perf(cuda): TFT + Mamba2 forward passes fully GPU-native
TFT (10 downloads eliminated):
- lstm_encoder: gpu_select_dim1 per-timestep, gpu_cat_dim0 assembly
- variable_selection: gpu_narrow_2d column select, gpu_broadcast_mul_col
weighted sum, gpu_mean_all per-column stats
- quantile_outputs: gpu_select_dim1 last timestep, gpu_stack_2d output,
GPU quantile loss (sub→abs→scale→add→mean_all)
- mod.rs: GPU 3D broadcast for static context
Mamba2 (10 downloads eliminated):
- loss.rs: GPU MSE (sub→sqr→mean_all), GPU directional MSE with sign
detection (mul→abs→div→scalar_sub→scale→mean_all)
- scan_algorithms: GPU sequential scan via gpu_select_dim1 per-step
- selective_state: GPU importance scoring (abs→mean_all)
- mod.rs: GPU state-space recurrence (select_dim1→matmul→add),
GPU accuracy computation (sub→abs→relu→clamp→mean_all)
169/169 ml-supervised tests pass.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
33864badf3