead of spread_velocity[0].
4 per-branch Q-gap EMAs with adaptive alpha drive continuous-time ODE.
Fast branch learning → large tau → slow adaptation → don't overshoot.
Stuck branch → small tau → fast adaptation → push harder.
spread_velocity infrastructure removed (-30 lines).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
launch_q_attention runs 2-head attention on 12 Q-values [B, 12].
Wired into reduce_current_q_stats after compute_expected_q.
Also fixes kernel function name: q_cross_branch_attn → cross_branch_q_attention.
Experience collection deferred — uses raw Q-values for now.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
d_h_bd → glu_backward → d_value + d_gate_pre. Two weight gradient
GEMMs per branch (value + gate). Upstream gradient sums both paths.
VSN backward DEFERRED (W_vsn init to identity, zero gradient for now).
Gradient flows to gate weights at indices 34-41.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Checkpoint ring buffer ranked by improvement rate, plateau detection
via liquid tau velocity, informed perturbation cycle (Adam reset,
shrink-perturb, temp boost, LR×2), depth-limited rewind. ~150 lines
in training_loop.rs, zero CUDA changes. Recovers from plateaus that
prevention mechanisms (spread gradient, liquid tau) fail to avoid.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
param_sizes[12] grows by 3*adv_h for direction conditioning.
mag_concat_buf [B, SH2+3] and d_mag_concat_buf allocated.
mag_concat_qdir and strided_accumulate kernels loaded.
launch_mag_concat and accumulate_d_h_s2_from_concat methods added.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
mag_concat_qdir: computes direction E[Q] from softmax over C51 atoms,
normalizes by delta_z for matched gradient scale, concatenates with h_s2
to produce [B, SH2+3] input for magnitude branch FC.
strided_accumulate: extracts first SH2 columns from d_mag_concat [B, SH2+3]
into d_h_s2 [B, SH2] with beta accumulation for backward pass.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Raw E[Q] ~ 0.1 while h_s2 ~ 1.0 post-ReLU. Without normalization,
Q_dir columns learn 10x slower. Dividing by delta_z scales to ~1-10.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
In-place ranking races across blocks: block X writes shaped rewards[i]
before block Y loads it in tile scan. __syncthreads is block-local.
Fix: read from rewards_out (raw), write to gpu_batch.rewards (clone).
The DtoD clone from collect_experiences_gpu provides the output buffer.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Magnitude FC input changes from h_s2[B, SH2] to [h_s2; Q_dir][B, SH2+3].
Direction E[Q] values (3 scalars per sample) are computed from branch 0
logits and concatenated. w_b1fc grows by 3*adv_h parameters (768 for AH=256).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Four improvements to the training pipeline:
1. Asymmetric spread gradient: challenger action (adjacent to taken) gets
pushed UP at half strength instead of DOWN. Creates a two-horse race
instead of single-action monopoly.
2. Per-branch spread scaling: spread_grad *= branch_scale. Direction branch
(high impact) gets more spread than urgency (low impact).
3. Distributional variance position sizing (Layer 4): Var[Q] = E[Z²] - E[Z]²
computed in compute_expected_q. Position scaled by 1/(1+sqrt(Var[Q_taken])).
High uncertainty → smaller position. Kelly criterion from C51 atoms.
4. Reward std guard: skip rank normalization when observed_reward_std ≈ 0
(epoch 0). Prevents zeroing all rewards before std is observed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a spread gradient to the C51 advantage logits that pushes the
taken action's distribution toward higher atoms and non-taken toward
lower. Scale = inv_batch * delta_z (adaptive to per-sample atom
resolution, zero hardcoded constants).
This gradient is ORTHOGONAL to the Bellman equation — it depends on
atom position, not target match. Active on ALL samples, providing
perpetual pressure to differentiate Q-values even when the C51
cross-entropy gradient vanishes at convergence. Prevents the Q-gap
plateau where all actions have identical Q-values.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Philox seed used batch_size+sample_id (constant across steps) so the
same action was always sampled for the same sample → at the fixed point,
identical target every step → zero gradient. Now uses t_buf (Adam step
counter, device-mapped, increments every step). Each training step
samples a DIFFERENT action per sample → target is truly stochastic →
gradient variance prevents convergence.
Also moves philox_uniform to common_device_functions.cuh for reuse.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Deterministic softmax mixture produced a fixed target that the model
matched exactly → zero gradient → Q-gap frozen at 0.167, val_Sharpe
frozen at -5.62 from epoch 22. Stochastic sampling picks ONE action
per sample from the softmax weights. Expected target over many steps
is identical (unbiased) but per-step target varies → model can never
match a moving target → perpetual gradient → no convergence trap.
Also moves philox_uniform to common_device_functions.cuh so all
kernels can use it.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Floor = max(|mean_Q| * 0.01, 1e-6). At Q-mean=0.005: floor=5e-5.
At Q-mean=1.0: floor=0.01. Fully adaptive, zero hardcoded constants.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
tau=Q_gap/3 gave 20:1 weight ratio (95%/4%/1%), making the target
essentially argmax. Once the online distribution matched this sharp
target, cross-entropy gradient vanished → Q-gap froze at 0.158,
val_Sharpe locked at 2.40 from epoch 22.
tau=Q_gap gives a CONSTANT e:1 ≈ 2.72:1 ratio (63%/23%/14% for 3
actions) regardless of Q-gap magnitude. The target always maintains
meaningful mixture → perpetual gradient pressure → Q-values keep
evolving. Scale-invariant: the softmax ratio depends only on relative
Q-differences, not their absolute magnitude.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All 4 branches (dir, mag, order, urgency) now use Boltzmann sampling
in eval mode instead of argmax. argmax(softmax(Q/tau)) = argmax(Q) —
temperature is meaningless with argmax. Every state picked the same
winning action → identical trades → Sharpe=0.00 from epoch 8 onward.
Boltzmann sampling with Philox seed gives deterministic but
action-diverse evaluation. When Q-values differentiate, Boltzmann
naturally sharpens toward the best action.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Softmax weights used implicit tau=1, producing near-uniform weights
when Q-gap=0.06: exp(0.06)≈1.06 → weights [0.35, 0.33, 0.32].
The mixture averaged all distributions → zero gradient to differentiate
actions → Q-gap frozen → val_Sharpe locked at 0.00 after epoch 19.
Now uses tau = max(Q_gap/3, 0.001). With Q-gap=0.06 and tau=0.02:
exp(0.06/0.02) = exp(3) ≈ 20:1 weight ratio between best and worst.
The 0.001 floor means even tiny Q-differences produce sharp weights,
preventing the uniform-averaging fixed point.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
read_total_loss() calls cuStreamSynchronize which is illegal during
CUDA Graph stream capture. Was inside submit_aux_ops (captured in
graph_mega/graph_aux). Moved to Step 4 (conditional ops outside graph).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
branching_action_select exposure head used hard argmax while order/
urgency used Boltzmann. experience_action_select eval mode used pure
argmax for direction. Both replaced with Boltzmann softmax:
tau = max(Q_max - Q_min, 0.01)
When Q-values differentiate: converges to argmax. When flat: spreads
evenly. Eval mode takes argmax of softmax probs (deterministic).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
