Files
foxhunt/crates
jgrusewski f782ec5a12 perf(cuda): NoisyLinear forward + copy fully GPU-native — zero CPU
noisy_layers.rs:
- forward(): cuBLAS sgemm for W=mu+sigma*epsilon matmul (was CPU loops)
- copy_params_from(): 6x DtoD async memcpy (was 12 PCIe roundtrips)
- Added noisy_add_bias CUDA kernel for bias addition

branching.rs:
- copy_weights_from(): DtoD memcpy for VarStore + NoisyLinear params
- register_mu_in_varstore(): DtoD clone (was host roundtrip)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-18 17:48:43 +01:00
..