Struct RecurLayer 

pub struct RecurLayer {
    pub conv_state: CudaSlice<f32>,
    pub ssm_state: CudaSlice<f32>,
    pub ssm_state_alt: CudaSlice<f32>,
}

Per-linear-attn-layer fixed recurrent state. conv_state and ssm_state are BOTH kept RESIDENT on GPU — the conv ring assemble + roll runs on-device (conv_assemble_and_roll), so there is no per-step dtoh/htod for either.

Fields

conv_state: CudaSlice<f32>

ssm_state: CudaSlice<f32>

ssm_state_alt: CudaSlice<f32>

PERSISTENT second SSM-state buffer for the gdn-scan double buffer (DECODE DETERMINISM FIX). gdn_scan needs DISTINCT in/out state buffers. The old eager path allocated a fresh state_scratch via e.uninit every step and swapped its pointer into ssm_state; that per-step alloc/free churned the stream-ordered async pool, and the freed prior ssm_state block was recycled by the next step’s scratch while a kernel referencing the swapped-in state was still in flight — a use-after-reuse that produced RUN-TO-RUN nondeterministic decode (two identical prompt primes diverged). We instead PING-PONG between two STABLE resident buffers (no per-step alloc/free, no pool churn): step writes into the spare, then swaps the two owned buffers in place. Stable pointers, identical math. Sized like ssm_state.