Expand description
Fused GPU kernel for DeltaNet g and beta computation.
g[t, vh] = softplus(alpha_logit[t, vh] + dt_bias[vh]) * (-ssm_a[vh])
beta[t, vh] = sigmoid(beta_logit[t, vh])
Replaces the CPU bridge in compute_g_and_beta_cpu.
For seq=1 this dispatches 1×nv threads — trivial GPU work but eliminates
2 CPU-GPU buffer downloads (alpha_logit, beta_logit) per delta-net layer.
Statics§
Functions§
- compute_
g_ beta_ gpu - Allocate output buffers, dispatch the kernel, commit, and return
(g, beta). - dispatch_
compute_ g_ beta - Dispatch
compute_g_betakernel intoencoder. - register
- Register
compute_g_beta_f32shader with the kernel registry.