jgrusewski
2ea96c4c00
fix(cuda): restore graph capture, fix SCRATCH1_DIM OOB, fix BF16 grad norm
- CUDA training kernel: SCRATCH1_DIM = max(SHARED_H1, VALUE_H + ADV_H)
fixes register array OOB in dqn_forward_loss_kernel distributional path
- gpu_dqn_trainer: restore CUDA graph capture/replay (was bypassed for
debug), remove 4 debug stream.synchronize() that serialized GPU pipeline
- gpu_dqn_trainer: disable/enable cudarc event tracking around graph
capture to prevent cross-stream event references from invalidating capture
- gpu_dqn_trainer: opt-in to >48KB dynamic shared memory when needed
- fused_training + gpu_backtest_evaluator: fork() dedicated stream for
graph capture (legacy stream 0 does not support begin_capture)
- gpu_weights: update struct comments from scalar to distributional C51
shapes, keep BranchingWeightSet::zeros() scalar-sized for experience
collector placeholder buffers
- 5 supervised adapters (tgnn/tlob/kan/xlstm/diffusion): cast BF16 grad
norm tensors to F32 before to_scalar::<f32>() extraction
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>